LAND COVER CHANGE AND HYDROLOGICAL REGIMES IN THE …

LAND COVER CHANGE AND HYDROLOGICAL

REGIMES IN THE SHIRE RIVER CATCHMENT,

MALAWI

Lobina Getrude Chozenga Palamuleni

Student Number 920418079

A thesis submitted in fulfilment of the requirements for the degree of Doctor of

Philosophy in the Department of Geography, Environmental Management and

Energy Studies, University of Johannesburg.

Supervisor: Professor Harold John Annegarn

Johannesburg, 20 March 2009

i

Declaration

I declare that the work contained in this thesis is my own original writing. Sources

referred to in the creation of this work have been appropriately acknowledged by

explicit references or footnotes. Other assistance received has been acknowledged. I

have not knowingly copied or used the words or ideas of others without such

acknowledgement.

Signed ___________________ Date __________________

Sections of this work have been presented at conferences and have been submitted for

Journal publication:

Palamuleni, L. G., T. Landmann and H. J. Annegarn, (2006), Land cover

mapping for the Shire River catchment in Malawi using Landsat

satellite data, Palamuleni, L. G., T. Landmann and H. J. Annegarn,

(2008), Awarded the Best Paper at the 6th African Association of

Remote Sensing of the Environment (AARSE) Conference, 30 October

-2 November 2006, Cairo, Egypt.

Palamuleni, L. G., T. Landmann and H. J. Annegarn, (2007), Mapping rural

savanna woodlands, a comparison of maximum likelihood and fuzzy

classifiers, International Geoscience and Remote Sensing Symposium

(IGARSS’07), 23-27 July 2007, Barcelona, Spain

Palamuleni, L. G., T. Landmann and H. J. Annegarn, (2008), An assessment

of land cover change using multi-temporal Landsat imagery for the

Shire River catchment, Malawi, 7th African Association of Remote

Sensing of the Environment (AARSE) Conference, 27 - 30 October

2008, Accra, Ghana.

Palamuleni, L. G., T. Landmann and H. J. Annegarn, (2008), Processing

changes in land cover using Landsat imagery for the Shire River

catchment, Malawi. Paper submitted to Journal of Applied Earth

Observation and Geoinformation.

Palamuleni, L. G., H. J. Annegarn, T. Landmann amd P. M. Ndomba, (2008),

Application of the AVSWATX Tool for the Shire River Sub-Catchment in

Malawi. Paper submitted to Hydrological Sciences Journal.

ii

Dedication

To my husband, Dr Martin E. Palamuleni, for showing me the path to greatness

and

to my sister, Rhoda, and my lovely daughters Tadala and Tamanda

for encouraging me against all odds

iii

Abstract

Land cover changes associated with growing human populations and expected

changes in climatic conditions are likely to accelerate alterations in hydrological

phenomena and processes on various scales. Subsequently, these changes could

significantly influence the quantity and quality of water resources for both nature and

human society. Documenting the distribution of land cover types within the Shire

River catchment is the foundation for applications in this study of the hydrology of the

Shire catchment.

The aim of this study is to investigate the relationships between the measured land

cover changes and hydrological regimes in the Shire River Catchment in Malawi.

Maps depicting land cover dynamics for 1989 and 2002 were derived from multi-

spectral and multi-temporal Landsat 5 (1989) and Landsat 7 ETM+ (2002) satellite

remote sensing data for this catchment. Other spectral-independent data sets included

the 90-m resolution Shuttle Radar Topographic Mission (SRTM) digital elevation

model (DEM), Geographical Information System (GIS) layers of soils, geology and

archived land cover. Core image-derived data sets such as individual Landsat bands,

Normalized Difference Vegetation Index (NDVI), Principal Components Analysis and

Tasseled Cap transformations were computed. From generated composite images,

land cover classes were identified using a maximum likelihood algorithm. Eight land

cover classes were mapped.

A hierarchical multispectral shape classifier with an object conditional approach

determined by the Food and Agriculture Organisation (FAO) Land Cover

Classification System (LCCS) legend structure was used to map land cover variables.

LCCS was used as a basis for classification to achieve legend harmonization within

Africa and on a global scale. Flexibility of the hierarchical system allowed

incorporation of digital elevation objects, soil and underlying geological features as

well as other available geographical data sets. This approach improved classification

accuracy and can be adopted to discriminate land cover features at several scales,

which are internally relatively homogeneous. In addition to compatibility with the

FAO/LCCS classification system, the derived land cover maps have provided recent

iv

and improved classification accuracy, and added thematic detail compared to the

existing 1992 land cover maps.

Fieldwork was conducted to validate the land cover classes identified during

classification. Accuracy assessment was based on the correlation between ground

reference samples collected during field exercise and the satellite image classification.

The overall mapping accuracy was 87%, with individual classes being mapped at

accuracies of above 77% for both user and producer accuracy. The combination of

Landsat images, vector data and detailed ground truthing information was used

successfully to classify land cover of the Shire River catchment for years 1989 and

2002.

Quantitative changes in the areas of various land cover categories and the direction of

change were determined. Land cover change detection was carried out by Multi-date

visual compositing, followed by Post-classification analysis. For the first step,

degradation of vegetation was chosen as the main indicator of change, while post

classification statistical analysis was employed to determine the specific nature of

changes in each land cover type. Multi-date visual composites were found to detect

areas of change and of no change better than post-classification. Using the post-

classification procedure, areal statistics and direction of change in each land cover

class were derived using a combination of both methods. This activity highlights areas

where there are major changes of land cover (i.e. "hot spots"), both in temporal and

spatial aspects. The study revealed significant changes in magnitude and direction that

have occurred in the catchment between 1989 and 2002, mainly in areas of human

habitation. Trends in land cover change in the upper Shire River catchment depict

land cover transition from woodlands to mostly cultivated/grazing and built-up areas.

Twelve per cent of the total land surface of the study area had been converted to

cultivation/grazing over a 13-year interval.

Positive changes (referring to reforestation of degraded areas) in woody closed areas

especially within the former refugee areas close to the Mozambican border, provides

some evidence of the ecological sustainability of the resource. However, the reversal

of the decreasing trend in woody open and savanna shrubs has raised some questions

regarding the possible continuation of the observed trends in future. As subsistence

farming continues to play a dominant role in land cover conversion, degradation, from

v

evergreen Brachystegia woodlands to more open, dry vegetation, and to grassland

formations, will continue.

Considering the present scale of temporal and spatial change of the land cover in the

area, more continuous and comprehensive land cover change monitoring is required

with multi-spectral and multi-temporal satellite data merging. This study has provided

insights into the kind of landscape transformations that have taken place over 13

years, and will serve as input for the monitoring and proper utilisation of the Shire

River catchment for sustainable socio-economic development and water resources

management.

The land cover mapping derived from satellite images served as input for hydrological

modelling within the Shire River catchment. A GIS interface for SWAT, the ArcView

Soil and Water Assessment Tool eXtendable (AVSWATX) tool was used to model the

hydrology of the Shire catchment. Input variables for AVSWATX included digital

elevation data, soil and land cover grids, and weather data (daily rainfall, temperature,

relative humidity and wind speed). Available catchment streamflow data from 1977 to

1981 (5 years) were used for model calibration, while data from 1984 to 1985 were

used for model validation. The calibration was done at daily time-steps, for which

observed and modelled outputs were compared at Liwonde gauging station, the outlet

point of the catchment. Statistical evaluation of simulated catchment streamflows for

the calibration yielded Nash and Sutcliffe efficiencies (ENS) of 86% and 42% for the

monthly and daily predictions respectively. The Nash and Sutcliffe efficiency for the

validation period were considered acceptable, since the model was capable of

capturing 64% of the variance on monthly, and 42% on daily, observed records.

This validated simulation for 2002 land cover was used as a baseline for scenario

development of three scenarios: (i) continued land cover change at current trends

(business as usual); (ii) accelerated land cover degradation, associated with extensive

deforestation; and (iii) land cover restoration, reflection large-scale land restoration

and reforestation. Average annual and monthly modeled outputs from the alternative

scenarios were compared to the business as usual values to compute percent change in

annual values of surface flow, baseflow, and total channel discharge.

The condition of water resources in the Shire River catchment, Malawi, has been

affected adversely by rapid changes in land cover over the last two decades. A record

vi

of land cover changes that have taken place in the Shire River catchment has been

produced. The study has quantified the relationships that exist between land cover

changes and long-term changes in streamflow yield. A cost-effective set of techniques

has been demonstrated, combining satellite remote sensing for land cover mapping

and hydrological monitoring, which can be used in the formulation of policies for

sustainable land and water resources management in Malawi, and similar

environments elsewhere in Africa.

vii

Acknowledgements

I would like to express my genuine gratitude and appreciation to my thesis supervisor,

Professor Harold Annegarn of University of Johannesburg. His willingness to take me in

despite his busy schedule and not only coming in to scrutinize finished products but also

working through each chapter with me taught me a lot about thinking scientifically. Thanks to

Dr. Tobias Landmann from the University of Würzburg, Remote Sensing Chair, Germany,

who guided me in shaping my proposal into a workable project and for his continued support

during the PhD period. In spite of their tight schedule, they were able to provide guidance,

advice and help in this research project. I also really appreciate the confidence they showed in

me for launching into a research area that was completely new to me. I am indebted to your

understanding and kindness for helping me throughout this study, from beginning to

submission of this thesis.

I would also like to extend my gratitude to the University of Malawi, Chancellor College, for

allowing me study leave to do this PhD degree. In addition, my heartfelt gratitude goes to my

colleagues in the Department of Geography and Earth Sciences for their support.

My special thanks go to Mrs Melanie Kneen, Research Assistant at the University of

Johannesburg, for her major contribution to this research study. She was a great help

throughout, always providing good suggestions, advice, guidance and showing great patience.

I appreciate the journey she took me through TNTmips software applications and for

proofreading my drafts. I am also grateful to Dr P. M. Ndomba from the University of Dar es

Salaam, Tanzania for the very useful help concerning the use of AVSWATX.

I would like to thank Mr David Stevens, from United Nations Office for Outer Space Affairs

(UNOOSA), and US Geological Survey, for providing access to the Landsat TM and ETM+

images used for this study.

I would also like extend my sincere gratitude to Fatima Ferraz, a colleague who first

appreciated my work in remote sensing and water resources management. Through her

inspiration, I was able to participate in the European Space Agency “Tiger Africa Initiative”.

This initiative brought a lot of experience and inspiration in my experience as a beginner in

remote sensing and hydrology.

viii

Fieldwork would have been impossible without the support of Mr Jonathan Gwaligwali, GIS

Technician at the University of Malawi, Chancellor College, for gracefully enduring the

July/August heat associated with overhead sun and bushfires in Malawi.

I shared many good moments with fellow PhD students in the Department of Geography,

Environmental Management and Energy Studies who made my stay at the University of

Johannesburg a memorable period of my life. Thanks to the many colleagues I shared offices

with for the support and friendship: Dr. Patience Gwaze, Julião Cumbane, Matthew Ojelede,

Olusola Ololade, Joseph Kanyanga, Charles Ntui, Micky Josipovic, Philip Goyns and Charles

Paradzayi.

I wish to acknowledge the silent presence of my dad, Donald A. Chozenga for his love and

guidance in life. I wish he were still alive to share this memorable academic achievement. My

deepest gratitude and appreciation go to my dearest mum, Rennie Chozenga. I would like to

thank my lovely sisters Rhoda, Rose, Grace and brothers Dalitso and Justice for always caring

for me.

My deepest thanks go to my husband, Dr Martin Palamuleni, for his great comprehension,

love, and support during this study. To my daughters, Tadala and Tamanda, thank you so

much for enduring my absence.

Special thanks to the Malawian community studying in South Africa, Johannesburg who has

been a special pillar of strength. Thanks for all the encouragement and the prayers. My friends

who have been there for me, with even a number of you physically coming over to check on

me, thank you for your love.

I would also like to appreciate the help from Kate Pendlebury for the detailed editing and

proofreading on the draft thesis.

During my PhD study I received support from a number of people and organizations. I thank

Deutscher Akademischer Austausch Dienst (DAAD) through the African Network of

Scientific and Technological Institutions (ANSTI) for the PhD fellowship and the University

of Johannesburg for support during my stay at the University. Support was also given for

operating expenses and conference travel from the National Research Foundation through a

Focus Area Grant: FA2005040600018 “Sustainability Studies Using GIS and Remote

Sensing” to Prof H Annegarn. I am also grateful to the African Association of Remote

Sensing of the Environment (AARSE) and ITC, Netherlands, for a conference/workshop

fellowship to attend the 6th AARSE conference and the refresher course on Innovative

Applications of Remote Sensing and Geoinformation Sciences for female professionals in

Earth Sciences, in Cairo, Egypt.

Above all, I am grateful to the Lord God Almighty who is my source of strength and

wisdom. “I can do all things through Christ who strengthens me”.

ix

Contents

Declaration........................................................................................................... i

Dedication ........................................................................................................... ii

Abstract.............................................................................................................. iii

Acknowledgements ........................................................................................... vii

Contents ............................................................................................................. ix

List of Figures .................................................................................................... xi

Table of Abbreviations and Acronyms...............................................................xiv

1 Introduction 1

1.1 Water resources and sustainable development in Africa...............................1

1.2 Challenges within the Shire River catchment...............................................3

1.3 Aim and objectives......................................................................................6

1.4 Concepts and definitions..............................................................................9

1.5 Structure of thesis......................................................................................10

2 Land Cover Dynamics in the upper Shire River Catchment 11

2.1 Shire River catchment ...............................................................................11

2.1.1 Vegetation................................................................................................14

2.2 Overview of land cover mapping...............................................................17

2.2.1 Classification system ...............................................................................19 2.2.2 Earlier land cover mapping in Malawi.....................................................21

2.3 Methods ....................................................................................................23

2.3.1 Selection of satellite images.....................................................................24 2.3.2 Image processing .....................................................................................27 2.3.3 Image classification .................................................................................33 2.3.4 Land Cover Classification System...........................................................36

2.4 Results and discussions .............................................................................36

2.4.1 Transformation results .............................................................................37 2.4.2 Land cover maps ......................................................................................40 2.4.3 Description of land cover classes.............................................................43 2.4.4 Distribution of land cover categories .......................................................49 2.4.5 Thematic accuracy assessment.................................................................53

2.5 Conclusion ................................................................................................54

x

3 Land Cover Change assessment 1989-2002 58

3.1 Land cover change ....................................................................................58

3.1.1 Land cover change and hydrological response ........................................59

3.2 Land cover change detection .....................................................................62

3.3 Methodology .............................................................................................65

3.3.1 Input for change detection .......................................................................65 3.3.2 Approaches ..............................................................................................65

3.4 Results and discussion...............................................................................67

3.4.1 Image overlay...........................................................................................67 3.4.2 Post classification and land cover change areas.......................................70

3.5 Conclusion ................................................................................................84

4 Hydrological Modelling based on the Land Cover analysis 86

4.1 Introduction...............................................................................................86

4.1.1 Land cover and hydrological processes ...................................................86 4.1.2 Hydrological modelling approaches ........................................................87 4.1.3 Overview of the AVSWATX model ..........................................................90 4.1.4 Application of the AVSWATX model to the Shire River

catchment .................................................................................................96

4.2 Methodology .............................................................................................97

4.2.1 Data..........................................................................................................97 4.2.2 Model setup............................................................................................114 4.2.3 Modelling the Shire River catchment ....................................................118 4.2.4 Testing effects of land cover change......................................................120 4.2.5 Scenario generation................................................................................120

4.3 Results and discussion.............................................................................124

4.3.1 Hydrological characterization of Shire River catchment using

hydrological variables ............................................................................124 4.3.2 Modelling of the Shire River catchment ................................................126 4.3.3 Scenario outcomes .................................................................................142

4.4 Conclusion ..............................................................................................153

5 Conclusion and recommendations 155

5.1 Conclusion ..............................................................................................155

5.2 Recommendations ...................................................................................159

5.3 Concluding remarks ................................................................................163

References.......................................................................................................165

xi

List of Figures

Figure 1: Shire River hydrograph at Liwonde, 1948 to 2002 - water year beginning November each year ..................................................................................................5

Figure 2: Total annual rainfall from five rainfall gauging stations within the Shire

River catchment.........................................................................................................6

Figure 3: Location map of Shire River catchment, Malawi....................................................11

Figure 4: Population growth in three districts located within the study area..........................12

Figure 5: Location of the sample sites for primary data collection.........................................30

Figure 6: False colour images - 1989 and 2002 ......................................................................37

Figure 7: Principal component images - 2002 and 1989 ........................................................39

Figure 8: NDVI images - 1989 and 2002................................................................................40

Figure 9: Land cover maps - 1989 and 2002 ..........................................................................41

Figure 10: Woody closed ..........................................................................................................43

Figure 11: Woody open ............................................................................................................44

Figure 12: Savanna shrubs ........................................................................................................44

Figure 13: Grasslands ...............................................................................................................45

Figure 14: Marshy area .............................................................................................................45

Figure 15: Cultivated or grazing lands......................................................................................46

Figure 16: Built-up areas ..........................................................................................................46

Figure 17: Fresh water body .....................................................................................................47

Figure 18: Land cover extents - 1989 .......................................................................................48

Figure 19: Land cover extents - 2002 .......................................................................................49

Figure 20: Location map of forest reserves ..............................................................................50

Figure 21: Image overlay for the Shire River catchment: 1989 — 2002..................................68

Figure 22: Expansion of Mangochi Township into previously vegetated areas .......................69

Figure 23: Post classification change map of the Shire River catchment between 1989

and 2002 ..................................................................................................................71

Figure 24: Expansion of cultivated or grazing areas into predominantly savanna areas ..........74

Figure 25: Forest fragmentation around forest reserve areas....................................................75

Figure 26: Built-up areas expanding around Mangochi Township...........................................78

Figure 27: Increase in woody closed areas ...............................................................................79

Figure 28: Increase in built-up areas and tourism expansion around Liwonde National

Park..........................................................................................................................80

Figure 29: Overview of SWAT hydrological structure (adapted from Arnold et al.,

1998)........................................................................................................................91

Figure 30: Spatial distribution of soils within the Shire River catchment ..............................100

Figure 31: Rainfall variability in the Shire River catchment – first pattern (Data from

Department of Meteorology, Malawi)...................................................................103

xii

Figure 32: Rainfall variability in the Shire River catchment – second pattern (Data from the Department of Meteorology, Malawi) ....................................................104

Figure 33: Rainfall variability in the Shire River catchment – third pattern (Data from

the Department of Meteorology, Malawi).............................................................105

Figure 34: Weather stations and river gauging stations in the Shire River catchment ...........106

Figure 35: Time series streamflow for the Shire River Mangochi (inflow) and

Liwonde (outflow) for the period 1976 - 1981: data as received ..........................108

Figure 36: Streamflow data for 1977 - 1981, data as received ...............................................109

Figure 37: Smoothed daily streamflow data from 1977 - 1981 ..............................................111

Figure 38: Smoothed catchment streamflow data, 1977 - 1981..............................................112

Figure 39: Grid based discretisation and concept of flow path used in a cell.........................116

Figure 40: Sub-basins for the Shire River catchment .............................................................117

Figure 41: Time series plots of catchment streamflow and rainfall........................................125

Figure 42: Comparison of measured and simulated average annual water yield (mm)

by calibration and validation period......................................................................129

Figure 43: Comparison of monthly streamflows for calibration period, 1977 - 1981 ............129

Figure 44: Comparison of monthly catchment streamflows for validation period,

1984 - 1985............................................................................................................130

Figure 45: Comparison of daily catchment streamflows for calibration period 1977 - 1981......................................................................................................................131

Figure 46: Comparison of daily catchment streamflow for validation period: 1984 -

1985......................................................................................................................133

Figure 47: Simulated annual catchment streamflow for 1989 and 2002 land cover...............134

Figure 48: Rainfall variability between 1977 and 1981, referenced against long-term

mean (1976 – 2002)...............................................................................................135

Figure 49: Monthly mean, standard deviations and maxima of daily simulated

catchment streamflows for 1989 and 2002 land cover simulations.......................137

Figure 50: Comparison of simulated daily catchment streamflows for 1989 and 2002

land cover data.......................................................................................................138

Figure 51: Baseflow simulation results obtained from land degradation scenarios................144

Figure 52: Surface flow simulation results obtained from land degradation scenarios ..........145

Figure 53: Baseflow simulation results from land conservation scenarios.............................147

Figure 54: Surface flow simulation results from land conservation scenarios .......................148

Figure 55: Baseflow and surface flow simulation results from land conservation

scenarios ................................................................................................................148

Figure 56: Monthly mean, standard deviations and maxima of daily simulated

catchment streamflows for 2002 land cover scenario 7 and 8 simulations ...........150

Figure 57: Monthly mean, standard deviations and maxima of daily simulated

catchment streamflows for 2002 land cover, scenario 9 and 10 simulations ........151

Figure 58: Monthly mean, standard deviations and maxima of daily simulated

catchment streamflows for 2002 land cover, scenario 11 and 12 simulations ............................................................................................................152

xiii

List of Tables

Table 1: Refugees statistics by December 1992 ....................................................................13

Table 2: Description of Landsat 5 TM ..................................................................................26

Table 3: Description of Landsat ETM+.................................................................................26

Table 4: Correlation matrix for Landsat 7 ETM+ reflective bands .......................................38

Table 5: Correlation matrix for Landsat 5 reflective bands...................................................38

Table 6: Percentage of variance and correlation mapped to each principal

components in study area ........................................................................................38

Table 7: Land cover classes and their definitions..................................................................42

Table 8: Spatial distribution of land cover classes – 1989 and 2002.....................................48

Table 9: Error matrix for land cover classes..........................................................................55

Table 10: Assignment of Change Classification Codes of land cover for the Shire

River Catchment for 1989 and 2002 .......................................................................66

Table 11: Land cover changes of the Shire River catchment during 1989 to 2002.................72

Table 12: Areas changed into cultivated or grazing areas between 1989 and 2002................73

Table 13: Areas changed into grassland areas between 1989 and 2002..................................76

Table 14: Areas changed into savanna shrubs areas between 1989 and 2002.........................77

Table 15: Areas changed into built-up areas between 1989 and 2002 ....................................77

Table 16: Areas changed into woody open areas between 1989 and 2002 .............................81

Table 17: Areas changed into woody closed areas between 1989 and 2002...........................82

Table 18: Examples of large-scale hydrologic model applications .........................................88

Table 19: Data sets and sources for input into the AVSWATX model......................................97

Table 20: Spatial distribution of land cover classes and SWAT land cover class

codes for 1989 and 2002 .........................................................................................99

Table 21: Major soil types of the Shire River catchment and percent area covered .............101

Table 22: Soil parameters required by AVSWATX ................................................................101

Table 23: Weather stations and available data ......................................................................106

Table 24: Daily river flow data..............................................................................................107

Table 25: Characteristics of 2002 land cover data and deforestation scenarios ....................123

Table 26: Characteristics of simulated land cover forestation scenarios...............................124

Table 27: Relative sensitivity values of the optimised parameters........................................126

Table 28: Parameter values calibrated in SWAT using the auto-calibration tool..................127

Table 29: Average annual volumes obtained from calibration for 1977-1981......................128

Table 30: Parameters obtained from annual simulations for 1989 and 2002 land cover ......................................................................................................................134

Table 31: Parameters obtained from daily simulations for 1989 and 2002 land cover .........138

Table 32: Surface run-off simulated from 1989 and 2002 land cover...................................139

Table 33: Baseflow simulated from 1989 and 2002 land cover ............................................140

Table 34: Simulation results from bounding cases scenarios ................................................143

Table 35: Simulation results from land degradation scenarios..............................................144

Table 36: Simulation results obtained from land conservation scenarios .............................146

xiv

Table of Abbreviations and Acronyms AVHRR Advanced Very High Resolution Radiometer

AVSWATX ArcView Soil and Water Assessment Tool eXtendable

CGIAR Consultative Group on International Agricultural Research

DEM Digital Elevation Model

ENSO El Niño Southern Oscillation

FAO/LCCS Food and Agricultural Organisation/Land Cover Classification System

FAO/UNESCO Food and Agricultural Organisation/United Nations Education

Scientific and Cultural Organisation

GIS Geographical Information System

GLCN Global Land Cover Network

GLCF Global Land Cover Facility

GPS Global Positioning System

HRUs Hydrological Response Units

IDA International Development Association

IDP Integrated Development Planning

IGBP International Geosphere Biosphere Programme

IWRM Integrated Water Resource Management

Landsat TM Land satellite Thematic Mapper

Landsat ETM+ Land satellite Enhanced Thematic Mapper Plus

LH-OAT Latin Hypercube One-Factor-At-a-Time

MDGs Millennium Development Goals

MSG METEOSAT Second Generation

MODIS Moderate Resolution Imaging Spectroradiometer

NDBI Normalised Difference Built-up Index

NDVI Normalized Difference Vegetation Index

NOAA National Oceanic and Atmospheric Administration

PR Precipitation Radar

SADC Southern African Development Community

SAR Synthetic Aperture Radar

SCE-UA Shuffled Complex Evolution - University of Arizona

SEVIRI Spinning Enhanced Visible and Infrared Imager

SPOT Satellite Pour l'Observation de la Terre

SWAT Soil and Water Assessment Tool

SWRRB Simulator for Water Resources in Rural Basins

TRMM Tropical Rainfall Measuring Mission

USDA-ARS United States Department of Agriculture - Agricultural Research Service

UNEP United Nations Environmental Programme

UN/FAO United Nations, Food and Agricultural Organization

USDA-SCS United States Department of Agriculture - Soil Conservation Service

USGS United States Geological Survey

UTM Universal Transverse Mercator

WSSD World Summit of Sustainable Development

1

Chapter 1

1 INTRODUCTION

Chapter 1 presents the framework for integrated water resources management while

conceptualising the challenges of water resources in Africa in an era of growing

population and climate change. This research addresses these challenging needs within

the upper Shire River catchment in Malawi. Fundamental issues relating to land cover

change and land surface hydrological response have been summarised. Within this

context, the research hypothesis and objectives are articulated. The chapter ends with

an outline of the thesis structure.

1.1 Water resources and sustainable development in Africa

Water resources are inextricably linked with climate change, population growth and

rainfall variability, so the prospect of global climate change has serious implications for

water resources and regional development in Africa [IPCC, 2001]. Efforts to provide

adequate water resources for Africa will confront several challenges, including population

pressure; problems associated with land use, such as erosion/siltation; and possible

ecological consequences of land-use change on the hydrological cycle [Riebsame et al.,

1994]. Assessment of management issues relating to the distribution and use of water

resources requires an integrated approach. Integrated water resources management is not

only required for analyzing consequences of the adverse natural conditions such as floods

or drought but also to assess possible strategies to make the area less vulnerable to

environmental constraints and changing climate [Tolba, 1982]. In addition, the notion of

water resources management requires the matching of water availability and water use in a

river basin [Terpstra and van Mazijk, 2001]. River basins are the preferred land surface

units for water-related regional scale studies because their drainage areas represent natural

spatial integrators or accumulators of water and associated material transports and thus

allow for the investigation of cumulative effects of human activities on the environment

[Lahmer et al., 2001]. This is the system of integrated river basin management endorsed in

Agenda 21 of Rio, 1992 and echoed at the World Summit of Sustainable Development

(WSSD), Johannesburg, South Africa in 2002.

Water availability is generally a consequence of precipitation and catchment run-off. An

aspect often forgotten in this respect is the impact of new patterns of land use and land

cover on the hydrological availability of water. There are many connections between land

surface characteristics and the water cycle. Firstly, land cover can affect both the degree of

2

infiltration and run-off following precipitation events. Secondly, the degree of vegetation

cover and the albedo (degree of absorption/reflection of sun's rays) of the surface can affect

rates of evaporation, humidity levels and cloud formation [Newson, 1992]. Any change in

land use and land cover will have correlated effects in the hydrological regimes, and

possible impacts on the habitat and ecological communities [Calder, 1992; Lorup et al.,

1998]. In essence, the degree and type of land cover influences the initiation of surface

run-off, the rate of infiltration and consequently the rate of ground water recharge [Calder,

1992].

Recent studies in the Southern African Development Community [SADC, 1995] region

have revealed that climatic and land cover changes threaten to undermine the integrity of

riverine habitats, the availability and quality of water, and agricultural productivity

[Headstreams Project, 2004]. Moreover, there are several indicators of water stress and

scarcity in the SADC region, including the amount of water available per person and the

volume ratio of water withdrawn and potentially available [IPCC, 2001]. This situation has

been attributed to increasing population, which translates into increased demand for water

supply (for domestic, agricultural and industrial use), as well as to climate change. Global

warming would induce changes in precipitation and wind patterns, changes in the

frequency and intensity of storms, ecosystem stress and species loss, reduced availability

of fresh water, and a rising global mean sea level [Ominde and Juma, 1991]. Although the

impacts may not be easily predicted, changes in weather patterns may lead to the

prevalence of severe drought conditions or extreme flood events in the SADC region. The

existence of prolonged drought periods vis-à-vis water scarcity will seriously affect

agricultural production and the socio-economic activities in the region. Malawi is one of

the southern African countries likely to experience absolute water scarcity by 2025 [SADC,

1995], which is a challenge for water resources management to sustain economic

development in the country.

To balance supply and demand for water resources and to reduce negative or undesired

effects for the environment and society, changes of actual land cover have to be studied at

all spatial scales. The land surface provides a critical role in the water cycle as it is the

level at which precipitation is redistributed into evaporation, run-off or soil moisture

storage [Verburg et al., 1999]. Thus, land use and land cover studies should be viewed as

responding to the complex interactions and feedbacks linking social and biophysical

3

processes that occur on the land [Dolman and Verhagen, 2003; Maidment, 1993]. With

increasing human activities vis-à-vis water conflicts, it is important to understand the

interactions between hydrological regimes and associated land use and land cover changes

in catchments [Rockström et al., 2002]. Such an understanding can be achieved by

integrating land use planning and water resources management. Land cover and land cover

change data represents a key variable in the management and understanding of the

environment, as well as driving many environmental models such as hydrological models

within large river basins or even for particular smaller catchments. Therefore, there is a

need to develop proper planning and management approaches within the context of

Integrated Water Resource Management (IWRM). IWRM as defined by the Global Water

Partnership [Global Water Partnership, 2005] is a process that considers the co-ordination

of development and management of water, land and related resources to enhance economic

and social welfare without jeopardising the sustainability of the ecosystem. Thus,

sustainable development of water resources is a key to the maintenance of the natural

ecosystem that supports the well-being of human populations.

1.2 Challenges within the Shire River catchment

The Shire River system, the only outlet of Lake Malawi, is probably the most important

water resource for Malawi. Hydro-electric power plants of about 200 MW generation

output, based on a firm flow of 170 m³ s-1

, have been developed on Shire River providing

98% of electricity produced and used in Malawi [Malawi Government, 2001]. This

electricity is the primary source driving the economic and industrial infrastructure and

services in the country. An estimated 20-25 m³ s-1

of water is abstracted for irrigation in the

Lower Shire valley and government sponsored smallholder schemes. Blantyre City

abstracts 1 m³ s-1

of water for both domestic and industrial use. The Shire River has also

led to the development of fisheries, water-transport and tourism industries. This translates

into an increased demand for water for diverse needs and values. When its supply is

limited in quantity or quality or its distribution is uneven, water can be a source of both

cooperation and contestation among its different users [Mulwafu et al., 2003].

Over the last three decades, the Shire River catchment has undergone considerable changes

in the structure and composition of land use and land cover [Malawi Government, 1998b].

The major driving forces are related to human population increases and rainfall variability.

The national population growth rate has been increasing from 2.7% during the 1977 census

4

to 3.2% in 1998 and is likely to double in the next twenty years [National Statistical Office,

2000]. The population density in Malawi is high – the national average density was 87

people km-² and 171 people km

-² of arable land during the 1977 and 1987 censuses

respectively [National Statistical Office, 1991]. Population density in settlements within

the Shire River catchment was recorded at over 275 people km-² during the 1998 census

[National Statistical Office, 2000].

The high population growth has translated into rapidly increasing demands from land in

terms of food, shelter, energy (fuelwood) and construction materials. Some of the

woodlands are now replaced by agricultural crops, while the grass-covered dambos have

been either overgrazed or cultivated and are left bare. Swamp vegetation has been drained

and cultivated. Studies done in 1967 estimated the woodland cover to be 74% while in

1990 the cover was estimated at 61% [Green and Nanthambwe, 1992]. These observations

suggest that woodland cover has declined by 13% between 1967 and 1990 and between

1981 and 1992 Mwanza district alone experienced 1.8% average annual deforestation rate

[Hudak and Wessman, 2000]. Much of the deforestation has been linked to conversion of

communally owned miombo woodlands into agricultural land [Desanker et al., 1997;

Place and Otsuka, 2001], while high wood demands for energy has exacerbated the

situation. Although land for agricultural production is limited to only 37% of the land area

under rain-fed cultivation at traditional management level, as much as 48% of the land was

found to be under cultivation by 1989/90 growing season [Green and Nanthambwe, 1992].

Inappropriate agricultural practices including overgrazing, mono-cropping, cultivation on

steep slopes and river banks and other marginal areas have degraded land through soil

erosion, reduced water retention and the loss of soil nutrients. Aggravating this situation is

the subsequent decrease in land holding sizes, estimated at 0.5 ha per household [Malawi

Government, 1998b].

Consequently, processes of the land hydrology such as run-off, infiltration,

evapotranspiration and interception have been modified. In most cases, this has resulted in

increased run-off, accelerated soil loss with sedimentation problems leading to reduction of

baseflows and increased incidences of flood disasters during heavy storms [Malawi

Government, 1998b]. In addition, hydropower supplies are threatened by low water flows

and sedimentation, hence power disruptions occur frequently especially in dry years

[Kaluwa et al., 1997]. Aggravating the situation is an increase in demand for water, by

5

different groups with diverse needs and values. Furthermore, it is important to note that the

flow from Lake Malawi into the Shire discontinued for a period of 22 years from 1915 to

1937 [Kidd, 1983] and almost dried up in 1997 [Malawi Government, 2001]. On the one

hand, it is hypothesised that the lack of outflow was due to the vegetation growth and

piling of sediments from the small tributaries near the source, while on the other hand, low

rainfall in the catchment area during the period prior to 1937 is said to be responsible for

the lowering of the lake levels [Sheila, 1995]. However, it is unlikely that sedimentation

would have affected the 1915-1937 occurrences due to low population and associated

agricultural activities and deforestation. A significant decline in the flow of Shire River has

been observed since 1992 (Figure 1) with mean flows as low as 130 m³ s-1

in 1997

compared to 825 m³ s-1

in 1980 and 634 m³ s-1

in 1990 [Malawi Government, 2001].

0

100

200

300

400

500

600

700

800

900

1948/1949

1951/1952

1954/1955

1957/1958

1960/1961

1963/1964

1966/1967

1969/1970

1972/1973

1975/1976

1978/1979

1981/1982

1984/1985

1987/1988

1990/1991

1993/1994

1996/1997

1999/2000

Streamflow (m3 s-1)

Figure 1: Shire River hydrograph at Liwonde, 1948 to 2002 - water year beginning November each year

In addition, recent droughts in southern Africa have been associated with the drop in river

flows [SADC, 1995]. In the last three decades, Malawi has experienced variability and

unpredictability in seasonal rainfall. There have been three significant droughts (in

6

1978/79, 1981/82, and most severe was in the 1991/92 season), frequent and increasingly

long dry spells, and an erratic onset and cessation of rainfall [Malawi Government, 2001].

Further discussion on rainfall variability within the Shire River catchment is in section

4.2.1 of this thesis. Rainfall data collected from gauges within the catchment (from 1977 to

1981) is plotted in Figure 2. This is the period that has been utilised for the hydrological

modelling in this study.

1977 1978 1979 1980 1981

500

1000

1500

2000

2500

Total annual rainfall (mm)

Ntaja

Salima

Mangochi

Balaka

Chancellor college

Figure 2: Total annual rainfall from five rainfall gauging stations within the Shire River catchment

Unprecedented rainfall variability means unexpected droughts or flooding which may in

turn produce changes in land use and land cover [Meyer and Turner, 1994]. Precipitation

variability, water scarcity and changes to the water regimes through land cover change will

seriously affect agricultural production and the socio-economic activities in the country. In

addition, such variations influence the temporal phenology and chlorophyll characteristics

of the vegetation in the area, which is a challenge in remote sensing studies.

1.3 Research question, Hypothesis, Aim and Objectives

Research question

This study is intended to address the following question: What are the effects of significant

land cover changes over the past two decades on river flow characteristics that are

important for water resources, environmental functioning and hydrological processes

7

within the upper Shire River catchment? Land cover changes associated with growing

human populations and expected changes in climatic conditions are likely to accelerate

alterations in hydrological phenomena and processes on various scales. Subsequently,

these changes could significantly influence the quantity and quality of water resources for

both nature and human society. This aim will be pursued in the context of developing

integrated land use planning and water resources management in Malawi.

Hypothesis

The following general hypothesis is proposed for the Shire River catchment in Malawi:

Unsustainable changes in land cover due to human activities are significantly

altering aggregate catchment conditions, giving rise to long-term, potentially

irreversible changes in river flow characteristics.

Aim and objectives

This research hypothesis will be tested through a structured sequence of land cover change

analyses and hydrological model simulations. Accordingly, the objectives are set out as:

• To map land cover within the upper Shire River catchment for 1989 and

2002, using Landsat TM and ETM satellite imagery;

• To quantify land cover changes in the catchment between 1989 and 2002;

• To model the hydrological regimes in the upper Shire River catchment and

its sub-basins;

• To challenge the thesis hypothesis by using the hydrological model to

evaluate effects of derived quantitative land cover changes on hydrological

processes;

• To simulate likely changes to hydrological processes in response to

continued land cover changes; and

• To discuss the implications of land use management on stabilising water

regimes of the Shire River catchment.

1.4 Research Design

Land cover mapping and change detection will be based on analyses of two Landsat

images captured 13 years apart. Supplementary digital mapping data sets were obtained

from the Department of Surveys in Malawi. To map land cover dynamics, pixel based

8

classification was undertaken using Maximum Likelihood algorithm. Accuracy assessment

was carried out using producer and user accuracies for each class along with overall

accuracies [Congalton and Green, 1999]. The UN Food and Agricultural Organisation

Land Cover Classification System [Food and Agriculture Organisation, 2005] was used to

label land cover variables to achieve legend harmonisation within Africa and on a global

scale. The new classification is also internally consistent, allowing for scalability and

flexibility that can be used at different scales and different levels of detail to distinguish

land cover features.

Two approaches were utilised to detect and compare changes in the upper Shire River

catchment, namely multi-date visual composite and post-classification analysis. In the

multi-date approach, vegetation reduction was chosen as the main indicator of land cover

change. This technique is not meant to be quantitative, but rather was used to identify and

explore areas of change. Using a post-classification approach, Landsat TM and Landsat

ETM+ images were classified and labelled individually. Later, classification results were

compared on a pixel-by-pixel basis using a change detection matrix where areas of change

were extracted. Quantitative statistics were compiled to determine specific changes

between the two images i.e. magnitude and direction of change in each land cover type

[Calder, 2002].

Hydrological responses were tackled using an existing physically based hydrological

model, the Soil and Water Assessment Tool (SWAT) [Arnold et al., 1994]. This model

incorporates key features of catchment properties, including links between land cover

hydrologic responses. A Geographical Information System (GIS) interface for SWAT, the

ArcView Soil and Water Assessment Tool eXtendable (AVSWATX) tool was used to prepare

parameter input values for the Shire catchment. Input variables for AVSWATX included

digital elevation data, soil and land cover grids and weather data (daily rainfall,

temperature, relative humidity and wind speed). Five rainfall gauges were used to provide

input daily rainfall data to AVSWATX. Available catchment streamflow data from 1977 to

1981 (5 years) was used for model calibration, while data from 1984 to 1985 was used for

model validation. The calibration was done at a daily time-step where observed and

measured outputs were compared at the same outlet point on the catchment, Liwonde

gauging station. Model runs were validated using the parameterised 2002 land cover data.

This validated simulation for land cover in 2002 was used as a baseline for scenario

9

development of three scenarios: (i) continued land cover change at current trends; (ii)

accelerated land cover change associated with extensive deforestation; (iii) reduced land

cover change due to management and reforestation. Average annual outputs from three

alternative futures were then differenced from the baseline values to compute percent

change in annual values of surface flow, baseflow, and total channel discharge.

1.5 Concepts and definitions

This section describes some of the basic terminologies used in land use and land cover

research. The definitions are fundamental to fully understand and apply research results to

a broader readership. The definitions are based mainly from FAO [2005].

Land is any delineable area of the Earths’ terrestrial surface involving all attributes of the

biosphere immediately above and below this surface. It encompasses the near-surface

climate, soils and terrain, surface hydrology and human settlements patterns and physical

results of human activities. Land can be considered in two domains: (i) land in its natural

condition, (ii) land that has been modified by human beings to suit a particular use or a

range of uses.

Land use is the manner in which human beings utilise the land and its resources. Examples

of land use include agriculture, urban development, grazing, logging, and mining.

Land cover describes the physical state of the land surface. Land cover categories include

cropland, forests, wetlands, pasture, roads, and urban areas. Land cover is taken to mean a

physical description of space, of the observed (bio)-physical cover of the Earths’ surface. It

indicates what covers the land such as forest, bushes, uncultivated areas and water bodies.

Land cover classification is the process of defining land cover and land use classes based

on well-defined diagnostic criteria. A classification describes the systematic framework

with the names of the classes and the criteria used to distinguish them, and the relationship

between classes. Such information is taken from ground surveys or through remote

sensing.

Land cover change can be categorised into two types: modification and conversion. Land

cover modifications entail the changes that affect the character of the land without

changing its overall classification and can either be human induced, for example, tree

removal for logging; or have natural origins resulting from, for example, flooding, drought

10

and disease epidemics. Land cover conversion is the complete replacement of one cover

type by another such as deforestation to create cropland or pasture.

1.6 Structure of thesis

This thesis is divided into five chapters. Chapter 1 comprise of the general introduction,

which also outlines the research hypothesis and the objectives. Chapter 2 is a discussion of

land cover dynamics within the upper Shire River catchment based on a supervised

Maximum Likelihood classification of Landsat 5 TM (1989) and Landsat 7 ETM+ (2002).

The variability in spatial land cover extents for each classified land cover class between the

two periods has been examined. The results are used in chapter 3 and 4. Chapter 3 is an

examination of land cover changes within the upper Shire River catchment. The adopted

change detection methods quantitatively reveal the major changes that have occurred in the

catchment between 1989 and 2002. Chapter 4 integrates the work of the previous chapters

(2 and 3) into the preparation of parameters for the Soil Water Assessment Tool (SWAT)

hydrological model. This chapter presents the calibration, validation, and application of

SWAT model for predicting the hydrological response from land cover activities within the

catchment. Critical land cover change simulations demonstrate the capability of the model

in guiding spatially distributed land cover change and precipitation events. The final

Chapter involves the critical discussion of the main research findings and recommends

future investigations to advance the field of physically based hydrological modelling for

the management of water resources.

11

Chapter 2

2 LAND COVER DYNAMICS IN THE UPPER SHIRE RIVER

CATCHMENT

Chapter 2 provides a description of the study area, including topographical, climatic

and hydrological characteristics. An overview of land cover mapping and concepts

with regard to application of satellite data are discussed. This is followed by analyses

of land cover dynamics within the upper Shire River catchment based on a supervised

Maximum Likelihood classification of two images captured 13 years apart:

Landsat 5 TM (1989) and Landsat 7 ETM+ (2002). Differences in spatial land cover extents for each classified land cover class between the two times are examined.

2.1 Shire River catchment

Location

The Shire River catchment lies in the southern part of the Great East African Rift Valley

system and is the outlet of Lake Malawi. The river flows approximately 400 km from

Mangochi on the southern extremity of Lake Malawi, to Ziu Ziu in Mozambique at the

confluence with the Zambezi River (Figure 3). The catchment area of the basin is

18,000 km2 and is divided into upper, middle and lower sections.

Salima

Chancellor College

Balaka

Mangochi

Ntaja

Liwonde

Lake

Malawi

Lake

Malombe

Lilongwe

Zomba

Mangochi

Liwonde

Catchment boundary

Country boundary

Lilongwe

Zomba

Mangochi

Liwonde

Lilongwe

Zomba

Mangochi

Liwonde

Catchment boundary

Country boundary

Catchment boundary

Country boundary

Salima

Chancellor College

Balaka

Mangochi

Ntaja

Liwonde

Lake

Malawi

Lake

Malombe

Salima

Chancellor College

Balaka

Mangochi

Ntaja

Liwonde

Lake

Malawi

Lake

Malombe

Salima

Chancellor College

Balaka

Mangochi

Ntaja

Liwonde

Salima

Chancellor College

Balaka

Mangochi

Ntaja

Liwonde

Lake

Malawi

Lake

Malombe

Lilongwe

Zomba

Mangochi

Liwonde

Catchment boundary

Country boundary

Lilongwe

Zomba

Mangochi

Liwonde

Lilongwe

Zomba

Mangochi

Liwonde

Catchment boundary

Country boundary

Catchment boundary

Country boundary

Figure 3: Location map of Shire River catchment, Malawi

12

The upper Shire River catchment is between Mangochi and Matope, with a total channel

bed drop of about 15 m over a distance of 130 km. The focus of this study is the uppermost

reach from Mangochi to Liwonde, which is almost flat at 465 – 600 m above mean sea

level over a distance of 87 km. It forms a catchment area of 4,500 km2, located between

latitudes 14° 20' S; 15° 12' S and longitudes 34° 59' E; 35° 30' E. The river flows through

Lake Malombe, which is 1.8 m below Lake Malawi.

Population

According to administrative boundaries, the study area is located within three districts

namely: Mangochi, Machinga and Balaka. The population within these three districts has

increased from 644 177 in 1977 to 1 011 843 in 1987 and 1 218 177 in 1998 [National

Statistical Office, 2000]. The southern region of Malawi, which forms the catchment area

of the Shire River, has the highest population density ranging between 53 and 275

people km-² of arable land, varying from district to district [National Statistical Office,

2000]. Population increased by 46% from 1977 to 1998 as depicted in Figure 4. Given that

a high proportion of the population is in subsistence agriculture, an increase of population

has serious implication for the degradation of the environment.

0

100

200

300

400

500

600

700

1977

1987

1998

Population ('000)

Mangochi Machinga Balaka

Figure 4: Population growth in three districts located within the study area

13

The increase in population between the 1977 and 1987 census is not only due to natural

population growth, but also due to the influx of refugees from Mozambique. Between 1976

and 1992, Mozambique was ravaged by civil wars such that many of its neighbouring

countries, including Malawi, became home to refugees. The 1992 refugee statistics for the

district within the study area are shown in Table 1.

Table 1: Refugees statistics by December 1992

District Malawians Mozambican Refugees

Mangochi 551 190 46 973

Machinga 577 860 33 300

Balaka 400 057 126 869

Source: Office of the President and Cabinet December 1992.

Rainfall

The dominant climate in Malawi is tropical savanna with distinct dry and wet seasons.

Rainfall is governed by the movement of the Inter-Tropical Convergence Zone (ITCZ) and

other belts of distribution. The onset of rain is not usually predictable but falls between

October/November and ends in April/May of the following year. Almost 90% of rainfall

occurs between December and March and most of the country receives 800 –

1,200 mm a-1

, with some exceptions [Hutcheson, 1998]. Rainfall statistics from stations

within the Shire River catchment receives an average rainfall of 996 mm a-1

, but there are

variations in the amount of rain, its onset, duration and intensity during the wet season.

Further discussion on rainfall variability within the Shire River catchment may be found

section 4.2.1 of this thesis.

Temperature

Temperatures vary with altitude. During the cold season from May to August, the

highlands have mean temperatures of 15 – 18°C and the low-lying rift valley of 20 – 24°C.

The margins of the river have long hot seasons and high humidity, with mean daily

temperatures ranging from 26°C in January to 21°C in July. During the hot season from

September to January, the highest temperatures are recorded in the Shire Valley and along

the lakeshore with the daily average maximum reaching approximately 32°C in October.

The lowest temperatures (26°C) are recorded over high altitude areas particularly the Shire

Highlands.

14

2.1.1 Vegetation

The natural vegetation in Malawi is part of the extensive dry forest Miombo woodland eco-

region, covering most of the southern and eastern parts of Africa [Abbot et al., 1995;

Desanker et al., 1997]. The Miombo woodland classification is characterised by mixed

deciduous woodlands, with dominant species from the family Caesalpinacea –

Brachystegia, Julbernardia and Isoberlinia. The vegetation types vary depending on

altitude, rainfall pattern, soil types and locations, and range from lowland rain forests,

mopane and sub-mopane woodlands, dry evergreen forests, wooded grasslands, wooded

farmlands, and swamp forests [Moyo et al., 1993].

The dominant vegetation in the study area is mopane woodland, which varies in density

from tall open woodland to dense scrub. Mopane woodlands may form pure stands

excluding other species, but are generally associated with several other prominent trees and

shrubs such as Kirkia acuminata, Dalbergia melanoxylon, Adansonia digitata, Combretum

apiculatum, C. imberbe, Acacia nigrescens, Cissus cornifolia, and Commiphora spp. The

herbaceous component of mopane communities differs according to soil conditions and

vegetation structure: dense swards are found beneath gaps in the mopane canopy on

favourable soils, while grasses are almost completely absent in shrubby mopane

communities on mopanosols [Low and Rebelo, 1996; Smith, 1998; White, 1983]. Within

the Shire highlands and the slopes along the catchment, lowland forest and Brachystegia

woodland are found.

Grasses include large tussocks of Festuca costata and Maximuella davyi interspersed with

cushions of Eragrostis volkensii and the Alloeochaete oreogena. Tall grasses are associated

with low altitude woodland, including Hyparrhenia gazensis, Hyparrhenia variabilis,

Hyparrhenia dichroa, Andropogon gayanus, Setaria palustris and Panicum maximum. In

densely settled and cultivated locations, tall reedy grasses are replaced by Urochloa

pullulans and Urochloa mosambicensis. Woodlands are characterised by Sterculia

africana, Colophospermum mopane, Acacia tortilis and Faidherbia albida according to

locality. Acacia woodland provides valuable grazing from pods to supplement grasses in

the dry season. Mature trees may stand within a dense understory, which includes

Commiphora spp., Bauhinia tomentosa, and Popowia obovata. The understory is likely to

be man-induced since lone-standing mature trees are found in open areas of cultivated

land, and in some cases trees are selectively retained by farmers to maturity (e.g.

Faidherbia albida). Base rich soils support Euphorbia ingens and Commiphora thicket,

15

whilst Hyphaene ventricosa, Hyphaene crinita and Borassus aethiopium palms occur

where the water table is high [Low and Rebelo, 1996; Smith, 1998; White, 1983].

Miombo woodlands form an integral part of the livelihood and farming systems of southern

Africa including the Shire catchment [Frost, 1996]. For most rural communities, the

woodlands are a primary source of energy in the form of fuelwood and charcoal and a

crucial source of essential subsistence goods [Dewees, 1994; Morris, 1995]. Households

rely on woodlands to supplement their food supply through the collection of wild food

plants, bushmeat, nuts, leaves and roots. Woodlands are also a source of income through

the sale of non-wood forest products such as mushrooms. Certain tree species are vital to

communities for their use as sources of traditional medicines. In commercially managed

Miombo forests, timber is a valuable product. People living in towns and cities throughout

the Miombo eco-region also depend on food, fibre, fuelwood and charcoal from miombo

woodlands [Bradley and McNamara, 1993; Dewees, 1994]. In addition, woodlands have

an ecological role in controlling soil erosion, providing shade, modifying hydrological

cycles and maintaining soil fertility [Desanker et al., 1997].

However, in recent years Miombo has been facing increasing pressure due to human

population expansion and intensive use. A large proportion of this eco-region has been

completely transformed. The various causes of deforestation include agricultural

expansion, shifting cultivation, overexploitation for fuelwood and poles, overgrazing,

excessive burning and a broad-spectrum of urban and industrial development. Although

habitat is fairly well conserved in protected areas, even national parks are affected by

people who increasingly encroach upon protected land to search for fuelwood or new

grazing and farming areas [Abbot et al., 1995]. Anthropogenic alterations together with

natural variations have transformed Miombo into open forests, thicket and grassland

formations [Frost, 1996].

Miombo woodlands present a number of cover types with similar physiognomies and their

land cover classes are usually heterogeneous at 1 km spatial resolution [Sedano et al.,

2005]. As a subtropical ecosystem, Miombo woodlands are characterised by a distinct dry

season. During this dry season, the vegetation experiences important phenological changes.

The magnitude and pace of these changes varies for every type of vegetation depending on

their capacity to reach water resources [Frost, 1996]. The characteristic phenology of every

land cover can be used in multi-temporal remote sensing approaches for land surface

16

classification. Conversely, the classification of such landscapes using remote sensing data

is likely to be affected by the reflectance patterns of different vegetation types that vary

during the dry season.

In addition, where woodlands are disappearing, landscapes are dominated by medium to

tall grasslands with forest relics and isolated stands of shrub-lands. In close association

with the woodlands, altered landscapes include village settlements (both clustered and

scattered) with grass-thatched roofing, mud walls and occasionally iron sheet roofing.

Consequently, the dormant living plants, natural leaf and grass litter, and human utilisation

of biomass products in structures exhibit intergraded classes that are internally

heterogeneous. Spectral responses recorded by remote sensing reveal similar spectral

responses that may relate to different classes. The mixture of spectral signal made up of

grass, leaf and twig litter, and bare soil constitutes a challenge for spectral image analysis

in the mapping of such landscapes. The ambiguities of land cover composition lead to land

cover classification errors.

Soils

The upper Shire River catchment is associated with the four main classes of soils: latosols,

lithosols, calcimorphic and hydromorphic soils [Malawi Government, 1998b].

Latosols and ferrosols are red-yellow soils, which include the ferruginous soils in the

upland areas of the catchment and are among the best agricultural soils in the country.

Ferralsols, both rhodic and orthic, cover large parts of the plains along the western border

of the catchment. Lithosols form shallow stony soils: they are immature soils originating

from sand. These occur in all areas of broken relief that are associated with steep slopes.

Calcimorphic soils are grey to greyish brown and occur on nearly level depositional plains

with imperfect drainage. This soil group includes alluvial soils of the lacustrine and

riverine plains: vertisols and mopanosols. The mopanosols are dominant in the upper

catchment. Hydromorphic soils are black, grey or mottled and are found in either

seasonally or permanently wet areas locally called dambos [Moyo et al., 1993]. These soils

are dominant in the valley below Lake Malombe.

Most of the soils in the Shire rift valley are of alluvial origin, rich in nutrients and ideal for

agricultural production. On the escarpment (slopes and plateaus), the soils are heavily

leached and of medium fertility. In hilly places the soils are shallow, and such areas are

17

used as catchments and for the protection of indigenous fauna and flora. The variability of

soil background reflection (per pixel) in satellite remote sensing can be a problem in

mapping vegetation in African savannas [Landmann, 2003]. In areas that are seasonally

dry, spectral differences are consistent with soil differences, senescent deciduous

woodland, dry grasses, rural settlements and bare fields. The presence of identifiable

chlorophyll in the vegetation can be used to differentiate dry grasses and soils. Hence, in

this study, mapping land cover with contextual data on soils and digital elevation was used

to minimise noise outliers.

2.2 Overview of land cover mapping

Identifying, delineating and mapping land cover is important for resource management and

planning programs. The review of land cover mapping indicates the significance of the

periodic determination of land cover distribution over an area of interest for scientific

research, resource management and planning and policy purposes [Cihlar, 2000]. Land

cover mapping is also considered an essential element for modelling the Earth as a system.

Environmental planning and management depends upon information concerning land

cover. This implies that sustainable livelihoods and food security depend on the effective

management of land resources. Hence, land cover classification forms a reference base for

resource managers in their decision-making processes to guide rural and urban growth,

through their decision-making, to determine changes to natural resources, and to develop

spatial trend analyses.

Increases in the Earth’s population have put more stress on the land to support the growing

human needs. Demands for agricultural land as well as economic initiatives have led to

changes in land cover. Globally, land cover has been altered by direct human use through

agriculture, forest harvesting and urban/suburban development. However, land cover is

also altered by non-anthropogenic forces. Natural events such as weather, flooding, fire,

climate fluctuations and ecosystem dynamics may also initiate changes in land cover.

There are also secondary indirect impacts on land cover from other human activities: for

example: forests and lakes may be damaged by acid rain from fossil fuel combustion, and

crops and natural vegetation near cities may be affected by elevated tropospheric ozone

resulting from automobile exhaust [Meyer, 1995].

18

Changes in land cover have resulted in changes in geochemical cycles, hydrological cycles

and atmospheric chemistry [Hunt, 2004; Lambin, 2006; Newson, 1992]. For example, land

cover can affect both the degree of infiltration and run-off following precipitation events.

The degree of vegetation cover and the albedo of the surface can affect rates of

evaporation, humidity levels and cloud formation [Calder, 1992]. Any change in land

cover will have effects on the hydrological regimes and possible impacts on habitat and

ecological communities [Calder, 1992; Lorup et al., 1998]. To balance supply and demand

for water resources and to reduce the negative or undesired effects upon the environment

and society, the characteristics of actual land cover should be studied at several spatial

scales.

Several important considerations determine the characteristics of any prospective land

cover classification. These may include purpose, scale, data and algorithms employed

[Geist and Lambin, 2002]. For example, specific models of vegetation-atmosphere

interactions require different types of land cover information [Sellers et al., 1996].

Similarly, productivity models [Liu and Smedt, 2004], hydrological models [Lahmer et al.,

2001], land use or land cover inventories and planning as well as other bio-physical

resource documentation require different forms of land cover information [Green et al.,

1994; Lambin and Strahler, 1994]. The thematic content also has strong influences on the

chosen frequency and type of land cover mapping. For instance, information may be

needed concerning few cover types (e.g. urban and non-urban areas), or for several, finely

distinguished cover types.

In addition, land cover information may be required locally, at a regional level or at

continental to global scales [Cihlar and Jansen, 2001]. Recent developments through

remote sensing have shown that large areas can be mapped using high resolution satellite

imagery. Time-series satellite images over the last three decades are used to derive the

location, extent and rates of land cover dynamics [Desanker et al., 1997] at synoptic scales

and facilitates the discerning of large scale ecosystem patterns [Roughgarden et al., 1991].

There are numerous satellite remote sensing sources which allow objective analyses of

ecological variables [Dauze et al., 2001]. Some of the currently available sources of

satellite remote sensing images include: the Moderate Resolution Imaging

Spectroradiometer (MODIS) for moderate spatial resolutions, which has pixel sizes

between 250 m and 1 km depending on the channel [Justice et al., 2002] and Advanced

19

Very High Resolution Radiometer (AVHRR), which has a pixel size of approximately

1 km [Loveland et al., 2000]. Finer spatial resolution satellites include: Land satellite

(Landsat), with a pixel size of 30 m; and Satellite Pour l'Observation de la Terre (SPOT),

with pixel sizes between 10 m and 20 m [Cihlar, 2000]. Very fine spatial resolution refers

to a new generation of satellite, such as IKONOS, launched in 1999, which has a 1 m pixel

size [Di et al., 2003]. All these sources provide multi-spectral and multi-temporal

information for characterising and monitoring land cover dynamics at various scales. The

scale, together with the resolution, determines the remote sensing data source appropriate

to any specific mapping problem.

2.2.1 Classification system

Land cover classification evaluates features on the land within the context of the

surrounding landscape. Information regarding standardised land use and land cover is

required for consistent and precise planning and management of infrastructure

development, environmental management, energy and resource development, urban

planning and industrial development [Jansen and Di Gregorio, 2002]. For local, regional

and national programmes it is appropriate to apply standardised classification systems [Di

Gregorio and Jansen, 2000].

A classification system is a tool designed to help an analyst make decisions about the

classification of an object. Usually, a classification system includes some type of decision

classification level. Thus, a classification system should have a hierarchical framework to

accommodate different levels of information, starting with structured broad-level classes

which allow further systematic subdivision into detailed sub-classes [Food and Agriculture

Organisation, 2005]. There is no one ideal land cover and land use mapping system but the

classification system used should meet the classification purpose [Anderson et al., 1976].

Many current classification systems are not suitable for mapping, and thus subsequent

monitoring purposes may not be applicable to the detailed classification at certain local

levels [Food and Agriculture Organisation, 2005]. Some of the classification systems

available include the International Geosphere Biosphere Programme (IGBP), the United

States Geological Survey (USGS), and the Food and Agricultural Organisation/Land Cover

Classification System (FAO/LCCS). While certain of these systems are more universally

applicable, none has been accepted as an international standard. The classification systems

20

use predefined legends and often resemble or incorporate other classification systems to

maintain cohesiveness and allow for data integration [Di Gregorio and Jansen, 2000].

In 1993, United Nations Environmental Programme (UNEP) and the FAO co-ordinated

action towards harmonisation of data collection and management, thus taking the first step

towards the internationally agreed-upon reference base for land cover and land use. The

first operational activity in this direction was the Africover Programme of the Environment

and Natural Resources Service (SDRN) of the FAO. The Programme developed an

approach for conceptualising, defining and classifying land cover: the Land Cover

Classification System (LCCS). According to Di Gregorio [2000], “LCCS provides a

standardised, hierarchical, consistent and flexible classification system with strict class

boundary definitions. LCCS enables comparison of land cover classes regardless of

mapping scale, land cover type, data collection method or geographic location”. LCCS has

been designed with two main phases: an initial dichotomous phase, in which eight major

land cover types are defined, followed by a subsequent modular-hierarchical phase. Land

cover classes are defined by a set of predefined classifiers tailored to each major land

cover type.

The dichotomous phase is the main level of classification used to define the major land

cover classes. Three classifiers are used, namely: presence of vegetation, edaphic

conditions and artificiality of cover. The modular-hierarchical phase is given by a set of

predefined pure land cover classifiers, each of which is different for the eight major land

cover classes. One of the basic principles adopted in the LCCS is that a given land cover

class is defined by the combination of a set of independent diagnostic attributes, the so-

called classifiers. The increase in detail in the description of a land cover feature is linked

to the increase in the number of classifiers used. In other words, the more classifiers added,

the more detailed the class. The class boundary is defined either by the different numbers

of classifiers or by the presence of one or more different types of classifiers. Thus, the

emphasis is no longer on the class name, but on the set of classifiers used to define this

class. For further definition, these pure land cover classifiers can be combined with so-

called attributes. Two types of attributes, which form separate levels of classification, are

distinguished as environmental and specific technical attributes. The use of diagnostic

criteria and their hierarchical arrangement in classes is generally a function of mappability,

i.e. the ability to define a clear boundary between two classes. Hence, LCCS uses

21

diagnostic criteria that are hierarchically arranged to ensure a high degree of geographical

accuracy at the highest levels of classification.

Due to the flexibility of the LCCS, the analyst defines the set of classifiers and attributes

appropriate for classification. In this context, “flexibility” should address the potential for

the classification system to describe enough classes to cope with real world conditions. At

the same time, however, the classification system should adhere to unambiguous strict

class boundary definitions. In addition, the classes in such a system should be as neutral as

possible in the description of a land cover feature to answer to the needs of a wide variety

of end-users and disciplines.

The purpose and thematic content of an exercise in land cover mapping helps to define the

classes that must be differentiated in the land cover product: in other words, to produce the

mapping legend. LCCS is a relatively new classification system and has been applied in

the Africover project [Di Gregorio and Jansen, 2000]. Through the Africover Programme

and the Global Land Cover Network (GLCN) initiative, a number of countries were

mapped using the LCCS classification system including: Burundi, DR Congo, Egypt,

Eritrea, Kenya, Rwanda, Somalia, Sudan, Tanzania and Uganda. This study serves as an

initial attempt to implement LCCS for legend categorisation in Malawi. Its appropriateness

will be critically evaluated in the Shire River catchment. Within the scope of the mapping

exercise, the applicability of LCCS will be analysed for particular applications such as

hydrological modeling.

2.2.2 Earlier land cover mapping in Malawi

Efforts have been made at various points in time to provide land cover information for

Malawi. Present information concerning land cover in Malawi includes Shaxson’s

Vegetation Classification Map of 1976 which was based on the Yang’ambi Vegetation

Classification System and the Biotic Community Map contained in the Atlas of Malawi

based on the IUCN classification system [Malawi Government, 1986]. The 1990 study by

Abbot focussed on woodland change in the Lake Malawi National Park [Abbot and

Homewood, 1999]. Aerial photographs were used to detect and monitor changes in the

Park woodlands over an 8-year period, 1982 – 1990. In 1993, forest cover and biomass

assessments were completed by the Government of Malawi and Satellitbild project. This

project was funded by the International Development Association (IDA) of the World

Bank under the World Bank Energy 1 project for Malawi [Malawi Government and

22

Satellitbild, 1993]. The emphasis was on forest cover information to facilitate the

evaluation of forest degradation in Malawi. This involved making land use/land cover

maps using visual interpretations of Satellite Image Maps (SIMs) from Landsat Thematic

Mapper (TM) imagery collected in 1990/91 [Malawi Government and Satellitbild, 1993].

The land cover information was compared with 1972/73 Landsat Multispectral Scanner

and was interpreted using the same visual method used in 1990/91. In this context, it can

be argued that the mapping procedure was dependent on the intended use of the land cover

map products.

Thorough examination of the available land cover information sets shows a lack of

systematic and reliable national coverage. What exists is fragmented in terms of time,

space, thematic detail and methodology. Land cover classes were evaluated manually using

the traditional visual interpretation methods of printed images. Databases derived using

visual interpretations are often subjective, time consuming and costly. The thematic classes

are too generalised and cannot be guaranteed for consistency and reproducibility [Cihlar,

2000]. These difficulties contribute to the ineffective periodical update of most land cover

maps.

Class labelling of the existing 1992 land cover data set applied a Malawian nationally

standardised classification system. The classification was tailored to the requirements of a

specific project, i.e. forest assessment. Hence, their utility for wider applications is limited.

In addition, map productions using national classification legends do not satisfy the

requirements of global mapping. The lack of compatibility characterises the information

gap between land cover generated in Malawi and other parts of Africa and the world at

large. Harmonised land cover information not only facilitates map-updating but also data

exchange since land cover types cross national boundaries [Di Gregorio and Jansen,

2005].

In the case of the Malawi Government/Satellitbild 1991 land cover map, there appears to

be a registration or projection error. This error has resulted in a 1 – 3 km shift in where the

land cover classes fall relative to the original images the classes were derived from [Orr et

al., 1998]. As such, using that land cover data set could result in the misclassification and

mis-representation of actual land cover information. In addition, appropriate mapping

procedures were not applied and this project did not produce satisfactory results to

23

represent adequately the diversity of applications needed, for example, in hydrological

modelling.

Any further modelling of the hydrological impacts of changes in land cover would thus

require a revision of these earlier classification attempts. For example, the present land

cover information does not include other important themes such as built-up areas.

Generating data sets with a thematic legend compatible with the FAO/Land Cover

Classification System was crucial to the outcomes of this project.

2.3 Methods

Information requirements necessitated accurate mapping of land cover classes in the Shire

River catchment to provide input data. This study serves as an initiative to produce

automated land cover maps less laboriously, more consistently, and rapidly compared to

those derived from visual interpretation. Automatic interpretation of multispectral satellite

imagery has proven to be an efficient procedure for consistent and accurate land cover

mapping. Map productions using satellite data benefit from all the advantages related to the

use of digital data, its periodical acquisition, and coverage of large areas at a relatively low

cost. A comprehensive assessment of the spatial and temporal distribution of land cover

dynamics between 1989 and 2002 was critical for understanding and documenting change,

as well as impacts on the land hydrological processes within the catchment.

Multi-temporal, high resolution satellite remote sensing images, Landsat 5 TM (1989) and

7 ETM+ (2002) were used for mapping land cover dynamics in the upper Shire River

catchment. The high spectral Landsat with 30 m resolution reveals detailed features of the

land surface. This is the scale of the characteristics needed for this study. Furthermore,

Landsat has provided an uninterrupted time series of such data since 1972 to present

[Global Land Cover Facility, 2005] that is critical for understanding and documenting land

cover dynamics within the upper Shire River catchment. Supplementary digital mapping

data sets were obtained from the Department of Surveys in Malawi to complement the

satellite information.

24

2.3.1 Selection of satellite images

Landsat images, level 1G, were obtained from the Global Land Cover Facility (GLCF),

FAO/Africover* project. To a large extent, the selected Landsat images were to provide

flexibility for other applications. It would have been advantageous to obtain multi-temporal

Landsat composites over time rather than snapshots for good land cover maps. However,

within the constraints of a limited number of suitable images in archive, a strategy for

selecting Landsat imagery for development of land cover database for the Shire River

catchment was governed by cost-free available multi-temporal images, vegetation

phenology and image quality (cloudiness, haze). The data set is derived from Landsat

satellite images spanning a 13-year period (July 1989 and May 2002).

Selection is based on seasonal land cover concepts to minimise changes in reflectance due

to phenological changes in vegetation. Seasonal land cover concepts provide a framework

for presenting the temporal and spatial patterns of vegetation in the database. The near-

anniversary dates are composed of relatively homogeneous land cover associations (for

example, similar floristic and physiognomic characteristics) which exhibit distinctive

phenology (that is, onset, peak and seasonal duration of greenness) and have common

levels of primary production. This pair of images ensures that radiometric differences are

due only to land cover change.

Climate of the study area is tropical continental, with 70-90% of precipitation occurring

during the summer and a marked dry winter season. Therefore, cloud-free satellite images

are mainly available during the cool dry period from May to August. Clouds introduce

significant noise to an image by obscuring reflectance of radiation from Earth surface

materials.

Rainfall intensity preceding the date of image acquisition strongly influences biomass

availability and distribution. Interannual variability of vegetation conditions in southern

Africa, contrasting the rainfall and NDVI measurements during contrasting years has been

reported [Anyamba et al., 2003]. During an El Nino episode, a dry phase with extreme

drought causes low vegetation and dry conditions, whereas a La Nina wet phase causes

extreme rainfall and flooding throughout the country, resulting in high vegetation and

* Acknowledgements to Mr David Stevens, UNOOSA and US Geological Survey for providing access to the

Landsat TM and ETM+ images

25

persistent NDVI anomalies. Rainfall conditions during the 1988/89 and 2001/02 season

have been classified as average for much of the country (Figure 2). As such, an assumption

was made that the chosen images were preceded by similar rainfall conditions yielding

comparable vegetation characteristics.

Natural fires in southern African ecosystems are ignited by lightening although human

activity is responsible for the majority of the fires in Africa particularly in grasslands and

savanna biomes [Silva et al., 2003]. The burning season lasts from May to October, with

its peak in July. Aerosol and trace gas emissions are strongly concentrated during the

period July to September [Swap et al., 2003]. Smoke from such fires contributes to

increased radiation scattering, which can reduce the information content of remotely

sensed data. For this study, Landsat images acquired before the peak of the fire season

were utilised to minimise atmospheric effects without compromising the need for

atmospheric corrections.

A critical component of the study was the acquisition of multi-temporal Landsat imagery.

It assumes that a distinct temporal trajectory of land cover dynamics can be identified

using multi-temporal remote sensing data, and that this information will provide increased

land cover identification capability. Considering the present scale of temporal and spatial

changes in land cover in the area, a 13 years period provides a comprehensive land cover

mapping and change monitoring zone.

Landsat 5 Thematic mapper

Landsat 5 TM is a multispectral scanning system and records reflected/emitted

electromagnetic energy in the visible, reflective-infrared, middle-infrared and thermal

infrared regions of the spectrum. The satellite was launched as a sun-synchronous with a

16-day repetitive cycle. It crosses the equator on the north-south portion of each orbit at

9:45 a.m. local time. It has a radiometric resolution of 8 bits, a spatial resolution of 28.5 x

28.5 metres for all bands except band 6, which has 120 metres resolution. Table 2 lists the

seven spectral bands of the TM along with a brief summary of the intended principal

applications of each.

26

Table 2: Description of Landsat 5 TM

Satellite Date of

acquisition Scene no.

Band Wavelength

(µm) Spectral location

Spatial resolution (m)

Landsat 5 TM

1 July 1989 167/70 1 0.45 - 0.52 Blue 30

2 0.52 - 0.60 Green 30

3 0.63 - 0.69 Red 30

4 0.76 - 0.90 Near IR 30

5 1.55 - 1.75 Mid-IR 30

6 10.4 - 12.5 Thermal-IR 120

7 2.08 - 2.35 Mid-IR 30

Landsat Enhanced Thematic Mapper Plus

Landsat 7 was launched in 1999 and uses the Enhanced Thematic Mapper Plus (ETM+) to

observe the Earth. Similar orbits and repeat patterns are used but with major improvements

that include an additional band, the panchromatic, which has a resolution of 15 m and a

60 m spatial resolution thermal channel. The spectral bands and their resolution are listed

in Table 3.

Table 3: Description of Landsat ETM+

Satellite Date of

acquisition Scene no.

Band Wavelength

(µm) Spectral location

Spatial Resolution (m)

Landsat 7 ETM+

26 May 2002 167/70 1 0.45 - 0.515 Blue 30

2 0.525 - 0.605 Green 30

3 0.63 - 0.690 Red 30

4 0.75 - 0.90 Reflective IR 30

5 1.55 - 1.75 Mid-IR 30

6 10.40 - 12.50 Thermal-IR 60

7 2.09 - 2.35 Mid-IR 30

Pan 0.52 - 0.90 15

Supplementary digital data sets were obtained from the Department of Surveys in Malawi

to complement information from satellite imageries [Malawi Government, 1998b]. From

these, a number of digital Geographical Information System (GIS) layers were created

including towns, road networks, administrative borders, soil types and hydrography (river

flow networks). Slope, aspect and altitude were delineated from digital elevation data

downloaded from the Consultative Group on International Agricultural Research website

[Consultative Group on International Agricultural Research]. These data sets formed the

27

bulk of the contextual information, which was used in the analysis together with

reflectance information in the images.

2.3.2 Image processing

Land cover mapping and subsequent quantitative change detection required geometric

registration between TM and ETM scenes, and radiometric rectification to adjust for

differences in atmospheric conditions, viewing geometry and sensor noise and response

[Jensen, 2005; Lillesand et al., 2004].

2.3.2.1 Geometric correction

The necessary geometric corrections can be applied in the early stages of the pre-

processing of the data usually by the ground station receiving the raw data from the

satellite [Milner, 1988]. Landsat images, level 1G, obtained from the GLCF,

FAO/Africover project had these necessary corrections. All the images had been resampled

using the nearest neighbour option and were projected to the Universal Transverse

Mercator (UTM) system.

The images were registered to the MalawiGP UTM Zone36/Arc1950 datum projection

system to match them with available in situ vector data [Malawi Government and

Satellitbild, 1993]. MalawiGP is the system currently being used for all topographic

mapping, and conforms to the UTM Zone 36, in use by several southern African nations

[Malawi Government and Satellitbild, 1993]. It defines a single metric plane reference

system for the entire country with error characteristics that do not exceed 1 part in 2500

(1:2500) with respect to discrepancies between this planar representation and the

ARC1950 model of the spheroidal nature of the Earth.

During change detection processes, it is important that both the images should have the

same spatial resolution. Further resampling was necessary to match the spatial resolution

and extents of the two images. The 1989 Landsat 5 image was resampled to the 2002

Landsat 7 ETM+ 30 meters using the nearest neighbour resampling method and a first

order affine transformation was applied [Microimages, 2005].

Radiometric correction

The images were pre-processed by converting the digital numbers to reflectance units [Culf

et al., 1995]. This correction aimed to minimise variation due to varying solar zenith

angles and incident solar radiation. DN values were converted to radiance values (L) using

28

the calibration coefficients “gain” and “bias” supplied in the imagery report file as shown

in Equation (1.

biasDNgainL +×= )(λ (1)

where λL is the at satellite radiance and DN is the digital number.

Once radiance values were calculated, the reflectance (ρ) was then calculated for each band

as described in Vermote et al. [1997]. Reflectance, as shown in Equation 2, was calculated

on a pixel by pixel basis for each scene using the corner coordinates and the overpass time

and acquisition date contained in the image report file. The λSunE values were taken from

the Landsat 7 science Data Users Handbook [NASA Goddard Space Flight Centre, 2002].

sSunE

dLP

θπ

ρλ

λ

cos

2

= (2)

where ρP is the dimensionless planetary reflectance, λL is the spectral radiance at the

sensor aperture, d is the Earth - Sun distance in astronomical units, λSunE is the mean solar

exo-atmospheric irradiances, and sθ is the solar zenith angle.

Removal of atmospheric and phase angle effects, however, requires information about the

gaseous and aerosol composition of the atmosphere and the bi-directional reflectance

characteristics of elements within the scene [Eckhardt et al., 1990]. It has been observed

that for multi-temporal Landsat images of flat areas, where the relative variation in solar

elevation angle is not significant, and when the images belong to the same season (e.g. dry

season), the atmospheric, phenological and seasonal factors can be assumed to be

comparable [Prakash and Gupta, 1998].

Based on the above reasons, no radiometric normalization was done for the Shire images

as the images belonged to the same phenological season. Images were acquired before the

fire season, thereby reducing atmospheric effects associated with smoke and radiation

scattering. Additionally, the methods used for land cover change detection in this study

involved only a general comparison of the statistics of the various land cover categories

(post-classification). Where pixel-to-pixel comparison was employed, appropriate

thresholds of the change detection image histograms were used [Jensen, 2005]. In addition,

the required atmospheric and bi-directional reflectance information was not available for

any of the two image scenes. The 2002 image was used for the field verification exercise.

29

Field verification exercise

Reference information is used to compare the classification with ground truth, and should

have a higher degree of accuracy than the information used for map production. Sources of

reference information include aerial photography, satellite imagery with better resolution

than that used in map production, and field work [Biging et al., 1998].

In this study, a field exercise was carried out to collect ground reference samples for use in

the assessment of the accuracy of classified images. The field verification exercise was

conducted within the study area between 17 and 31 July 2006. The fieldwork started with a

two-day reconnaissance visit to the study area, and proceeded with data collection. The

2002 image, being the most recent, was selected for the exercise. A colour printout of the

reference image was taken into the field to examine the relationship between the image and

the ground cover classes. Furthermore, the field exercise was done during the dry season,

when the images were acquired, to correlate spectral features of the image with features on

the ground.

Field sampling locations where ground information would be collected were selected

within the study area. Location and number of samples taken during the primary data

collection are illustrated in Figure 5.

The location of the samples was based on the classified image interpretation of the 2002

Landsat image, allowing for the unsystematic sampling of mapping units of similar land

cover characteristics. The interpretation of the image was done by automated classification

with Maximum Likelihood. These samples were identified on the image before going into

the field. In the sampling scheme, the area covered by each land cover class was

considered a cluster of polygons. The sampling design encompassed type, size, location of

samples and number of samples. Polygon sample areas of sizes 100m x 100m and in some

cases 120 m x 120 m were used. Each polygon corresponds to an area more than 30 m2

(one pixel) and a minimum mapping unit of one ha. Across the entire study area,

approximately 20 to 35 samples were selected for each land cover type. There were 228

reference points in total.

30

Fieldwork pointsFieldwork points

Figure 5: Location of the sample sites for primary data collection

The choice of the sample sites was further restricted by accessibility within the study area,

which constituted a strong limiting factor during the data collection. The existing access

consists of a single tarmac main road, secondary roads in poor condition and paths. In view

of this, and considering the dimensions of the study area and the time available, the main

criteria for the location of samples was the distance to the closest accessible road and path.

Choice of sampling areas was biased by proximity to passable roads.

The sites were geo-referenced in the field using a Garmin Global Positioning System

(GPS) receiver. This was accomplished by tagging the sample polygons with location

information (latitude and longitude) which could be targets for the field verification. Field

notes and digital ground photographs were taken for each field reference point. At each site

a GPS reading was recorded and qualitative descriptive information of the land cover was

entered (e.g. whether the land cover consisted of cultivated or grazing areas, grassland,

built-up areas, open water, woodland etc). Such qualitative land cover descriptions have

been shown to be adequate for the purposes of image classification and are sometimes the

only practical option when there are time constraints [Treitz et al., 1992].

31

Global Positioning System units could help in assessing variability encountered in

accessing sample points. Proximity to a sample point could be quantified (GPS error

incorporated) and used in the determination of map accuracy. Samples of 228 polygon

areas were visited in the field to verify their land cover status with a mean error of ±6 m.

The ground data were compared to data derived from image classification. The ground

reference label was paired with the remote sensing-derived label for assignment in the

error matrix [Jensen, 2005; Lillesand et al., 2004]. Indices of thematic accuracy were

derived for the land cover classification. The indices computed included overall accuracy,

and for each class the producer accuracy and the user accuracy were estimated.

Land cover validation limitations

This study was limited by insufficient reference data for the interpretation of Landsat

satellite images. A few factors might have caused errors in the validation of land cover

using satellite images. The user should be aware of the following concerns: 1) limited data

available to validate the classification, 2) the fact that recent available data are a level 1

classification done in 1991 differing in terms of the classification system used. Data sets

based on Landsat 5 TM and Landsat 7 ETM, obtained in 1989 and 2002, are inappropriate

to carry out field verifications after fifteen and four years of land cover changes

respectively.

Classification accuracy assessment

Accuracy assessment of land cover mapping is an important step in the process of

analysing remote sensing data. It offers information about errors in the data, for

improvement of the mapping procedure, and allows for future users of the classification to

assess the suitability of data for particular applications. This may be done by using various

methods. A relatively common one is by expressing overall and category accuracies

[Meijerink et al., 1994]. Overall classification accuracy provides a general picture of how a

specific classifier is performing while the user and producer accuracies help to ascertain

whether the classified pixel actually represents the relevant information class on the ground

[Tso and Mather, 2001].

Overall classification accuracy is the ratio of number of correct classifications to total

number of samples evaluated [Congalton and Green, 1999; Corves and Place, 1994]. The

overall accuracy constitutes the percentages of correctly classified classes lying along the

diagonal and is determined as in Equation 3:

32

∑∑=

)(

)(

totalcolumnortotalRow

diagonalalongclassesclassifiedcorrectlyaccuracyOverall (3)

Producer's accuracy is a measure of how accurately the analyst classified the image data by

category (columns). Producer accuracy details the errors of omission. An error of omission

results when a pixel is incorrectly classified into another category. It is an important

measure because the producers of spatial data are interested in knowing how well a

particular area on the Earth surface can be mapped. Producer accuracies result from

dividing the number of correctly classified pixels in each category (on the major diagonal)

by the number of reference pixels used for that category (the column total). Producer

accuracy is calculated as in equation 4:

rowainverifieditemsofnumbertotal

columnainitemclassifiedcorrectlyofnumberaccuracyProducer = (4)

User accuracy represents the probability that a sample from the classified image actually

represents that category on the ground. User accuracy details errors of commission. An

error of commission indicates the probability that a pixel classified into a given category

actually represents that category on the ground. User accuracy is important for users of

spatial data because users are principally interested in knowing how well the spatial data

actually represents what can be found on the ground [Congalton, 1991; Lillesand et al.,

2004]. The user accuracy is computed by dividing the number of correctly classified

samples of the relevant class by the total number of samples that were verified as

belonging to that class. User accuracy is determined as in Equation 5:

rowainverifieditemsofnumbertotal

rowainitemclassifiedcorrectlyofnumberaccuracyUser = (5)

Accuracy assessment of the classified maps was based on the independent field data set

consisting of observations at the 228 homogeneous sampling areas as described above. The

results were then tabulated in the form of an error matrix. The columns of the error matrix

table define the reference ground data and the rows define the classified image classes. The

values in the cells of the table indicate how well the classified data agrees with the

reference data. The diagonal elements of the matrix indicate correct classifications. The

higher the proportion of the pixels within the user and producer accuracies for the

individual class in question, the more accurate the classified maps are.

33

2.3.3 Image classification

Selection of training areas

The location of training sample locations is normally done by fieldwork or the use of aerial

photographs and map interpretation [Mather, 1993]. However, in the absence of reliable

ground data, the choice of training data has to be considered somewhat subjective. For the

upper Shire River catchment, no reliable ancillary data are available; hence, signatures

were chosen visually through combining spectral and contextual information. It was then

necessary to determine appropriate band sets as input for land cover classification. Several

transformations were carried out on the original bands to achieve this.

Transformations and indices

Landsat image bands were combined in transformations and indices with physical

meaning, enhancing certain characteristics of the land surface. A number of methods can

be applied to perform image enhancement. The process of visually interpreting digitally

enhanced imagery attempts to optimize the complementary abilities of the human mind and

the computer. Different bands of a multi-spectral image were combined to accentuate

different land cover areas given the nature of the landscape involved. This was necessary

for the selection of training data for subsequent classification of the images. In this study,

the following types of transformations and indices were applied:

False colour composites (FCC)

Digital images are typically displayed as additive colour composites that are usually

composed of three bands, each assigned to one of the basic colours: red, green and blue

(RGB) [Jensen, 2005; Lillesand et al., 2004]. The RGB displays are used extensively in

digital processing to display normal colour, false colour infrared and arbitrary colour

composites [Jensen, 2005; Lillesand et al., 2004]. Two types of colour composites i.e. a

false colour composite (FCC) and a natural colour composite (NCC) are distinguished

here. To create a clear feature on the Landsat images, it was necessary to know the

reflection characteristics of the basic cover types of the earth surface. The best FCC

depends on the purpose of the study. For the upper Shire River catchment, from several

FCC combinations of the Landsat bands (Table 2 and Table 3) produced for visual

interpretation, the best combinations were: {band 7, band 4, band 2}; and {band 4, band 3

and band 2} in red, green and blue respectively.

34

Principal Components Analysis

Principal component analysis (PCA) is a statistical method used for compressing the

original data set without losing too much information. PCA involves reducing the

dimensionality of a set of spectral bands to a smaller number of orthogonal axes. The

number of principal components is equal to the number of bands retained after rotational

transformation of the original set of axes. The first Principal Component (PC) is defined by

maximum variance of the original data set; the last PC defines the leftover variance

[Meijerink et al., 1994]. Principal components were computed from the original six bands

of each image to reduce redundancy. The first two components were combined with the

Normalized Difference Vegetation Index (NDVI) to generate image composites for

signature development and classification.

Tasseled Cap transformation

Tasseled Cap (TC) orthogonal transformation of the original six bands of each image was

computed [Kauth and Thomas, 1976]. TC transformation rotates the data such that the

majority of information is contained in three components or features that are directly

related to physical scene characteristics. The first three bands are conventional indices used

for land applications, corresponding to soil, green vegetation and moisture indices

respectively. TC was computed for the two Landsat images to generate the physical feature

characteristics of brightness, greenness and wetness. The computed images were used in

the development of training signatures and subsequent classification. The following

equations (Equations 6, 7 and 8) were used to transform the image data into channels of

brightness (B), greenness (G) and wetness (W):

B = 0.304*B1 + 0.279*B2 + 0.474*B3 + 0.559*B4 + 0.508*B5 + 0.186*B7 (6)

G = (-0.285*B1) + (-0.244*B2) + (-0.544*B3) + 0.724*B4 + 0.0840*B5 + (-0.180*B7) (7)

W = 0.151*B1 + 0.197*B2 + 0.328*B3 + 0.341*B4 + (-0.711*B5) + (-0.457*B7) (8)

Normalised Difference Vegetation Index

NDVI is calculated from a scene by taking the ratio of the difference of the near infrared

and red reflection and the sum of these two bands, as shown in Equation 9 [De Jong,

1994]:

NDVI=( )( )34

34

TMTM

TMTM

+−

(9)

35

where NDVI is the Normalised Difference Vegetation Index; TM 4 is the TM spectral

band 4; and TM 3 is the TM spectral band 3.

Normalised Difference Built-up Index

In monitoring the spatial distribution of built-up areas, the Normalised Difference Built-up

Index (NDBI) [Jensen, 2005] was used to map the possible areas under settlements

(Equation 10).

45

45

TMTM

TMTM

NIRMidIR

NIRMidIRNDBI

+−

= (10)

where NDBI is the Normalised Difference Built-up Index, MidIR is the Mid-Infrared,

NIR is the Near Infrared, TM is the Thermal Band.

After computing the Normalised Difference Built-up Index (Equation 11), a Boolean

image is created by subtracting the Normalised Difference Built-up Index from Normalised

Difference Vegetation Index i.e through differencing two real number images with a value

(0-1).

NDVINDBIupBuilt area −=

(11)

where NDBI is the Normalised Difference Built-up Index and NDVI is the Normalised

Difference Vegetation Index.

Maximum Likelihood Classification

Various methods are used for classifying land cover such as supervised and unsupervised

classifications. Supervised classification uses area statistics based on sample training to

classify an image, whereas unsupervised classifications involve algorithms that examine a

large number of unknown pixels and divide them into a number of classes based on natural

groupings [Lillesand et al., 2004]. In this study, the Maximum Likelihood algorithm was

used for the classification. Maximum Likelihood is a supervised classifier, i.e. the analyst

supervises the classification by identifying representative areas, called training areas.

These areas are then described numerically and presented to the computer algorithm,

which classifies the pixels of the entire scene into the respective spectral class that appears

to be most alike. In a Maximum Likelihood classification, the distribution of the response

pattern of each class is assumed to be normal (Gaussian). The data should include all

spectral variation within each class. In theory, a statistically based algorithm requires a

minimum of n+1 pixels for training in each class, where n is the number of wavelength

36

bands. However, in practice, the use of a minimum of 10 n to 100 n is advised by Lillesand

et al. [2004].

The Maximum Likelihood algorithm consists of two steps:

Estimation of model parameters, which determines the a posteriori probability that a given

pixel belongs to class i, given that the pixel has feature f. This probability is calculated

using Bayes Rule of conditional probability (Equation 12):

( ) ( ) ( )( ) ( )∑

=

j

ipifp

ipifpfip (12)

where p(f|i) is the probability of a pixel having feature f, given that it belongs to class i,

and p(i) is the probability that class i occurs on the image of interest (also referred to as the

a priori probability).

Discriminant functions, which establish the decision rule to classify a pixel is usually set to

be equal to the a posteriori probability for optimal results. It is stated as follows (Equation

13): if a pixel satisfies the equation, then it is assigned to class i.

( ) ( )fDfD ki ≤= (13)

2.3.4 Land Cover Classification System

To achieve land cover harmonisation within Africa and on a global scale, the FAO/LCCS

legend structure Land Cover Classification System (LCCS) [Food and Agriculture

Organisation, 2005] was used as has been discussed in section 2.2.1. All modifications in

the categories were done keeping in view the area under investigation and application of

the derived land cover maps i.e. hydrological modelling.

2.4 Results and discussions

This section describes results obtained from false colour composite and principal

components analysis. Visual interpretation of the images during land cover classification

was improved by means of these transformations. Land cover classification results are

presented with a display of land cover maps generated from the 1989 and 2002 Landsat

images for the Shire River catchment. This is followed by a description of the

characteristics of each classified land cover class based on the LCCS classification system.

Related information portraying spatial extents and distribution of each classified land cover

37

class has been highlighted. The section ends with an interpretation of the accuracy

assessment results.

2.4.1 Transformation results

To aid visual interpretation, visual appearances of the objects in the image were improved

by image enhancement techniques. The goal of image enhancement was to improve visual

interpretability of an image by increasing the apparent distinction between features.

Training areas were derived from the enhanced images in combination with an

interpretation of the indices.

False Colour Composite

A false colour composite image of bands {7, 4, 2} (RGB) was selected as providing the

best visualization, as discussed in Section 2.3.3. Results of this transformation applied to

the 1989 and 2002 Landsat images are presented in Figure 6.

1989

2002 1989

2002

Figure 6: False colour images - 1989 and 2002

This combination distinctively distinguished different types of land cover features:

cultivated/grazing areas are shown in magenta; fresh water bodies emerge in bluish or

black; built-up areas are revealed as shades of pink; and vegetated areas appear in different

shades of green.

38

Principal components

The six standardised principal components (PC1-6) were calculated from the

transformation of the original TM and ETM+ reflective bands 1 – 5, and 7 and the

correlations are shown in (Table 4 and Table 5). The percentage variance that was mapped

to each component after PCA shows that PC1, PC2 and PC3 had higher variance

percentages. These first three components accounted for 98.4% for TM and 98.3% for

ETM of the variability in the data as shown in Table 6. Thus, higher order components (3 –

6) were dropped from further analysis as they constitute noise in the data set.

Table 4: Correlation matrix for Landsat 7 ETM+ reflective bands

b1 b2 b3 b4 b5 b7

b1 1.0 0.9 0.8 0.3 0.7 0.7

b2 1.0 0.9 0.5 0.8 0.8

b3 1.0 0.5 0.9 0.9

b4 1.0 0.9 0.6

b5 1.0 0.9

b7 1.0

Table 5: Correlation matrix for Landsat 5 reflective bands

b1 b2 b3 b4 b5 b7

B1 1.0 0.9 0.8 0.4 0.7 0.8

B2 1.0 0.9 0.6 0.9 0.9

B3 1.0 0.6 0.9 0.9

B4 1.0 0.8 0.6

B5 1.0 0.9

B7 1.0

Table 6: Percentage of variance and correlation mapped to each principal components

in study area

Percentage of variance for ETM+

Percentage of variance for TM

PC1 83.1 86.7

PC2 15.2 11.7

PC3 0.83 0.79

PC4 0.67 0.61

PC5 0.13 0.13

PC6 0.07 0.06

39

Images compiled from the retained principal components are displayed in Figure 7, for

1989 and 2002 respectively. It is apparent that the principal component transformation

displays the land cover classes better than the false colour composite. The PCA images

show more clearly differences between bare soil, green vegetated areas and water bodies.

Water bodies appear as dark blue to black; recently burned grasslands appear in bright

blue; cultivated areas appear as purple; and built-up areas appear in pinkish red colour.

Varying shades of yellow and light green allow discrimination of stages of phenology of

the different types of vegetation.

1989 2002

1989 2002

Figure 7: Principal component images - 2002 and 1989

NDVI

NDVI maps were produced as a measure of biomass distribution over the landscape. The

maps provide an insight of vegetation distribution in the area as shown in Figure 8.

Vegetation indices provide values that are indicative of the spectral reflectance of the

vegetation at a given place. Vegetated areas yielded high values of NDVI because of their

relatively high near infrared reflectance and low visible reflectance. In contrast, water,

clouds and iron roofed built-up areas have larger visible reflectance than near-infrared

reflectance. Thus, these features yielded negative index values. Rock and bare soil areas

have similar reflectance in the two bands (3 and 4) and result in vegetation indices near

40

zero shown as pale brown to brown. Varying shades of green depict different types of

vegetation. For purposes of this study, our focus was on categorical vegetation mapping, as

such NDVI values derived from these images were not used for quantitative analysis.

Results from NDVI was an initial procedure in the classification process enabling land

cover features to be grouped into two broad categories: vegetated and non-vegetated areas.

1989 20021989 2002

Figure 8: NDVI images - 1989 and 2002

2.4.2 Land cover maps

The upper Shire River catchment land cover classification presented here is a result of the

Maximum Likelihood classifier. Eight land cover classes were identified in the upper Shire

River catchment as displayed in Figure 9.

41

19892002

19892002

Figure 9: Land cover maps - 1989 and 2002

The classified land cover categories in the upper Shire River catchment were woody

closed, woody open, savanna shrubs, grasslands, marshes, cultivated or grazing areas

built-up areas and fresh water. Land cover classes and their definitions developed from

FAO/LCCS classification system are shown in Table 7.

42

Table 7: Land cover classes and their definitions

Land Cover LCCS Code LCCS user label

LCCS Own description LCCS Label attributes

Built-up Areas 5003-14 Built-up areas

Includes village settlements (both clustered and scattered) with grass thatched roofing and mud walls

Medium Density Urban Area(s)

Natural Waterbodies

8011-1 Fresh water Lake and river water Deep To Medium Deep Perennial Natural Waterbodies (Flowing)

Cultivated and managed

terrestrial areas

11291-12771-S0305S0403S0

503

Cultivated/ grazing areas

Cultivated areas during the wet season, communal land for grazing (cattle and goats) after harvesting before the next growing season. Land holding sizes >0.5 ha per

household, on average.

Small Sized Field(s) of Rain fed Graminoid Crop(s) (Two Additional Crops) (Two Herbaceous Terrestrial Crops both with Simultaneous Period) Dominant Crop: Cereals - Maize (Zea mays L.) 2nd Crop: Roots & Tubers -Sweet potato (Ipomoea batatas (L.) Lam) 3rd Crop: Pulses & Vegetables - Beans (Vigna spp.)

Forest 20599-13225-L23L8N2N4P1

0

Woody closed

Gazetted as protected areas but prone to encroachment

Broadleaved Evergreen High Trees with Dwarf Shrubs Major Landclass: Sloping Land, Medium-Gradient Escarpm. Zone, Slope Class: Hilly Soils: Soil Surface, Stony (5 - 40%) Altitude: 1000-1500 m

Grasslands 40397-4732 Marshes Mostly drained and cultivated during dry season in the shallow areas.

Closed Medium Tall Grassland On Permanently Flooded Land

Woodland 20808-97744-L15L5N2N1109

P8

Woody open Remnants of the vegetation mostly in cultivated areas available as fruits trees and agroforestry trees (msangu and masau and mango

trees)

Semi-Deciduous (40 - (20-10)%) Woodland with Open Medium to Tall Herbaceous Layer and Sparse Medium High Shrubs Major Landclass: Level Land, Valley Floor, Slope Class: Flat to almost flat Soils: Soil Surface, Subsurface: Ferralsols Altitude: 300-600m

Grasslands 20441-12289-L11L5N2N4P7

Grasslands Prone to annual dry season burning Medium Tall Grassland with Medium High Trees Major Landclass: Level Land, Plain, Slope Class: Flat to Almost Flat Soils: Soil Surface, Stony (5 - 40%), Altitude: 100-300 m

Shrubland 21104-40064-L11L5N2N1109

P8

Savanna shrubs

Dominated by scattered baobab trees (woodlands) and man induced

thicket common in cultivated parklands

Broadleaved Deciduous Medium High Shrubland with Closed Medium to Tall Herbaceous and Medium High Emergents Major Landclass: Level Land, Plain, Slope Class: Flat To Almost Flat Soils: Soil Surface, Subsurface: Ferralsols Altitude: 300-600 m

43

2.4.3 Description of land cover classes

Detailed descriptions of each of the land cover class characteristics follow:

Woody closed (LCCS code: 20599-13225-L23L8N2N4P10)

This woodland category covers those communities referred to as Miombo in which

Brachystegia-Jubernalia species are dominant (Figure 10). It is predominantly composed

of mixed evergreen or deciduous forest; and dry evergreen broadleaf types Combretum,

Acacia and Piliostigma tree species. The few emergent shrubs and grass layers are

depressed by the relatively light-crowned trees. From a phenological point of view, this

type of deciduous forest looks evergreen, as the diverse deciduous species do not shed their

leaves at the same time. Woody closed areas are common in the hilly escarpments

dominated by stony lithosols.

Figure 10: Woody closed

Woody open (LCCS code: 20808-97744-L15L5N2N1109P8)

Woody open consists of plant formations comprising a continuous woody stratum and an

herbaceous stratum (Figure 11). The woody trees dominate the herbaceous layer, with a

total woody cover of between 35 and 60 percent. This category is associated with short

grass and deciduous broad-leaved woodlands, common on areas of ferralsols soils.

Selective felling has given rise to various communities with single species dominance of

trees such as Zizyphus jujube (masau), Faidherbia albida (msangu) and Mangifera indica

(mango) common in cultivated parklands. These trees are left intentionally - farmers

collect the fruits for sale or domestic use (masau and mango). Faidherbia albida is rich in

nitrogen and mostly left in the farmlands for nitrogen fixing purposes. This vegetation

exemplifies the regrowth after exploitation by over-logging, clear-cutting or cultivation.

44

woody open with mature trees of F.albida woody open with shrubs and regrowthswoody open with mature trees of F.albida woody open with shrubs and regrowths

Figure 11: Woody open

Savanna shrubs (LCCS code: 21104-40064-L11L5N2N1109P8)

These are savanna areas where the woody stratum comprises mainly shrubs with a cover of

between 5 and 40 percent (Figure 12). When cover is between 5 and 10 percent, the shrub

savanna is referred to as “open” and, when cover is 30 to 40 percent, “closed”. This class is

made up of shrubs generally found in flat areas where undergrowth is predominantly grass.

It includes a substantial portion of unstocked forest with less than 20% crown coverage.

Tree species are dominated by scattered baobab trees. Disturbances within savanna-

dominated areas have given rise to continuous grassy ground cover and scattered thickets

of shrubs and trees.

closed savanna shrubs open/sparse savanna shrubsclosed savanna shrubs open/sparse savanna shrubs

Figure 12: Savanna shrubs

45

Grasslands (LCCS code: 20441-12289-L11L5N2N4P7)

Grasslands constitute communities dominated by medium to tall grasslands with forest

relics and isolated stands of shrub-lands. The grasslands are dense and tussocky tall on the

higher slopes and short on the lower slopes. The pattern is controlled by annual grass fires

(Figure 13).

Figure 13: Grasslands

Marshes (LCCS code: 40397-4732)

Marshes represent edaphic communities under the control of a high water table, which may

give rise to permanent or seasonally inundated grasslands (Figure 14).

Marshy area undisturbed Marshy area under cultivationMarshy area undisturbed Marshy area under cultivation

Figure 14: Marshy area

Cultivated/grazing (LCCS code: 11291-12771-S0305S0403S0503)

Cultivated/grazing areas consist mainly of rainfed subsistence agricultural fields. It

corresponds also to less vegetated surfaces that include abandoned fields and exposed soil

during the dry period. The fields are open customary land where goats and cattle are left to

graze during the dry season or immediately after harvest (Figure 15).

46

Figure 15: Cultivated or grazing lands

Built-up areas (LCCS code: 5003-14)

This class contains substantial amounts of constructed surface mixed with substantial

amounts of vegetated surface. Small buildings (such as single-family housing, farm

outbuildings, and sheds), streets, roads, and cemeteries typically fall into this class. It

includes village settlements (both clustered and scattered) with grass thatched roofing, mud

walls and occasionally iron sheet roofing (Figure 16). Landscape categories usually

associated with barren conditions such as rock outcrops, bare earth, bare soil, gravel road

were not considered. This is in part due to the factor of difficulty in differentiating between

such areas with similarly high reflectance values in multi-spectral imagery.

Grass-thatched dwelling Iron-roofed dwellingGrass-thatched dwelling Iron-roofed dwelling

Figure 16: Built-up areas

47

Fresh water (LCCS code: 8011-1)

This class includes all areas of open water with less than 30% cover of trees, shrubs,

persistent emergent plants, emergent mosses, or lichens (Figure 17).

Figure 17: Fresh water body

The spatial extents of each land cover type for 1989 and 2002 are shown in Table 8. The

graphical representation of spatial distribution of land cover classes for 1989 is shown in

Figure 18.

Spatial distribution of land cover classes 1989

The classification indicates large areas of savanna shrubs mainly located in the lower

escarpments occupying 152 791 ha (34%). Savanna shrubs are associated with scattered

trees and bush thickets (less than 15m of height) as well as open grasslands. Cultivated

areas, which are simultaneously used as grazing areas after crop harvest, formed the next

dominant covering 95 428 ha of the total study area representing 21%. Woody open class

formed another class with up to 70 612 ha indicating 15.5% of the study area. The woody

closed areas covering 37 930 ha (8.3%) form mostly the protected areas with Miombo

forest (tall trees <30m and less shrubs or no undergrowth) being the dominant vegetation.

Built-up areas formed another class occupying 39 813 demonstrating an extent of 8.7% of

the total land surface in the upper Shire catchment. The fresh water body occupied

38 353 ha of the land surface representing 8.4%. Minor extents of the land surface were

occupied by grasslands, 15 127 ha (3.3%) and marshes occupied 6 444 ha (1.4%).

48

Table 8: Spatial distribution of land cover classes – 1989 and 2002

1989 2002 Class

Area (Ha) Percent Area (Ha) Percent

Fresh water 38 353 8.4 37 178 8.1

Built-up areas 39 813 8.7 14 326 3.1

Cultivated/grazing 95 428 20.9 117 071 25.7

Marshes 6 444 1.4 29 490 6.5

Grasslands 15 127 3.3 63 664 14.0

Savanna shrubs 152 791 33.5 112 356 24.6

Woody open 70 612 15.5 38 446 8.4

Woody closed 37 930 8.3 43 967 9.6

Total 456 498 100 456 498 100

0

50

100

150

200

Fresh water

Built_up areas

Cultivated/graz

Marshes

Grasslands

Savanna shrubs

Woody Open

Woody closed

Area ( 103 Ha)

Figure 18: Land cover extents - 1989

Spatial distribution of land cover classes 2002

The magnitude of the coverage of land surface features for 2002 was slightly different as

graphically portrayed in Figure 19. Among all the land cover classes, cultivated or grazing

areas had the largest extents covering about 117 071 ha (25.7%) of the land surface. The

next large categories included savanna shrubs 112 356 ha (24.6%), grasslands 63 664 ha

49

(14%) and woody closed areas 43 967 ha (9.6%). Woody open represented 38 446 ha

(8.4%) of the study area. Fresh water covered about 37 178 ha totalling 8.1% of the land

surface area and marshes represented 29 490 ha (6.5%). The smallest category was

occupied by built-up areas with 14 326 ha (3.1%) representation.

0

20

40

60

80

100

120

140

Fresh water

Built_up areas

Cultivated/grazing

Marshes

Grasslands

Savanna shrubs

Woody Open

Woody closed

Area (103Ha)

0

20

40

60

80

100

120

140

Fresh water

Built_up areas

Cultivated/grazing

Marshes

Grasslands

Savanna shrubs

Woody Open

Woody closed

Area (103Ha)

Figure 19: Land cover extents - 2002

2.4.4 Distribution of land cover categories

Documenting the distribution of land cover types within the Shire River catchment is the

foundation for applications not only in the study of surface water redistribution and run-off

but also in monitoring environmental trends. The advancement in the capabilities of

Landsat satellite has enabled the measurement of the amount of impervious surface and

vegetation representation with a 30 m pixel in the Shire River catchment. Impervious

surface and vegetation cover distribution can be combined to measure and model

hydrological impacts. Impervious surfaces include all surfaces (fabricated or natural) that

inhibit infiltration by rainfall. Vegetation acts as an integrator of many physical and

biological properties of an area. Characterising the complex cover types created by

variation in vegetation cover proportions provide spatial data that can be used to model

hydrological impacts in the Shire catchment. The classified land cover classes indicate

marked variations in their spatial distribution [Palamuleni et al., 2006].

50

Woody closed

In reference to the altitudinal distribution, this type of forest is mostly confined to the

higher elevations. The majority of woody closed areas constitute the bulk of forested area

in the south east (Machinga Hills), north east (Namizimu forest) and mid north west

(Phirilongwe) regions of the catchment (Figure 20).

Machinga Hills

Namizimu Forest

Phirilongwe Forest

Machinga Hills

Namizimu Forest

Phirilongwe Forest

Figure 20: Location map of forest reserves

Woody closed areas are found mostly in designated forest reserves with dissected rocky

steep escarpment and gorges. Most of the vegetation is associated with shallow and stony

lithosols. These factors are unattractive to subsistence farming and as such, these areas

have not been utilised as much for subsistence cultivation. However, due to increasing

population growth there is increasing land pressure and demand on forest resources. Forest

transitions of shrubs and grasslands dominate the lower and middle sections of the

escarpment especially in areas close to settlements. These changes are discussed further in

Chapter 3.

Also classified as woody closed are Eucalyptus globulus (bluegum), Gmelina aborea

(gmelina) and Acacia nigrescens (acacia) woodlots mainly located in the eastern part of the

study area and within school grounds. Annual tree planting exercises promote planting of

such trees. The woodlots had similar spectral reflectance to woody closed as verified by

field data and site context.

51

Woody open

Apart from forest reserves, which have been mapped as woody closed areas, the woody

open category is found in the lower escarpment and the valley regions within the study

area. Remnants of woody open forests are located in the river valleys, residual hills and

cultivated or grazing areas. The land use practices are the major causes of deforestation

and degradation of this type of forest. Trees of economic importance such as Zizyphus

jujube (masau), Mangifera indica (mango) and Faidherbia albida (msangu) are left on the

cultivated fields and around homesteads.

Savanna shrubs

Savanna shrubs constitute the dry deciduous group, which is more pronounced in the

calcimorphic alluvial terraces of the catchment. Savanna shrubs occur in close association

with grasslands and cultivated or grazing areas class, and are referred to as mopane

woodlands. This category has wider altitudinal distribution than any other class. A majority

of the savanna shrubs are found from 100 m to as high as 800 m above sea level and

commonly distributed in various sections in the catchment.

Grasslands

Continuous tree felling has given rise to open and sometimes patchy grasslands found

almost everywhere in the study region. This class occurs in close association with savanna

shrubs and cultivated fields where grass is left to thrive after harvesting. The grass heights

may range from 1.5 m to 3 m commonly Acacia macrothyrsa often in the well-drained

calcimorphic soils while the short grasses, Chloris gayana (Luba) are found in the eroded

low-lying sites. However, discriminating the different grass species is rather difficult in

remote sensing due to similarity in spectral reflectance. Cattle often graze the short grasses

and the tall grasses are harvested for thatching houses. The grasses are often burned during

the dry season. If grasslands are overgrazed or subjected to uncontrolled burning, patches

of bare ground are created, leading to increased evaporation and accelerated run-off.

Marshes

Seasonally inundated grasslands found along the shores of Lake Malombe and the Shire

River. Field exercise indicated dry season cultivation in the marshes when water levels

were generally very low.

52

Cultivated/grazing areas

During the fieldwork exercise, it was noted that there is a clear indication of continuous

cultivation in most farming areas (Figure 9). Fire is the most commonly used land-clearing

practice, which contributes substantially to the degradation. Because of low yields and

insufficient areas for subsistence cultivation, farmers encroach into neighbouring hills,

which are primarily forests and steep slopes. This is the most damaging form of activity

because such areas are vulnerable to accelerated soil erosion and run-off.

As noted by the Department of Environmental Affairs, people in Malawi and the upper

Shire River catchment in particular, will continue to be dependent on natural resources for

their livelihood in the foreseeable future [Malawi Government, 1998a]. At present, the

country has a large population, which is not yet consumer-oriented, has low energy

consumption, small-scale industrialisation and subsistence pattern of farming without

rotational cycles. However, since subsistence agriculture has low productivity, there has

been an inevitable tendency to increase the amount of land being farmed by using marginal

areas.

Patches of cultivated or grazing areas classified in the Liwonde National Park could be

attributed to misclassification and differences in water level between the two periods.

Built-up areas

Built-up areas in the study area are both clustered and scattered depending on population

densities. Mangochi Township forms the major township located near the source of Shire

River while village settlements are spread throughout the study area. However, detailed

settlement distribution could not be easily mapped due to spectral similarity between

building materials (grass roofing and mud walls) and the surrounding grasslands and

cultivated areas (Figure 16). A detailed discussion of the changes in built-up areas has been

given in Section 3.4

Fresh water

The mapped water bodies are mostly reservoirs and rivers. Water volumes and surface

areas change in response to seasonal and climatic variations. The largest continuous water

body includes Lake Malawi, Lake Malombe and the Shire River. Most of the rivers

flowing from Machinga hills and Phirilongwe forest into the Shire River have continuous

water during the wet season. During the dry season, most of the small rivers have no flow

practically invisible with Landsat 30 m resolution.

53

2.4.5 Thematic accuracy assessment

Accuracy assessment was based on the correlation between ground reference samples

collected during field exercise and the satellite image classification to give an indication of

the overall agreement between ground-truthing data and processed classifications. Based

on the ground truth observations and the classification, the error matrix in Table 9 was

constructed. The overall mapping accuracy was 87%, with individual classes being

mapped at accuracies of above 77% (user), and above 77% (producer).

Some land cover classes presented particular problems for mapping from the imagery.

These problems were overcome using additional data when available [Palamuleni et al.,

2007]. The most significant confusion was between built-up areas and cultivated or

grazing lands. This yielded a user accuracy of 77%. Built-up areas (especially grass-

thatched) and cultivated or grazing areas occur in similar environments and are often in

adjacent or mixed stands. During dry periods when there is little chlorophyll in the

vegetation, grazing causes exposure of soil between remaining vegetation resulting into

similar spectral values making it difficult to distinguish the two classes. To separate these

types, a combination of vector layers of towns and roads, Normalised Difference Built-up

Index (NDBI) and false colour composites were used. From the NDBI, a Boolean image

was created that contained only built-up and barren pixels having positive values while all

other covers had a value of 0 or –254. This technique was reported to be 92% accurate

[Jensen, 2005] whereas in this research a producer accuracy of 83% was attained which is

highly acceptable.

Marshes were often in confusion with woody open and woody closed areas resulting into

user accuracy of 79% and producer accuracy of 77%. This could be attributed to the

phenology of the covers both having high chlorophyll during the dry season. The woody

trees: Zizyphus jujube (masau), Mangifera indica (mango) and Faidherbia albida (msangu)

are in leaf during the dry season. The plausible way to separate them was using the DEM

where marshes are located in very low areas close to water bodies while woody trees are

mostly located in upper areas in association with cultivated areas.

Shrublands and grasslands in the catchment form a complex matrix that is patchy in some

places and homogeneously mixed in others, and spectral separation was difficult. To

separate these types, transformations and contextual knowledge of the area were

incorporated. Digital elevation and soil data helped to identify boundaries between woody

54

open and woody closed areas. The closed woodlands are dominant in hills and plateau

areas while the open woodlands are common in the lower escarpment. From a remote

sensing point of view, discriminating open mixed deciduous with an area cleared under

shifting cultivation or logging activities proved difficult. Because of the dry non-growing

season, the field supposed to belong to cultivation category appeared mostly as shrubland.

The most accurately classified land cover was the fresh water category due to its

contiguous nature that exhibited a unique spectral signature. Fresh water had 100%

accuracy for both producer and user accuracy. The user accuracy of 100% for grassland

could be explained by the fact that the land cover class was broad and large homogeneous

polygons were purposively selected for the exercise.

2.5 Conclusion

The basis of this research comprised multi-temporal classification of Landsat satellite

imagery to provide a recent perspective of land cover types within the upper Shire River

catchment. Results from this study indicate successfully the synergy between Landsat data,

vector data and detailed ground information in mapping land cover features. It was found

that by integrating contextual information and ancillary data, discrimination of

heterogeneous land cover classes dominant in savanna woodlands was improved. Eight

land cover classes were mapped for the Shire River catchment, with an overall accuracy of

87%. Thus, the classification procedure not only separated land cover into different

classes, but also illustrated the patterns that exist across the landscape.

In addition, this project has produced land cover maps at 1:50 000 scale for the Shire River

catchment, covering 4,500 km2. These maps were prepared using Landsat satellite data,

acquired in 1989 and 2002 as the main data source and thus represent the land cover

existing at that time. The land cover maps have compatible digital formats hence they can

easily be applied to a variety of future GIS applications. Additional themes can be

incorporated as more resource information becomes available, or as new management

needs are identified.

55

Table 9: Error matrix for land cover classes

Ground truth data

Classification Fresh water

Built-up areas

Cultivated /grazing

Marshes Grass-lands

Savanna shrubs

Woody open

Woody closed

TR Accuracy

Fresh water 11 0 0 0 0 0 0 0 11 100.0%

Built-up areas 0 20 5 0 0 0 1 0 26 76.9%

Cultivated /grazing 0 2 40 0 1 2 5 1 51 78.4%

Marshes 0 2 0 23 0 0 0 0 25 92.0%

Grasslands 0 0 0 0 29 0 0 0 100 100.0%

Savanna shrubs 0 0 0 0 0 20 2 0 22 90.9%

Woody open 0 0 1 6 0 0 31 1 39 79.5%

Woody closed 0 0 0 1 0 0 0 24 25 96.0%

TC 11 24 46 30 30 22 39 26 228

Total 100% 83.3% 87% 76.7% 96.7% 90.9% 79.5% 92.3%

where TD = sum of major diagonal, TC = column totals, TR = row totals. The overall classification accuracy was TD/TR (198/228) = 87%. Errors were considered consistent with limits of the available technology and ancillary data.

56

The present study has adopted the hierarchical legend structure determined by the Food

and Agriculture Organisation (FAO) Land Cover Classification System (LCCS) to label

land cover variables [Di Gregorio and Jansen, 2005]. LCCS was used as a basis for the

classification to achieve legend harmonisation within Africa and on a global scale.

Currently, the system serves as an internationally agreed reference base for land cover

mapping. LCCS is a relatively new classification system and has been applied in the

Africover project [Di Gregorio and Jansen, 2005]. Through the Africover Programme and

the GLCN initiative, a number of countries were mapped using the LCCS classification

system including: Burundi, Democratic Republic of Congo, Egypt, Eritrea, Kenya,

Rwanda, Somalia, Sudan, Tanzania and Uganda. No land cover mapping study to date has

been conducted in Malawi using LCCS as a coding classification system. The adoption of

LCCS carried out in this research may thus be an important step towards a rigorous update

and translation of the existing land cover maps for Malawi.

The new classification system is also internally consistent, allowing scalability and

mappability that can be used at different scales and levels of detail to discriminate land

cover features. The methodology is applicable at any scale and is comprehensive in the

sense that any land cover identified anywhere in the world can be readily accommodated.

In addition to compatibility with the FAO/LCCS, the derived land cover maps have

provided recent and improved classification accuracy, added thematic detail compared to

the existing 1992 land cover maps.

LCCS presents an advantage for mapping heterogeneous landscapes such as those present

within the Shire River catchment in Malawi. The distributions of savanna woodlands, rural

residential areas, both clustered and scattered (with grass thatched roofing, mud walls and

occasionally iron sheet roofing) and cultivated or grazing areas, represent classes which

have similar spectral signatures (especially during the dry season). They occur in similar

environments and are often in adjacent or mixed stands. Flexibility of LCCS allows

incorporation of geographical data sets to facilitate definition of class boundaries that are

explicit and clear. In this study, digital elevation objects, soil and underlying geological

features were used to identify boundaries between spectrally similar land cover types, for

example, open woodlands and closed woodlands. The closed woodlands are dominant in

hills and plateau areas associated with lithosols while open woodlands are common in the

lower escarpment. Overall, the hierarchical classification technique presents an advantage

57

for mapping heterogeneous landscapes such as those present within the Shire River

catchment.

In terms of classification methodology, the overall approach developed and applied in the

study comprises a variety of techniques ranging from integration of contextual information

such as terrain height and soil characteristics to Maximum Likelihood classifier. The

combination of analytical tools grants considerable insights in land cover mapping of

heterogeneous savanna woodland environments. The approach has the potential of being

adapted and replicated in other similar regions of the world, hence enhancing knowledge of

land cover mapping and dynamic processes at the policy-relevant meso-scale.

The derived land cover maps have added thematic detail enabling spatial and temporal

analysis of patterns and trends in land cover in the upper Shire River catchment. Mapping

of built-up areas was significant as it provided information for establishing the association

between the spreading pattern of the built-up areas with the intensity of cultivation system

and other changes in land cover classes. As such, the characteristically localised nature of

catchment changes can easily be accounted for, making it possible for interpretation and

prediction of the effects of land cover change. These changes are further discussed in

Chapter 3. Results of land cover dynamics reported here will be used as input data for

hydrological modelling in Chapter 4 to establish relationships between fluctuations in the

hydrological regimes and each land cover class.

58

Chapter 3

3 LAND COVER CHANGE ASSESSMENT 1989-2002

Chapter 3 examines of land cover changes in the upper Shire River catchment. Using

Landsat classifications, a discussion of the changes in landscape diversity and

fragmentation between 1989 and 2002 is presented. Areal statistics and the direction

of change in each land cover class were derived. Combination of both image overlay

and post-classification change detection methods reveals significant changes that have

occurred in the Shire River catchment between 1989 and 2002. Changes in land cover

have potential effects on the hydrological processes of the catchment.

3.1 Land cover change

Rapid changes in land cover have a significant impact on conditions of catchment

ecosystems. Accurate information on the status of and trends in land cover changes is

needed to develop strategies for sustainable development and to improve the livelihood of

communities. The ability to monitor catchment land cover changes is highly desirable by

both local communities and policy decision makers. With increased availability and

improved quality of multi-spatial and multi-temporal remote sensing data as well as new

analytical techniques, it is now possible to monitor catchment land cover changes and their

potential hydrological responses in a timely and cost-effective way.

In Malawi, as in many other countries, the landscape is continually changing under the

influence of several factors (including population growth, agricultural expansion and

urbanisation), and as a result, land cover maps rapidly become out of date. In the area

under consideration - the Shire River catchment - two factors may be related to land cover

changes in the past two decades: population growth and the expansion of subsistence

agriculture. Population growth causes changes in land cover through the expansion of

human settlements. The conversion of land to cropland for food security is a cause of

major concern. Forest clearance has resulted from an increased demand for forest products

such as fuelwood, commercial logging and construction materials. Many current land

cover practices in the catchment have the potential to adversely effect and degrade the

environment with respect to forests, soil, water, and biodiversity resources. Previous

studies show increasing land cover change without quantifying the degree and direction of

change in terms of land cover categories [Green and Nanthambwe, 1992]. The impact of

such land cover changes on land surface hydrological processes is of major interest in this

thesis.

59

3.1.1 Land cover change and hydrological response

Land cover plays a critical role in the hydrological cycle, as water availability is generally

a consequence of the distribution of precipitation into evaporation, run-off and soil

moisture storage [Dolman and Verhagen, 2003]. There are many connections between land

surface characteristics and the water cycle. Firstly, land cover can affect both the degree of

infiltration and run-off following precipitation events. Secondly, the degree of vegetation

cover and the albedo of the surface can affect rates of evaporation, humidity levels and

cloud formation. Any change in land cover will have correlated effects in the hydrological

regimes, and possible impacts on the habitat and ecological communities [Calder, 1992;

Lorup et al., 1998].

Agricultural developments have resulted in the widespread deterioration of soil structure, a

process which favours soil sealing and crusting, and related modifications in the rates of

infiltration and storage [Boardman and Favis-Mortlock, 1993; De Roo et al., 2001;

Robinson, 1990]. Specifically enhanced grazing pressure and intensive cultivation

practices have led to soil compaction, reduced infiltration and ground water recharge and

excessive run-off [Boardman and Favis-Mortlock, 1993; Evans, 1990; Forher et al., 2001;

Moussa et al., 2002].

In catchments which were traditionally cultivated or grazing areas that have been changed

to woodland, increases in evaporative loss and decreases in discharge to the outlet have

been observed [Dagnachew et al., 2003; De Roo et al., 2001; Robinson, 1990]. The effect

of land drainage on river discharge can enhance peak flows by increasing the density of

rivers or inhibiting water flow (infiltration) and storage within the soil matrix [Maidment,

1993]. Consequently, seasonal variations in river discharge have been associated with

decreased infiltration during the wet season, which can influence the streamflow during the

dry season by inhibiting ground water recharge.

The consequences of forestation include an increase in infiltration and a reduction of the

incidence of surface run-off influencing the seasonal regime of rivers. However,

infiltration generally depends upon the combined effect of soil properties, climatic

conditions as well as land cover and the developmental stage of the vegetation.

Demonstrations from catchment studies showed that the pine forestation of former

grassland not only reduces annual streamflow by 440 mm but also reduces the dry season

flow by 15 mm [Bosch, 1979]. Thus, benefits gained by the forestation of degraded or

eroded catchments will be dependent on the situation and the management methods

employed [Calder, 1998].

60

Various microclimate changes resulting from land cover changes have been documented.

Deforestation alters the disposition of radiant energy by increasing surface albedo and

daytime long-wave emission by the land surface resulting in lower net radiation [Bastable

et al., 1993; Culf et al., 1995; Larkin, 2002]. The characteristics and variable state of the

land surface control the proportions of net radiation extended upon latent

evapotranspiration and sensible energy flux. A drastic change in vegetation cover, such as

clear cutting in the Pacific north-west, can produce 90% more run-off than in catchments

unaltered by human practices [Franklin, 1992]. Replacing forests with other land cover

types changes the energy partitioning because of differences in leaf area, aerodynamic

roughness, root depth, and stomata behaviour. Field studies have verified that the

replacement of forest by pasture or non-irrigated crops reduces evapotranspiration which

generally increases streamflow, possibly increasing flood hazards [Giambelluca et al.,

2000; Jipp et al., 1998; Wright, 1992]. However, as Bruinjzeel [2001] emphasises, the

effects of deforestation on evapotranspiration and streamflow are not uniform, and depend

on the original forest type, characteristics of replacement land cover, climate, exposure and

soil depth.

Canopy characteristics such as canopy depth, crop height and spacing, leaf area and shape

will have an influence on the hydrological performance of vegetation cover through its

effects on interception. Studies conducted in wet conditions [Bouten et al., 1991] indicate

that interception losses will be higher from forests than from shorter crops resulting in the

reduction of run-off from forested areas compared with those under shorter vegetation. In

addition, the amount of vegetation controls the partitioning of incoming solar energy into

sensible and latent heat fluxes thereby affecting evapotranspiration rates. Yet in deforested

areas, interception and evaporation rates will be lower while increasing surface run-off and

sediment production which consequently lower flow discharge to rivers [Calder, 1992;

Newson, 1992; Shaw, 1990]. However, the amount of precipitation intercepted and

evaporated will vary depending on the vegetation type and species [Hall and Calder,

1993].

Land cover conversions such as the change from forests or vegetated areas into urban land

uses have many hydrologic effects with significant ecological and sociological

ramifications [Booth, 1991; Hollis, 1975; Leopold, 1968]. These changes reduce

interception, infiltration, subsurface flow, evapotranspiration, storm water storage on hill

61

slopes, and the time required for storm water to travel over and through a hill slope to a

stream [Burges et al., 1998; Dinicola, 1990]. The percentage of the catchment surface that

is impermeable due to urban and road surfaces influences the volume of water that runs

and increases the amount of sediment that can be moved [Arnold and Gibbons, 1996].

Urbanisation is known to transform permeable areas such as forest and farmland into

impermeable areas, thereby reducing the amount of water infiltrated into the soil and the

amount of water on the ground surface [Cheng and Wang, 2002; DeFries R. and

Eshleman, 2004]. As run-off volumes in urban channels increase, the duration of high flow

decreases because groundwater is no longer contributing to the flow. In addition, urban

development causes a decrease in lag time between rainfall and run-off by increasing the

hydraulic efficiency of the drainage system (water can reach the channel more swiftly

when it travels over smooth, hard surfaces). In addition, in urban areas, which are largely

covered by roads and buildings, water evaporation from the ground surface and plants into

the atmosphere is generally lower than in agricultural areas or forests. Recent results

regarding the impact of land cover change on hydrological regimes in urban areas have

been discussed elsewhere [Acreman et al., 2000; De Roo et al., 2001; Jennings and

Jarnagin, 2002; Lahmer et al., 2001]. Theoretically, urban built-up areas - especially those

with impervious surfaces - could affect the flow behaviour within a catchment.

The size of a catchment will also influence the hydrological response of precipitation

where small catchments show obvious responses to specific land uses. In large catchments,

meanwhile, hydrological responses are affected by the complex water storage and release

mechanisms [Calder, 1992; Forher et al., 2001]. In small catchments, land use changes to

urbanisation will reduce evapotranspiration and infiltration and increase surface run-off

[De Roo et al., 2001]. Land use change to forest will increase evaporation which may

result in low soil moisture and possibly increase infiltration during rainfall resulting in a

decreased run-off [Calder, 1992; Dagnachew et al., 2003; De Roo et al., 2001]. However,

management activities associated with forestry such as cultivation, drainage, road

construction and soil compaction during logging are more likely to influence hydrological

responses than the presence or absence of the forests themselves.

The magnitude of land cover effects will be dictated by the nature and scale of the land

cover change. This is also linked to physiographic, climatic and management differences.

Land cover change studies should be viewed as responding to the complex interactions and

62

feedbacks linking social, or indirect and direct biophysical processes that occur on the land

[Dolman and Verhagen, 2003; Maidment, 1993]. To balance the supply and demand for

water resources and to reduce negative or undesired effects for the environment and

society, changes in actual land cover have to be studied at all spatial scales due to the

heterogeneous patterns of the distribution of terrestrial vegetation and soils over land

surfaces [Verburg et al., 1999]. Therefore, establishing clear linkages between specific

land cover changes and hydrological responses at the catchment scale poses a challenging

problem in hydrological studies.

3.2 Land cover change detection

The use of multi-date satellite remote sensing data to detect land cover change began in the

early 1970s [Singh, 1989]. Change detection is the process of identifying differences in the

state of an object or phenomenon by observing it at different times [Singh, 1989]. The

process encompasses the quantification of multi-date imagery to derive changes over two

time periods [Coppin et al., 2004]. It is an important process in monitoring and managing

natural resources and urban development because it provides a quantitative analysis of the

spatial distribution of the population of interest. Change detection is useful in such diverse

applications as land cover change analysis, the monitoring of shifting cultivation, the

assessment of deforestation, the study of changes in vegetation phenology, seasonal

changes in pasture production, damage assessment, crop stress detection, disaster

monitoring, day/night analysis of thermal characteristics as well as other environmental

changes [Singh, 1989].

Techniques to perform change detection with satellite imagery have become numerous as a

result of the increasing versatility of digital data and increasing computing power. A wide

range of approaches to change detection analysis have been reported [Fung, 1990; Green et

al., 1994; Howarth and Wickware, 1981; Jensen, 1986; Jensen et al., 1993; Lillesand et al.,

2004; Milner, 1988; Mouat et al., 1993; Singh, 1989]. Approaches include transparency

compositing [Crapper and Hynson, 1983], image differencing [Dale et al., 1996;

Muchoney and Haack, 1994; Price et al., 1992], image ratio-ing, classification

comparisons [Dale et al., 1996; Jensen, 1995; Muchoney and Haack, 1994], image

enhancement techniques, such as principal component analysis [Fung and LeDrew, 1987],

Normalized Difference Vegetation Index [Mikkola, 1996], image algebra change detection

[Green et al., 1994], and post-classification comparison [Rutchy and Vilchek, 1999].

63

Although these methods have been successful in monitoring change for a myriad of

applications, there is no consensus as to the ‘best’ change detection approach. The type of

change detection method employed is largely dependent on the purpose of the change

investigation, data availability, the geographic area of study, time and computing

constraints, and the type of application [Coppin et al., 2004].

Rather than attempt to review all published change detection techniques, only techniques

that have been selected for use in this study are reviewed, namely image overlay and post-

classification analysis. The purpose of change detection in this exercise is focused on

factors that may influence the behaviour of the hydrological regimes in the catchment.

Image overlay

This change detection algorithm uses a digital enhancement technique for on-screen

change delineation. It provides a simple mechanism to display changes between two dates

of imagery quickly and efficiently [Jensen, 2005]. The image is prepared by making a

photographic two-colour composite showing the two dates in separate colour overlays. The

colours of the resulting image indicate changes in reflectance values between the two

dates. For instance, features which are bright (high reflectance) on date one, but dark (low

reflectance) on date two, will appear in the colour of the first photographic overlay and

vice versa. Features which are unchanged between the two dates will be equally bright in

both overlays and hence will appear as the colour sum of the two overlays.

To represent the bi-temporal variations within one single image product, Alwashe and

Bokhari merged TM bands 2, 4, and 5 of two different acquisition dates via an intensity-

hue-saturation (IHS) transformation. On such an image, vegetation differences showed up

in distinctly different colours [Alwashe and Bokhari, 1993]. Similarly, Sunar [1998]

performed change detection between Landsat TM data of 1984 and 1992 and observed that

areas of no change were represented by values of 127 (mid-grey), while areas that were

darker in 1992 than they were in 1984 had values between 128 and 255. However, this

technique provides the analyst with little information regarding the nature of the change

[Jensen, 2005].

Post-classification analysis

Post classification analysis is the most commonly used method of land cover change

detection. It is a comparative analysis of spectral classifications for times t1 and t2 (or

more) produced independently [Singh, 1989]. Classification results are compared on a

64

pixel-by-pixel basis using a change detection matrix, and the areas of change extracted

[Jensen, 1996; Jensen, 2005; Singh, 1989; Yuan and Elvidge, 1998]. The classified images

are combined to create a new change image classification. The new image indicates the

changes “from” and “to” that took place and the kind of landscape transformations that

have occurred are calculated and mapped. The advantage of this method includes the

detailed “from-to” information that can be extracted. Individual classification of two image

dates minimises the problem of normalizing for atmospheric and sensor differences

between two dates [Jensen, 2005; Singh, 1989] although accurate geo-referencing and

accurate classifications are crucial to ensure precise change-detection results [Augenstein et

al., 1991; Foody, 2001; Rutchy and Vilchek, 1999].

A useful example of the use of the post-classification approach is the work of Munyati

[2000] in which Landsat images acquired in 1984, 1988, 1991 and 1994 were used to

assess change on a section of the Zambian Kafue Flats floodplain wetland system. Similar

supervised Maximum Likelihood classification procedures were employed on all images.

The classified images produced were analysed for change in each land cover category by

overlaying them in a GIS framework, and transition rates between the classes were

calculated. The change detection results provided a reliable indication of the long-term

change that the Kafue Flats wetland area has undergone.

Land cover maps derived from Landsat 5 and Landsat 7 ETM+, discussed in Chapter 2,

form the input data for change detection. Two approaches were used to detect changes in

the upper Shire River catchment, namely image overlay and post-classification analysis

and their results compared. Using image overlay, vegetation characteristics were chosen as

the main indicator of land cover change. Post classification statistical analysis was

employed to determine the specific nature of changes between dates of imagery (1989 and

2002) in each land cover class. Image overlay was found to detect areas of change and no

change within the catchment. By using post classification procedures, areal statistics and

the direction of change in each classified land cover class were derived. A combination of

both methods reveals the significant changes that have occurred in the Shire River

catchment between 1989 and 2002. These activities also highlighted areas where there are

major changes in land cover (i.e. "hot spots"), in both temporal and spatial aspects.

65

3.3 Methodology

3.3.1 Input for change detection

Land cover maps derived and discussed in Chapter 2 form the basis of data for the land

cover change detection.

3.3.2 Approaches

Image overlay

In this approach, vegetation reduction was chosen as the main indicator of land cover

change. Therefore, band 4 of Landsat TM and ETM+ (near-infrared) respectively were

selected. For the 1989 image, band 4 was displayed through the R and B (blue) channels of

the computer monitor. Band 4 of the 2002 image was displayed through the G (green)

channel of the computer monitor. This technique is not intended to be quantitative but

rather was used to identify qualitatively and explore the areas of change in the Shire River

catchment.

Post-classification

The starting point for the post-classification approach is a pair of images for which land

cover classification and labelling have been carried out, in this case the Landsat TM and

Landsat ETM+ images discussed in Chapter 2. The next step is to compare the resultant

classification images on a pixel-by-pixel basis and to extract areas of change using a

change detection matrix. A change detection matrix is an {N x N} table of change

detection classification codes produced by assigning a unique integer to each possible

change from one land cover category to every other category, where N is the number of

thematic land cover classes in the study.

For this study, eight land classes were identified. Each of the eight identified land classes is

assigned a base value in the sequence {20, 40 …160}. A 64-cell table (8 x 8) of change

detection codes, was created for each possible pair of “changed from” and “changed to”

combinations by incrementing a base value by unity, e.g. from the base value 20 assigned

to built-up areas, the vector {21, 22 …28} is created. The full set of assigned codes is

presented in Table 10. A change from savanna shrubs to built-up areas, for example,

would be represented by the code 26 shown in column 1. Diagonal elements of the matrix

are assigned to pixels that have not undergone change. Certain changes are physically

improbable, e.g. open water to woody closed, but values are assigned to complete the

matrix.

66

By applying these change detection codes to the comparison of the two classified images, a

new layer is produced, the post classification change layer, with each pair of pixels

represented by a change detection code. From the change detection layer, change detection

maps may be produced to visualise areas of change by assigning colours to single or

grouped categories of change. For example, any pixel in the 1989 map that changed to

built-up area by 2002 is assigned the value red; any pixel that changed into cultivated or

grazing land by 2002 is assigned the value pink. The assigned colours to be used in the

results section are superimposed on the matrix shown in Table 10.

Table 10: Assignment of Change Classification Codes of land cover for the Shire River

Catchment for 1989 and 2002

Built-up areas

Cultivated/grazin

g

Fresh Water

Grass-lands

Marshes Savanna shrubs

Woody open

Woody closed

To 2002

From 1989 20 40 60 80 100 120 140 160

Built-up areas 21 41 61 81 101 121 141 161

Cultivated/grazing 22 42 62 82 102 122 142 162

Fresh water 23 43 63 83 103 123 143 163

Grasslands 24 44 64 84 104 124 144 164

Marshes 25 45 65 85 105 125 145 165

Savanna shrubs 26 46 66 86 106 126 146 166

Woody open 27 47 67 87 107 127 147 167

Woody closed 28 48 68 88 108 128 148 168

� no change in land cover between dates � new built-up areas � land cover change to cultivated or grazing areas � water related changes not selected for display � water � change in land cover to vegetated areas � vegetation regeneration from previously built-up areas � land cover change to grassland areas � change in land cover to vegetated areas

This table is interpretable by presenting tabular results quantifying change between classes,

such as the change from woody open to cultivated or grazing land. The table shows the

degree of change, the pixel prior classification, and the pixel post classification. This

matrix can be used to distinguish the entire change analysis in one view and is thus a

powerful tool for examining landscape change.

67

Land cover change areal statistics were also summarised to express the net land cover lost

and gained by each class. This was accomplished by subtracting the total amount of land

cover gain from the total amount of loss. Statistics were compiled in the Microsoft Excel

programme to determine the specific nature of changes between dates of imagery, i.e. the

size of the differences between the dates and the direction (positive and negative) in each

land cover type.

3.4 Results and discussion

3.4.1 Image overlay

The visual overlay composite image (Figure 21) indicates visual changes in land cover

change that took place in the upper Shire River catchment between 1989 and 2002. This

image highlights the changes in vegetation and qualitatively assesses the extent of the

change. This technique showed clearly areas of no change, and areas where changes have

taken place, in the upper Shire River catchment. The resulting image is grey in areas where

no change has occurred, such as to the west of the Shire River in the Liwonde National

Park. Where vegetation has decreased, the image is magenta/purple and where vegetation

has increased, it is green.

In Figure 21, the magenta regions represent vegetation that has been cleared for new

cultivation and settlements. Major vegetation reductions are apparent within Mangochi

Township near the outlet of the Shire River (Figure 22).

Mangochi is one of the fastest growing towns in Malawi due to the development of tourism

around the Shire River and Lake Malawi [National Statistical Office, 2000]. Its population

had increased from 496 578 in 1987 to 599 935 in 1998 as captured by population censuses

[National Statistical Office, 2000]. The high population growth has translated into rapidly

increasing demands from land in terms of food, shelter, energy (in particular, fuelwood)

and construction materials. Disappearing vegetation cover lessens the landscape's ability to

intercept, retain and transport precipitation. Instead of trapping precipitation, which then

percolates to groundwater systems, deforested areas become sources of surface water

run-off, which moves much faster than subsurface flows.

68

Figure 21: Image overlay for the Shire River catchment: 1989 — 2002

69

Mangochi TownshipMangochi Township

Figure 22: Expansion of Mangochi Township into previously vegetated areas

Another area of interest is the riverline below Lake Malombe, which shows a decreasing

trend in marshes, in contrast to an increase of marshland above the lake. Marshes are

critical because they recharge groundwater supplies and moderate streamflow. This is an

especially important function during periods of drought. The presence of marshes in a

catchment helps to reduce damage caused by floods by slowing and storing flood water. It

was verified during the field exercise (Section 2.3.2) that the grass-covered marshes have

either been drained and cultivated or overgrazed. These agricultural activities are part of an

initiative by Save the Children (UK), a non-governmental organisation, as a means of

increasing food security in the area.

In some parts of the south-eastern and north-western side of the catchment, some degree of

vegetation recovery is depicted. These areas were previously occupied by Mozambican

refugees during the civil war between 1976 and 1992. After the civil war ended, most

refugees returned to Mozambique. The harvesting of natural forest products by the

occupants of these refugee camps resulted in significant loss of forest area and increase in

fragmentation as captured by the 1989 image. By 2002, vegetation re-growth can be

observed in most of what were previously refugee camps. The green regions represent

70

higher reflectance in the near-infrared wavelength signifying vegetation re-growth through

natural processes and government intervention through reforestation programmes. At the

initial stage of the reforestation programmes, exotic trees such as bluegum and gmelina

were planted. However, water resources were being compromised due to these fast

growing species. The Malawi Government and NGOs working in this area, now

rehabilitate the camps with indigenous trees [Kafakoma, 1996]. As a result, the presence of

vegetation can change the quantity of water on the surface, in the soil or groundwater

recharge. This in turn affects run-off rates and the availability of water for either ecosystem

functions or human services.

3.4.2 Post classification and land cover change areas

Results from post-classification analysis are presented using a series of change maps for

visualisation, and statistical tables to provide quantitative measures of change. The change

maps use the colour coded from-to indices of change defined in Table 10.

Change Map

A map of overall land cover changes that have occurred within the Shire River catchment

is shown in Figure 23. The change information is overlaid onto the 2002 Landsat natural

colour image {3, 2, 1} for orientation purposes. Shades of bright green depict areas that did

not change between 1989 and 2002. Areas that have changed to cultivated or grazing lands

are shown in purple. Changes from other types of land cover to vegetated areas are

displayed in yellow and green. This map confirms the qualitative changes indicated in the

change overlay image of Figure 21.

71

no land cover change

� new built up areas

� change cultivated or grazing areas

�water related changesnot selected for display

??� water

� change to vegetated areas

vegetation regeneration from previously built up areas

� change to grassland areas

� change to vegetated areas

no land cover change

� new built up areas

� change cultivated or grazing areas

�water related changesnot selected for display

??� water

� change to vegetated areas

vegetation regeneration from previously built up areas

� change to grassland areas

� change to vegetated areas

no land cover change

� new built up areas

� change cultivated or grazing areas

�water related changesnot selected for display

??� water

� change to vegetated areas

vegetation regeneration from previously built up areas

� change to grassland areas

� change to vegetated areas

Figure 23: Post classification change map of the Shire River catchment between 1989 and 2002

Overall change statistics

Changes in each land cover category over the thirteen-year period 1989 to 2002 are

summarised in tables (Table 11 to Table 17) and discussed in this section. Table 11 shows

the net gains and net losses for the various land cover categories in hectares and

percentages. There are significant increases in the spatial extent of grasslands, cultivated

or grazing, and of woody closed classes, while savanna shrubs, woody open and built-up

areas decreased in extent. Unsurprisingly, the analysis shows insignificant changes in the

extent of fresh water and marshes.

72

Table 11: Land cover changes of the Shire River catchment during 1989 to 2002

Total area in land

class at 1989 ( ha )

Total area changed to per class

(ha)

Total area changed from per land class

(ha)

Net gain (loss) (ha)

Net change (%)

Built-up areas 39 813 456 2 005 -1 549 -3.9

Cultivated or grazing

95 428 4 622 3 158 1 464 1.5

Fresh water 38 353 0 108 -108 -0.3

Grasslands 15 127 3 616 469 3 147 20.8

Marshes 6 444 108 180 -72 -1.1

Savanna shrubs 152 791 2 584 4 370 -1 786 -1.2

Woody open 70 612 1 312 3 017 -1 705 -2.4

Woody closed 37 930 1 652 1 043 608 1.6

Total area 456 498 14 350 14 350 0

Changed areas for each land cover type are calculated by summing the respective numbers

of from-to change pixels occurring in the change detection matrix for each change

classification code (Table 10). Reference is made to the changed land area of the

catchment (excluding water bodies) amounting to 14 350 ha. The entire catchment

excluding open water bodies is 418 145 ha, thus the total changed area represents 3.43% of

the land area of the catchment. The following sections discuss changes within each land

cover class.

Cultivated or grazing areas

The total land area within the upper Shire catchment that was converted from various

classes to cultivated or grazing land amounted to 4 622 ha (Table 12). The increase in

cultivated or grazing land occurred mainly at the expense of savanna shrubs, woody open

areas and built-up areas: cultivated or grazing areas expanded by 1 671 ha (1.0%) from

previously savanna shrubs region; and 1 359 ha (1.9%) from previously woody open areas.

Surprisingly, there is a significant change from built-up areas (1 156 ha) to cultivated or

grazing land.

73

Table 12: Areas changed into cultivated or grazing areas between 1989 and 2002

From Total area in land class at 1989

( ha )

Area changed to cultivated ( ha )

% change

per class♣♣♣♣

Built-up areas 39 813 1 156 2.9

Grasslands 15 127 153 1.0

Marsh 6 444 24 0.4

Savanna shrubs 152 791 1 672 1.0

Woody open 70 612 1 359 1.9

Woody closed 37 930 259 0.7

Total 322 717 4 622 1.4

♣ Percentages calculated as fraction of the original area of the class.

The expansion of cultivated or grazing land indicates increased subsistence agriculture.

Savanna shrubs have experienced degradation and exploitation due to cultivation and

demand for wood resources. The rich calcimorphic alluvial soils render the savanna more

vulnerable to agricultural expansion. Most of the soils in the Shire rift valley are of alluvial

origin, rich in nutrients and ideal for agricultural production. The 2002 land cover map

shows that most of the remaining savanna shrubs are found in the Liwonde National Park

(protected area) and south east of Lake Malombe, which is sparsely populated (Figure 24).

A study of the two classification images shows that cultivated or grazing areas are also

expanding from the lowlands into the higher areas predominantly occupied by woodlands.

These processes are closely related to increases in the demand for food production.

Inevitably, cultivation and encroachment for exploitation of wood for construction and

fuelwood has spread onto such areas as can be noted from the 1989 compared to the 2002

land cover map. This has given rise to a transition from closed savanna woodlands to open

and sparse savanna, and it is mostly economic shrubs that are left in the fields (Figure 24).

74

1989

2002

1989

2002

Figure 24: Expansion of cultivated or grazing areas into predominantly savanna areas

An interesting result is the loss from built-up areas amounting to 1 156 ha which was

converted to cultivated or grazing land. Overall, this represents a net loss of -1 549 ha

(-3.9%) into other land cover categories (Table 11). Built-up areas (especially grass-

thatched) and cultivated or grazing areas occur in similar environments, often in adjacent

or mixed stands and have similar spectral signatures. The decrease could be an aspect of

misclassification due to similarity between roofing materials and the surrounding pixels

dominated by shallow soils and rock outcrops.

Additionally, the ravages of civil war in Mozambique had left large scars on the landscape,

particularly in the forest areas of the eastern mountains, which form the international

boundary between Malawi and Mozambique (Figure 25). Some bare patches where once

there were refugees camps have not yet been reforested, as can be seen in the 2002 image.

These bare patches from dismantled refugee camps are indistinguishable from fallow

cultivated fields and have probably been classified as change to cultivated or grazing areas.

75

Declining forest cover poses serious threats to water supplies within the area. A cause of

concern for this trend in land cover conversion is the significant role the forest reserves

play in catchment management. The forest reserves, Namizimu forest and Machinga hills

(Figure 20) serve some of the big rivers flowing into the Shire River. Overexploited

forestlands are threatened by accelerated run-off and soil erosion, and consequently

degrade the agricultural productivity of the land. Unless monitored, this trend may impact

negatively on the livelihoods of people and the survival of other natural resources in this

area, including water.

Machinga Forest

Namizimu Forest

Forest fragments Machinga Forest

Namizimu Forest

Forest fragments

Figure 25: Forest fragmentation around forest reserve areas

The predominant subsistence agricultural practices signify a gradual but chronic

degradation of the landscape along the valleys and adjoining slopes. These findings concur

with the observations by Kalipeni [1996], which indicates that pressures on land are

increasing, resulting in cultivation of marginal land, more fragile soils and steeper slopes.

Grasslands

From the post-classification land cover change results, 3 616 ha of land was lost from other

classes and converted to grasslands (Table 13). Taking into account losses of grasslands to

other types, the results in a net gain of 3 147 ha (+20.8%) in the extent of grasslands

76

(Table 11). High proportions of change originated from savanna shrubs (1 368 ha) and

cultivated or grazing lands (1 073 ha). Other changes include transition from woody open

land (553 ha), built-up areas (439 ha) and woody closed (158 ha). This is a considerable

change, in view of the short period over which the land cover changes have occurred.

Table 13: Areas changed into grassland areas between 1989 and 2002

From Total area in land class at 1989

( ha )

Area changed to grasslands

( ha )

% change

per class♣♣♣♣

Built-up areas 39 813 439 1.1

Cultivated or grazing 95 428 1 073 1.1

Marshes 6 444 25 0.4

Savanna shrubs 152 791 1 368 0.9

Woody open 70 612 553 0.8

Woody closed 37 930 158 0.4

Total 403 018 3 616 0.9

♣ Percentages calculated as fraction of the original area of the class.

The increase in grasslands from previously woody closed, woody open and savanna shrub

areas is a result of grass thriving best without canopy cover. Increasing demand for wood

resources places substantial strain on landscape integrity. However, the loss of built-up

areas to grassland could be a result of misclassification where the rural building materials

are from dry grass yielding similar spectral response to the natural grassland. In addition,

the surrounding pixels are made of bare soil, which increases the uniformity of spectral

appearance. As a result, spectral classification would assign all pixels including

grasslands, built-up areas and bare soil as a single class.

Changes in fire frequency and timing are known to transform vegetation structure and

composition. Frequent late dry season fires eventually transform woodland into open

grasslands with only isolated, fire-tolerant canopy trees, scattered understory and shrubs

[Desanker et al., 1997]. As the landscape becomes altered, run-off velocity and flow rates

in river systems are also affected.

Savanna shrubs

By analysing the changes that occurred to the landscape, a total of 2 586 ha (0.9%) was

converted to savanna shrub (Table 14). This increase does not compensate for the loss of

77

savanna shrubs to cultivated or grazing areas as discussed above. However, it could be a

result of misclassification as savanna shrubs and cultivated or grazing areas occurring in

similar environments.

Table 14: Areas changed into savanna shrubs areas between 1989 and 2002

From Total area in land class at 1989 ( ha )

Area changed to savanna shrubs ( ha )

% change

per class♣♣♣♣

Built-up areas 39 813 217 0.5

Cultivated or grazing 95 428 1 373 1.4

Grasslands 15 127 267 1.8

Marshes 6 444 32 0.5

Woody open 70 612 377 0.5

Woody closed 37 930 320 0.9

Total 265 354 2 586 0.97

♣ Percentages calculated as fraction of the original area of the class.

Built-up areas

Spatial extents in built-up areas show an overall subtle increase in settlements. Changes in

this land cover category are shown in Table 15 depicting higher land cover changes from

woody open areas (127 ha), cultivated or grazing areas (187 ha) and savanna shrubs

(87 ha). A detailed examination of built-up areas category across the catchment revealed

four important subclasses of change, which are discussed below.

Table 15: Areas changed into built-up areas between 1989 and 2002

From Total area in land class at 1989 ( ha )

Area changed to built-up areas ( ha )

% change

per class♣♣♣♣

Cultivated or grazing 95 428 187 0.47

Grasslands 15 127 28 0.18

Marshes 6 444 7 0.11

Savanna shrubs 152 791 87 0.06

Woody open 70 612 123 0.17

Woody closed 37 930 25 0.06

Total 378 322 457 0.12

♣ Percentages calculated as fraction of the original area of the class.

The spatial distribution of change in built-up areas is of particular interest in this thesis.

Most of the emerging towns, for example Liwonde Township, Mangochi Township and

Balaka Township portray an increase in built-up area signifying an aspect of rural-urban

78

migration (Figure 26). With the national annual urbanisation growth rate estimated at 6.7%

in 1998, many rural dwellers are migrating to urban and expanding district centres in

anticipation of wage employment [National Statistical Office, 2000]. There is a sparse

distribution of spontaneous rural settlements on the lower escarpments, across valleys and

flat lands. With the increasing population, slight expansions into the higher escarpment are

visible on the change image.

Mangochi Township

Built up areas

Mangochi Township

Built up areasBuilt up areas

Figure 26: Built-up areas expanding around Mangochi Township

A decrease in built-up areas is apparent in areas especially formally occupied by

Mozambican refugees. Most of the refugee camps were located to the north western part of

the study area around Mangochi district and Namwera Township. This area is close to the

border with Mozambique such that in 1989 during the civil war most refugee camps had

occupied this area. However, by 2002 the civil war was over and upon the repatriation of

Mozambican refugees, the Malawi Government embarked on land rehabilitation at the

79

former refugee camps, which has resulted in the slight increase of woody closed areas

(Figure 27).

Woody closed areasWoody closed areas

Figure 27: Increase in woody closed areas

Another note of significance is the effects of the electric fence constructed around Liwonde

National Park, a government reserved area, between 2002 and 2004. During the time of the

satellite observation in 1989, there were built-up areas (settlements) close to Liwonde

National Park. Due to prowling animals from the park, many settlements were abandoned,

leading to the regeneration of savanna shrubs as seen in the 2002 image (Figure 9). Field

observations and consultation with members of the surrounding community (in 2006)

indicated that since the containment of animals by the game fence, people have started

moving back to areas close to the park boundary [Palamuleni et al., 2008]. The area is

endowed with calcimorphic alluvial soils, which are very attractive for subsistence farming

(Figure 28).

The boundaries of the reserved area are assumed not to have changed, though subtle

increases in built-up areas do occur within the reserve. These negligible changes are due to

80

the expanding tourism industry within Liwonde National Park, which are visible in Figure

28. New built-up areas along the rivers are primarily composed of tourist resorts and camp

sites.

Liwonde National

Park

Tourist resorts and camp sites

Built up areas expanding outside Liwonde National

park

Tourist resorts and camp sites

Built up areas

Liwonde national park boundary

Liwonde National

Park

Liwonde National

Park

Tourist resorts and camp sites

Built up areas expanding outside Liwonde National

park

Tourist resorts and camp sites

Built up areas

Tourist resorts and camp sites

Built up areas expanding outside Liwonde National

park

Tourist resorts and camp sites

Tourist resorts and camp sites

Built up areas expanding outside Liwonde National

park

Tourist resorts and camp sites

Built up areas

Liwonde national park boundary

Figure 28: Increase in built-up areas and tourism expansion around Liwonde National Park

According to the image classification, the size of built-up areas has decreased between

1989 and 2002 with a net loss of -1 549 ha (-3.9%) into other land cover categories (Table

11). The decrease could be a result of misclassification because the population increased

from 1 011 843 in 1987 to 1 218 177 in 1998 as recorded by population censuses [National

Statistical Office, 2000] and population increases result in the expansion in built-up and

cultivated areas. The misclassification may be a result of local building materials, which

include grass (thatching) and sand (hand made clay bricks) such that built-up areas and

cultivated or grazing areas where there is little green grass and significant exposed soils

are spectrally identical. During summer, most of the areas are non-vegetated with exposed

soils due to ploughing by farmers in preparation for the next crop season. This condition

yields similar spectral values to those of cultivated areas; grass thatched houses and

81

exposed bare areas. Another plausible explaination could be increase in cluster and

nucleated settlement pattern as opposed to scattered settlements.

Percentage of the catchment surface that is impermeable due to settlement and road

surfaces influences the volume of water that runs and increases the amount of sediment

that can be moved [Arnold and Gibbons, 1996]. Generally, such trends in the landscape

could significantly alter the hydrology of the system.

Woody open

A total land area of 1 312 ha (0.38%) was converted from other land cover categories to

woody open land (Table 16). The transitions to woody open areas originated from land

areas previously occupied by marshes (455 ha), cultivated or grazing areas (377 ha),

woody closed areas (281 ha) and built-up areas (166 ha). Although this was a positive

trend, the extent of the transition from for example cultivated or grazing areas (377 ha) to

woody open areas was negligible in comparison to the 1 359 ha loss of woody open areas

to cultivated or grazing areas.

Table 16: Areas changed into woody open areas between 1989 and 2002

From Total area in land class at 1989

( ha )

Area changed to woody open

( ha )

% change

per class♣♣♣♣

Built-up areas 39 813 166 0.42

Cultivated or grazing 95 428 377 0.39

Grasslands 15 127 12 0.08

Marshes 6 444 455 0.30

Savanna shrubs 152 791 21 0.33

Woody closed 37 930 281 0.74

Total 347 533 1 312 0.38

♣ Percentages calculated as fraction of the original area of the class.

Classification errors could be attributed to this transition, which may not be a true

indication of the actual trends in this land cover category. Marshes, woody closed and

woody open areas have similar spectral characteristics as they both have high chlorophyll

during the dry season. The woody trees: Zizyphus jujube (masau), Mangifera indica

(mango) and Faidherbia albida (msangu) are in leaf during the dry season whereas

marshes are mostly evergreen due to presence of water throughout the year.

82

Woody closed

There was a slight increase of 1 652 ha (0.48%) in the area of the woody closed areas

(Table 17). A unique aspect observed in this land cover class is the reclamation of areas

previously occupied by Mozambican refugees and reforestation of school compounds. This

process is illustrated by regeneration from fallow fields especially those areas formerly

occupied by the Mozambican refugees in 1989 (Figure 27). Additionally, annual tree

planting exercises involve the planting of both indigenous and exotic trees such as

Eucalyptus globulus, Gmelina aborea and Acacia nigrescens in woodlots around schools.

As a result, there is a net gain of 608 ha (+1.6%) in woody closed areas across the

catchment area (Table 11). The woodlots had a similar spectral reflectance to the woody

closed areas as verified during field data collection and site context (Section 2.4.5).

Table 17: Areas changed into woody closed areas between 1989 and 2002

From Total area in land class at 1989

( ha )

Area changed to woody closed

( ha )

% change

per class♣♣♣♣

Built-up areas 39 813 28 0.1

Cultivated or grazing 95 428 149 0.2

Grasslands 15 127 10 0.1

Marshes 6 444 70 1.1

Savanna shrubs 152 791 789 0.5

Woody open 70 612 606 0.9

Total 347 533 1 652 0.5

♣ Percentages calculated as fraction of the original area of the class.

The proportion of rainfall that directly reaches the ground surface is affected by the

coverage of vegetation cover. Tree root systems hold the soil together and thus slow the

rate of run-off and reduce erosion. Trees also absorb water during the rainy season.

Because of the good ground cover, surface run-off is retarded and water can infiltrate the

topsoil, leading to high levels of infiltration and recharging of the aquifer. The positive

change in the extent of woody closed areas signifies ecological sustainability for water

resources management.

Although the increase in woody closed areas is a positive development in relation to

hydrological processes, this amount may be somewhat misleading because of the presence

of shadows in some parts of the images, which created difficulties in image classification.

83

It is possible to misclassify topographic shadows and woodlands as marshes. The

absorption effects of water in the marshes are spectrally similar to the lack of light

collected from the shadowed areas, while the strong reflection from the surrounding semi-

deciduous Miombo woodlands and the evergreen Eucalypts from woodlots exhibit spectral

similarity with the marshes. Most likely, misclassifications were expected. Nonetheless,

these areas have been identified in the DEM data for the region and they do not constitute a

large area. This potential misclassification is a result of the inherent limitations of the

technology and ancillary data at hand since during pre-processing no shadows were

removed. For instance, the 70 ha of marshland converted to woody closed area may not be

an accurate indication of trends in land cover change. The increase in the area classified as

woodlands during the period 1989-2002, however, is a positive development.

Marshes

The size of the marshy areas appears to be increasing, although it is possible non-marsh

areas were misclassified. However, water bodies do change in response to climatic

conditions and records of rainfall distribution in the Shire River catchment depict

continuous decreasing variations [Malawi Government, 1999]. As such, the water levels

along the shores of the river and the lake have been decreasing giving rise to waterlogged

areas, hence the increase of 108 ha.

Because marshes retain water, they support the vigorous growth of grass and provide good

dry season grazing areas when other forms of grazing are in short supply. Marsh margins

are also used for gardens (during the dry season) providing a more reliable crop output to

supplement rainfed harvests. However, marshy areas act as hydrological stores, holding

water and releasing it as base flow to the headwater streams during the dry season.

Continuous water extraction for dry season cultivation may be subject to uncertainty in the

variability of seasonal river recharge.

Fresh water

The changes in the extents of fresh water and marshes are insignificant, possibly due to

rainfall variations between the two time periods. The 1989 image was preceded by a La

Nina episode while the 2002 it was strong La Nina episode [SADC: Drought Monitoring

Centre]. In the 1989 image some very small dams were mapped which appear to have

dried up or have significantly reduced in surface area during the peak of the dry season in

2002.

84

3.5 Conclusion

Scientific approaches and holistic decisions regarding land management are required for

the sustainable development of the Shire River catchment. Change detection techniques

using temporal remote sensing data provide detailed information for detecting and

assessing land cover dynamics. Landsat classifications were used to produce accurate

landscape change maps and statistics. Two change detection techniques (multi-date image

overlay and post classification analysis) were applied to monitor land cover changes based

on data from the two dates (1989 and 2002). The result of this study documents a

significant combination of deforestation and fragmentation of forests in the study area.

General patterns and trends of land cover change in the Shire River catchment depict

transition towards degradation of woodlands and an increase in patches within the forest.

The change analysis showed an increase of 4 622 ha (7.9%) in cultivated or grazing lands

emanating from declining natural vegetated areas. The main driving factor is subsistence

agricultural expansion and demand for wood resources as evidenced by current agricultural

practices and population growth. As agriculture continues to play a dominant role in land

cover conversion and degradation from Brachystegia woodlands to more open and dry

vegetation formations will continue to evolve. To this end, it is apparent that the rapid

increase in cultivated areas within the catchment will not only decrease the amount of

forests and vegetated areas but also increase run-off potential thereby diminishing the

overall quality and quantity of water resources. Distinguishing and quantifying where

potentially risky changes occur is critical to the initiation of regular monitoring of

resources and the environment in general.

The patterns of land cover change observed in this study do not provide evidence of

impacts of industrialisation and commercialisation in the catchment over the last decade.

The strongest signals of landscape change detected thus far appear to correspond to an

apparent expansion of subsistence agricultural practices and changes in the amount and

extent of forest resources. Hence, the rapid growth in population, coupled with an absence

of industrialisation, will continue to put severe strains on the sustainability of

environmental resources.

Change data can be used to generate information tied to geographic coordinates and thus to

update of maps. The selected algorithms used in this study allow for the efficient thematic

updating of land cover change maps vis-à-vis land cover information for the Shire River

85

catchment and other areas in Malawi. The maps can also be used to estimate the rates of

land cover change. Recent information about land cover variables and the nature of the

transformation of land cover can provide a valuable guide for formulating appropriate

policies and for the effective implementation of programs for natural resource allocation,

land husbandry, conservation, management, sustainable use and combating deforestation.

Results from this study have demonstrated that satellite remote sensing approaches provide

a cost-effective alternative when more information is needed, but financial resources are

limited. Land cover change information generated from this study could serve the

requirements of the Department of Environmental Affairs in Malawi in their goal of

characterising the existing conditions of the Shire River catchment, both as a baseline for

later research and as the starting point for the development of future scenarios. Overall,

considering the present scale of temporal and spatial land cover change in the Shire River

catchment, a more continuous and comprehensive land cover change monitoring system is

required.

Land cover change information derived from remotely sensed data discussed in this

chapter will be used to examine the relationship between land cover variables and

hydrological regimes in the upper Shire River catchment. Chapter 4 provides an analysis of

the hydrological response from land cover activities.

86

Chapter 4

4 HYDROLOGICAL MODELLING BASED ON THE LAND COVER

ANALYSIS

Chapter 4 integrates the findings of the previous chapters into a conceptually based

spatially distributed ArcView Soil and Water Assessment Tool eXtendable version

(AVSWATX) hydrological model. This chapter presents the calibration, validation, and

application of the AVSWATX model for predicting the catchment hydrological

responses to changes in land cover in the Shire River catchment. Land cover change

simulations and scenarios demonstrate the model’s ability to integrate spatially distributed land cover change and precipitation events into output responses – in the

form of run-off from the catchment and total water yield at the outlet. Model

predictions are compared to observed hydrological records. The AVSWATX model provides a new approach for evaluating relative land cover changes across catchment

landscapes.

4.1 Introduction

4.1.1 Land cover and hydrological processes

Land surface heterogeneity is characteristic of many regions of the world [Lahmer et al.,

2001]. It plays an important role in partitioning incoming radiation at the land surface into

latent and sensible heat, and in partitioning precipitation into percolation, run-off and

evaporation [Calder, 2002]. At the larger scale typified by orographic features such as

mountains and valleys, topographic variation directly affects precipitation and surface

temperature. Precipitation and surface temperature are the main drivers of surface

hydrology through their effects on local and regional atmospheric circulation, the vertical

distribution of atmospheric moisture and temperature, and condensation [Maidment, 1993].

At a smaller scale, topographic variation can modify surface and subsurface run-off

through down-slope redistribution of soil water.

Precipitation and land hydrological processes maintain the water balance in a river basin.

Land surface performs a role in the hydrological cycle, as water availability is generally a

consequence of precipitation redistributed into evaporation, run-off and soil moisture

storage [Dolman and Verhagen, 2003]. The majority of precipitation must pass over the

land surface or drain through the soil and bedrock to translate into river flows. The spatial

heterogeneity associated with land cover, soil properties and localized precipitation

influences soil moisture and surface fluxes. Land cover change and the effects of land

management on the hydrological response of a catchment are most likely where the change

87

alters the surface characteristics of a basin. The degree and type of land cover influences

surface run-off and the rate of infiltration, and consequently the rate of ground water

recharge [Calder, 1992; Shaw, 1990]. Changes in these hydrological variables may have

implications for water resources.

Studies of the relationships between changes in land cover, environmental change, the

amount of run-off and percolation at the landscape scale can be used to compare

catchments, identify at-risk communities, and aid management attempts to limit undesired

effects. Hydrological models predict hydrological responses (streamflow, run-off,

percolation) in response to various inputs (precipitation timing and intensity, landscape

surface characteristics, vegetation cover change, and surface temperature), and in so doing

facilitate the understanding of hydrological events [Brooks et al., 1991].

4.1.2 Hydrological modelling approaches

Models are increasingly used in hydrology to simulate changes in catchment management,

to extend datasets and to evaluate the impact of external influences (such as climate

change) [Dingman, 2002]. A model is a conceptualisation of a real system that retains the

essence of that system for a particular purpose [Maidment, 1993]. Every model is an

attempt to capture the complex nature of hydrological processes, but it is important to

recognise that this conceptualisation involves a considerable degree of simplification.

Hydrological models can be divided into two main types according to how they treat the

spatial component of catchment hydrology: lumped and distributed models. The main

difference between these two groups of models is that the lumped models do not take

account of the spatial distribution of physical data of the basin (e.g. soil, land cover,

topography) or of the spatial variation of the climate (e.g. precipitation, evaporation), while

distributed models do. The lumped models have the advantage that they are easier to

operate and require less data than distributed models. However, they can only be applied to

basins with measurements and require long-term historical data for calibration [Maidment,

1993]. Distributed models require a great deal of detailed data concerning the basin and

have a large number of parameters to optimise. These models are spatially distributed since

they have been developed to incorporate the spatial patterns of terrain, soils, and vegetation

as estimated using remote sensing and Geographical Information Systems [Famiglietti and

Wood, 1994; Star et al., 1997; Wigmosta et al., 1994]. The spatial variation of data in these

types of models is represented by sub-basins or grids.

88

There are several studies of large-scale hydrological model applications in tropical regions

[Andersen et al., 2001; Bormann, 2005; Molicova et al., 1997; Ndomba, 2007; Ndomba et

al., 2005; Perrin et al., 2001]. In the case studies mentioned (Table 18), the objective of

modelling was to approximate the distribution and movement of water over land,

underground and in-stream and sediment yield. The models were able to estimate the

quantity of water stored in the soil and in natural bodies and the exchange between the two.

In addition, the models estimated changes in rates and quantities over time. Soil and Water

Assessment Tool (SWAT) model, a GIS-based distributed hydrological model based on the

ArcView GIS software platform, is one of the more widely used models [Kepner et al.,

2004; Miller et al., 2003; Ndomba, 2007; Ndomba et al., 2005].

Table 18: Examples of large-scale hydrologic model applications

Hydrologic model River basin Scale (km²)

Year Reference

Topmodel Sinnamary, French Guiana

15 000 1997 Molicova et al.

19 daily hydrological model Brazil 50 600 2001 Perrin et al.

MIKE SHE Senegal River, Senegal 375 000 2001 Anderson et al.

UHP Queme, Benin 14 000 2005 Bormann et al.

SWAT2000 Simiyu, Tanzania 10 659 2005 Ndomba et al.

SWAT2000 Pangani, Tanzania 7 280 2007 Ndomba et al.

The SWAT model has been employed to evaluate the effects of historic land cover change

on catchment response of the Upper San Pedro Basin, USA [Miller et al., 2002]. Simulated

catchment response in the form of run-off volume, peak run-off rate, and total sediment

yield were used as indicators of catchment condition. Hydrological modelling results

indicated that catchment hydrological response in the Upper San Pedro Basin has been

altered to favour increased average annual surface run-off due to land cover change during

the period from 1973 to 1997, and consequently it is at risk for decreased water quality and

related effects upon the local ecology.

ArcView Soil and Water Assessment Tool eXtendable version (AVSWATX) is an enhanced

version of the earlier SWAT model. Several useful results have been demonstrated in

previous work [Arnold and Fohrer, 2005; Kepner et al., 2004; Miller et al., 2003; Nedkov

and Nikolova, 2006]. AVSWATX model was used to assess the floods hazard in Yantra

river basin, Bulgaria and evaluate the influence of landscape changes on the hazard

89

[Nedkov and Nikolova, 2006]. The results show that particular changes to land cover in the

basin could increase the flood hazard in some areas. These could be used to make

recommendations for management measures directed to reduce the damages caused by

floods. The model was used with success in several other basins worldwide, primarily in

the United States and in many European countries, the Motueka basin (2 075 km²) in New

Zealand [Cao et al., 2003], the Alban Hills basin (1 000 km²) in Central Italy [Benedini et

al., 2003], and the Celone Creek basin (24 072 km²) in Italy [Pappagallo et al., 2003].

Similar successful applications of AVSWATX have been reported from Africa although

researchers have questioned the applicability of complex models such as AVSWATX to

regions with limitations such as data scarcity, arguing that such models offer too many

parameters [Ndomba, 2007]. AVSWATX was applied to the Tafna wadi basin, which

represents a major potential water resource for western Algeria [Yebdri et al., 2007]. The

purpose of modelling was to provide data required for sustainability in water resources

management. The results of this application demonstrate that the model reproduces and

generates the correct climatic variables and produces accurate water resources assessment

in the basin. Related results have been reported in the case of SWAT2005, which was

applied successfully to the Upper Ouémé catchment in Benin (~14 500 km2) to quantify

water and sediment yield [Busche et al., 2005]. In West Africa, the AVSWATX model was

applied to a four million km2 area which included Niger, Senegal and Volta River basin

[Schuol and Abbaspour, 2005]. Similarly, SWAT was used in the Hare River watershed,

Southern Rift Valley Lakes Basin, Ethiopia to investigate land use and land cover

dynamics and their effects on streamflow [Tadele and Förch, 2007].

Previous research on rainfall run-off modelling in Pangani River basin in the North-Eastern

part of Tanzania has applied complex models (i.e. physically based and distributed models)

such as SWAT2000 and Landpine [Birhanu, 2005; Rohr, 2003]. The SWAT2000 model

has been used in the Simiyu sub-basin in Lake Victoria, Tanzania [Mulungu and Munishi,

2007]. Use of GIS and remote sensing were found to be helpful tools to detect and analyse

spatio-temporal land use and land cover dynamics. SWAT2005 was useful for analyzing

the impacts of land use and land cover changes on streamflow as it provides an accurate

hydrological performance model.

The above studies demonstrate the successful application of the SWAT model in areas with

limited data availability. SWAT can be used with confidence in similar watersheds.

90

However, it has been pointed out that the model needs to be customized to various

hydrological conditions for its suitability generalisation [Ndomba et al., 2005]. The

University of Dar es Salaam is currently customizing the SWAT model in various

catchments in the Eastern African region for the same purpose {personal communication

with P. M. Ndomba, 2007}. The general performance of the SWAT model in tropical

regions in Eastern Africa is summarized by Ndomba [Ndomba, 2007].

4.1.3 Overview of the AVSWATX model

ArcView Soil and Water Assessment Tool (AVSWATX) is a river basin, or watershed, scale

model developed by Arnold et al. [1998] for the United States Department of Agriculture -

Agricultural Research Service (USDA-ARS). The model is a modification of the Simulator

for Water Resources in Rural Basins (SWRRB) model for application to large basins

[Arnold et al., 1990]. SWAT is a semi-distributed, process oriented hydrologic model. It is

a continuous time model, which simulates both the water balance and the nutrient cycle

with daily time steps. This model incorporates key features of catchment properties,

including links between land cover hydrological responses. According to Arnold et al.

[2003]:

SWAT model has been developed to predict the response to natural inputs as well as

the manmade interventions on water and sediment yields in un-gauged catchments.

The model (a) is physically based; (b) uses readily available inputs; (c) is

computationally efficient to operate and (d) is continuous time and capable of

simulating long periods for computing the effects of management changes. The major

advantage of the SWAT model is that unlike the other conventional conceptual

simulation models can be used on ungauged watersheds.

The model structure is shown in Figure 29.

For modelling purposes, a macro-watershed or catchment is considered to consist of a

number of watersheds. The use of a number of discrete watersheds in a simulation is

particularly beneficial when different areas of the macro-watershed are dominated by land

uses or soils different enough in their properties to have different effects on the

hydrological response. Within SWAT, input information for each watershed is grouped

with respect to weather, unique areas of land cover, soil and management, and each such

area with a unique combination is identified as a hydrologic response unit or HRU (the

basic modelling unit).

91

Figure 29: Overview of SWAT hydrological structure (adapted from Arnold et al., 1998)

SWAT model is available with various interfaces, such as DOS, GRASS, ArcView,

and GRAM++. The most versatile interface is ArcView, which has been used in the

present study, as AVSWATX.

The model combines empirical and physically based equations, uses readily available

inputs, and enables users to study long-term impacts. The hydrological model is based on

the water balance equation (Equation 14):

( )iiiii

t

it QRPETQRSWSW −−−−+= Σ

=1 (14)

where SW is the soil water content minus the total content of the soil layer at wilting point

(-1.5MPa); t is the time in days; and R, Q, ET, P, and QR are the daily amounts (mm) of

92

precipitation, run-off, evapotranspiration, percolation and return flow respectively. Since

the model maintains a continuous water balance, complex basins are subdivided to reflect

differences in ET for various crops and soils. Thus, run-off is predicted separately for each

sub-area and routed to obtain the total run-off for the basin. This increases accuracy and

gives an accurate physical description of the water balance.

Surface run-off, or overland flow, is flow that occurs along a sloping surface. Surface run-

off occurs whenever the rate of water application to the ground surface exceeds the rate of

infiltration. When water is initially applied to a dry soil, the application rate and infiltration

rates may be similar. However, the infiltration rate will decrease as the soil becomes

wetter. When the application rate is higher than the infiltration rate, surface depressions

begin to fill. If the application rate continues to exceed the infiltration rate, once all surface

depressions have filled, surface run-off will start [Neitsch et al., 2005].

Using daily and/or sub-daily rainfall amounts, SWAT simulates surface run-off volumes

and peak run-off rates for each HRU. Surface run-off is estimated with a modification of

the SCS curve number method [USDA-SCS, 1986] (Equation 15) or the Green & Ampt

infiltration method [Green and Ampt, 1911]:

( )

SR

sRQ

8.0

2.02

+−

= SR 2.0< (15)

where Q is the daily surface run-off (mm), R is the daily rainfall (mm), and S is the

retention parameter (mm). The retention parameter S varies (i) among catchments, because

of changes in soils, land-use, and slope; and (ii) with time, because of changes in soil water

content. The parameter S is related to curve number by the SCS equation [USDA-SCS,

1986] (Equation 16):

−= 1100

254CN

S (16)

The constant 254 in Equation 16 gives S in mm. In the curve number method, the curve

number varies non-linearly with the moisture content of the soil. The curve number drops

as the soil approaches the wilting point and increases to near 100 as the soil approaches

saturation. The Green & Ampt method requires sub-daily precipitation data and calculates

infiltration as a function of the wetting front matrix potential and effective hydraulic

conductivity. Water that does not infiltrate becomes surface run-off.

93

Flow in a watershed is classified as overland or channelised. The primary difference

between the two flow processes is that water storage and its influence on flow rates is

considered in channelised flow. Main channel processes modelled by SWAT include the

movement of water, sediment and other constituents (e.g. nutrients, pesticides) in the

stream network, in-stream nutrient cycling and in-stream pesticide transformations.

Optional processes include the change in channel dimensions with time due to down-

cutting and widening.

Open channel flow is defined as channel flow with a free surface, such as flow in a river or

partially full pipe. SWAT uses Manning’s equation to define the rate of flow. Water is

routed through the channel network using the variable storage routing method or the

Muskingum river routing method. Both the variable storage and Muskingum routing

methods are variations of the kinematic wave model. A kinematic storage routing

technique that is based on saturated conductivity is used to calculate lateral subsurface

flow simultaneously with percolation. A detailed discussion of the kinematic wave flood

routing model can be found in Chow et al. [1988].

A shallow aquifer recharged by the percolation from the bottom of the root zone is

incorporated. Baseflow is allowed to enter the channel reach only if the amount of water

stored in shallow aquifer exceeds the threshold value defined through the calibration

process.

The water balance for a shallow aquifer in SWAT is calculated as follows (Equation 17):

shpumpdeeprevapgwrchrgishish WWWQwaqaq ,1,, −−−−+= − (17)

where ishaq , is the amount of water stored in the shallow aquifer on day i (mm), 1, −ishaq is

the amount of water stored in the shallow aquifer on day i-1 (mm), wrchrg is the amount of

recharge entering the aquifer on day i (mm), Qgw is the groundwater flow, or baseflow, into

the main channel on day i (mm), wrevap is the amount of water moving into the soil zone in

response to water deficiencies on day I (mm), wdeep is the amount of water percolating

from the shallow aquifer into the deep aquifer on day i (mm), and wpump,sh is the amount of

water removed from the shallow aquifer by pumping on day i (mm).

Evapotranspiration is the primary mechanism by which water is removed from a

catchment. Evapotranspiration includes all processes by which water at the earth’s surface

94

is converted to water vapour. It includes evaporation from the plant canopy, transpiration,

sublimation and evaporation from the soil. Three options for estimating potential

evapotranspiration (PET) are included in the model: Penman-Monteith [Monteith, 1965],

Priestley-Taylor [Priestley and Taylor, 1972] and Hargreaves [Hargreaves and Samani,

1985] are included in the model. The Penman-Monteith method requires information

concerning solar radiation, air temperature, relative humidity and wind speed. The

Priestley-Taylor method requires inputs of solar radiation, air temperature and relative

humidity. The Hargreaves method requires air temperature as input. The Penman-Monteith

method was selected to calculate PET for the catchment. Penmann-Monteith method takes

account both the feedback effects of surface evaporation on atmospheric moisture content

and the stomatal feedback effects between atmospheric moisture content and surface

evaporation. In the case of impact studies (land use changes), it is assumed to be the most

rigorous method of estimating effects of land-use change on evapotranspiration [Calder,

1992].

A brief description of AVSWATX interface

The AVSWATX interface consists of three segments: a main interface, a pre-processor and a

post processor. The Main Interface handles the creation of a new SWAT project, opening

an existing project, copying an existing project, deleting an existing project and exiting the

ArcView.

The pre-processor is the backbone of the interface. AVSWATX model (run from an

executable file) requires extensive input files in their respective formats. The pre-processor

helps the user create the necessary formats. The basic input required is the Digital

Elevation Model (DEM) for the area under consideration. The pre-processor generates the

Stream Network, identifies the outlet points for a given threshold value and delineates the

main watershed and sub-watersheds within it. Watershed characteristics like area, slope,

perimeter and channel characteristics are also calculated. Land cover and soil grids are then

overlaid and the basic modelling units are extracted. The other input files (including soil,

water use, management practices, pesticide and water quality) for each sub-basin are

written. Default values are used in many files, which could be modified using the EDIT

FILES menu. The sequence of input data creation is best followed using the

enable/disabled menu item. AVSWATX model is run using SWAT RUN menu.

95

The post-processor reads the results of the simulation run for the watershed as a basin file

and channel routing file in tabular form and facilitates the viewing of the output created

after SWAT model run. The basin table and channel routing table are viewed at daily,

monthly and yearly frequencies.

A detailed model description is given in [Neitsch et al., 2005]. Additional information

about AVSWATX and model updates can be found at http://www.brc.tamus.edu/swat.

Concepts and definitions for hydrological modelling

This section describes some of the basic terminologies used in Geographical Information

Systems (GIS) hydrologic modelling based research. Definitions of these terms are given in

the context of the AVSWATX model [Neitsch et al., 2005]:

Catchment is the drainage area in which water from various stream networks runs down to

the lowest point following the natural slope and collects to form a river.

Catchment outline forms an enclosed drainage boundary representing all cells of a given

Digital Elevation Model (DEM) that drain to the specified outlet location.

Sub-basins are delineated sub-drainage areas from within a catchment.

Streams are identified as lines of cells whose flow accumulation exceeds a specified

number of cells and thus a specified upstream drainage area.

Stream map is a theme containing all streams for a given DEM.

Stream network is a shapefile representing a set of connected water flowlines through

channel reaches and water bodies at the specified contributing source area for a given

catchment outline.

Channel lengths characterise the maximum water flow length of individual stream

elements.

Contributing source area (CSA) is the threshold drainage area required to define a

channel.

Curve number is a dimensionless parameter determined based on the following factors:

hydrologic soil group, land cover, land treatment and hydrologic conditions. Curve

Number values range from 1 (minimum run-off) to 100 (maximum run-off).

Antecedent moisture condition (AMC) is an indicator of catchment wetness and the

availability of soil moisture storage prior to a storm, and can have a significant effect on

run-off volume.

96

Hydrological response unit (HRU) is a region within a sub-basin that has unique land

cover attributes and soil characteristics.

Hydrological Soil Group (HSG) is the soil classification according to the minimum

infiltration rate obtained for a bare soil after prolonged wetting. Soils are classified into

four hydrological soil groups (A, B, C, and D).

4.1.4 Application of the AVSWATX model to the Shire River catchment

Patterns and trends in land cover change in the Shire River catchment indicate degradation

of woodlands including patches within the forest regions. Satellite data acquired for the

Shire catchment spanning the period 1989 to 2002 was classified into eight land cover

classes and changes between the dates determined (see Chapter 3). The main driving factor

behind woodland degradation is subsistence agricultural expansion and an increasing

demand for wood resources. As agriculture continues to play a dominant role in land cover

conversion, degradation from Brachystegia woodlands to open dry vegetation will

continue to occur. To this end, it is apparent that the rapid increase in cultivated areas

within the catchment will not only decrease the number of forests and naturally vegetated

areas, but also increase potential for localised run-off and erosion events, thereby

potentially diminishing the overall quality and quantity of water resources. In the absence

of observation studies, modelling within the Shire catchment could provide a better

understanding of hydrological responses to changes in land cover, and play a role in the

formulation of policies and programs for land use planning.

The major shortcoming for modelling within the Shire catchment is the lack of long-term

hydrological observations (run-off data, evapotranspiration data, sunshine hours and

intensity) with sufficient spatial coverage. However, hydrologically, the Shire catchment

falls within the tropical savanna climate with distinct dry and wet seasons. For the

successful modelling of the Shire catchment, adequate land cover data (generated in this

research) and long-term precipitation, soil and elevation data are available. Data input for

AVSWATX is limited to digital elevation, land cover, soils and weather (including daily

precipitation, daily temperature, windspeed, humidity and monthly solar radiation) data.

Since this study is aimed at assessing the contribution of different land cover types to

surface hydrological parameters and exploring the integration of new technologies

(including GIS) and natural sciences to improve catchment management and

environmental decision-making, the AVSWATX model was selected for application after a

97

review of a range of other models. Generally, most of the hydrological models mentioned

have evolved to accommodate the latest data sources including remote sensing and GIS

data. However, they require comprehensive input data and do not have a land cover change

scenario analysis. For this study, AVSWATX was selected because it has the ability to

characterize large complex watershed representations to account for the spatial variability

of soils, rainfall distribution and vegetation heterogeneity. It has the ability to show the

effects of land cover on surface run-off and sediment yield and the ability to characterize

large surface run-off and sediment yield producing mechanisms. AVSWATX was selected

over other hydrological models largely because of its forecast option. AVSWATX provides

the best and most comprehensive description of the effects of land cover change on

hydrological regimes and land cover change scenarios for anticipated future changes.

4.2 Methodology

4.2.1 Data

Data required for modelling the Shire River catchment were collected from various

sources. Spatial input data used included the DEM, land cover, soil and meteorological

data. The types and sources of input data are listed in Table 19.

Table 19: Data sets and sources for input into the AVSWATX model

Category Data Type Data source Spatial

discretisation

Catchment boundary Extracted from USGS# as DEM 1 km Boundary

conditions Sub-basin boundary Extracted from USGS as DEM 1 km

Elevation Topography/DEM Extracted from USGS 1 km

Land cover Land cover maps Derived from Landsat imagery

(chapter 2, this work) 28.5 m

Soil Soil types and depth Digital FAO Soil map of the

World (FAO, 1998) 1 km

Precipitation Rainfall zones Department of Meteorological

Services, Malawi 5 stations (Figure 34)

Stream Discharge Stream discharge rate Department of Water, Malawi 2 stations (Figure 34)

# USGS – United States Geological Survey (source:http://edc2.usgs/geodata)

Digital Elevation Model data

DEM data were sourced from the U.S. Geological Survey (USGS) at 1-km resolution. The

USGS Digital Elevation Model (DEM) data files are digital representations of cartographic

information in raster format. This data set provides reliable global coverage of topography

98

and stream networks. The DEM was used to delineate the topographic characterisation of

the catchment and determine its hydrological parameters of, such as slope, flow

accumulation, flow direction, and stream network. To capture heterogeneity in physical

properties, the catchment was divided into thirteen sub-basins. Each sub-basin was

individualised into Hydrologic Response Units (HRUs) according to relatively

homogeneous characteristics.

Land cover data

Land cover affects surface erosion, water run-off and evapotranspiration in a catchment.

Classified land cover, derived from Landsat remote sensing (see Chapter 2), has been used

as input for the modelling. Figure 9 shows the land cover classification map of the study

catchment, derived from a 1989 image, which was used for this part of the study.

For purposes of compatibility in AVSWATX, the FAO/Land Cover Classification System

(LCCS) land cover codes used in Chapter 2 were converted to the SWAT land cover/plant

codes. The corresponding land cover categories and the area covered under each category

for the Shire basin are presented in Table 20. From this conversion resulted the following

land cover distribution of the main land cover categories: Rangeland brush (RNGB) 38.9%

and Agricultural Land Generic (AGRL) 22.6%. The minor landuse classes are Forest

Mixed (FRST) 7.6%; Residential Medium Density (URMD) 12.0%; Forest Deciduous

(FRSD) 9.5%; Water (WATR) 8.3% and Wetland (WETF) 1.1%.

99

Table 20: Spatial distribution of land cover classes and SWAT land cover class codes for

1989 and 2002

1989 2002 LCCS Classification

SWAT Land/Plant cover classes Area (Ha) % Area (Ha) %

Fresh water Water 40 573 8.3 37 100 7.6

Built-up areas Residential medium density 58 835 12.0 5 775 1.2

Cultivated/grazing Agricultural Land Generic 110 874 22.6 200 570 40.9

Marshes Wetland 4 982 1.1 24 412 4.9

Savanna shrubs Rangeland range brush 190 795 38.9 110 495 22.6

Woody open Forest mixed 37 446 7.6 36 629 7.5

Wood closed Forest Deciduous 46 595 9.5 75 119 15.3

Total 490 100 100.0 490 100 100.0

Soil data

The soils database describes soil characteristics of the surface and upper subsurface of the

watershed. These data are used to determine a water budget for the soil profile, daily run-

off and erosion. The AVSWATX model requires textural properties and physical-chemical-

properties for each of the soil layers. Soil data were obtained from the United Nations,

Food and Agricultural Organization (UN/FAO) Digital Soil Map of the World

[FAO/UNESCO, 2003]. The Food and Agricultural Organisation/United Nations Education

Scientific and Cultural Organisation (FAO/UNESCO) soil map of Africa, in vector format

(ARC/INFO Export), with a 10x10 minute resolution (scale 1:100,000) was used. The soils

classification of Africa was based on the agronomic characteristics of 133 soil types. It is a

generalised soil dataset but contains the attribute data required by the model, which are not

available in any other form for Malawi. Other researchers in Tanzania have successfully

used this type of data in the SWAT hydrological model [Birhanu, 2005; Ndomba, 2007].

The major soil classification map of the study area according to the FAO/UNESCO Soil

Classification System is shown in Figure 30. The areas of all soil types in the study are

summarized in Table 21.

100

Figure 30: Spatial distribution of soils within the Shire River catchment

To integrate the soil map within the AVSWATX model, it is necessary to make a User Soil

Database containing the textural properties and physical-chemical properties for each of

the soil layers. In this database, all the soil types in the area are represented coupled with

their characteristics (Table 22).

To use the soils dataset in AVSWATX, it was first re-projected to a geographical projection

with decimal degrees, and then to Universal Transverse Mercator (UTM) with units in

meters, to match the Malawi co-ordinate system. Then, within the AVSWATX modelling

package, the catchment was intersected with the soil and land cover data sets, after which

hydrological parameters necessary for the model runs were determined and added to the

polygon and stream channel tables.

101

Table 21: Major soil types of the Shire River catchment and percent area covered

Soil unit Area (ha)

Percentage coverage

Hydrological soil group

Soil texture

572 26 722 5.5 C Sandy_clay

644 253 355 51.6 C Clay_loam

688 168 835 34.5 C Clay_loam

1972 41 188 8.4 D Water

Total catchment 490 100 100

Table 22: Soil parameters required by AVSWATX

Name of soil parameter Description

NLAYERS Number of layers in the soil (min 1, max 10)

HYDGRP Soil hydrologic group (A, B, C, D)

SOL_ZMX Maximum rooting depth of soil profile

ANION_EXCL Fraction of porosity from which anions are excluded

SOL_CRK Crack volume potential of soil (optional)

TEXTURE Texture of soil layer (optional)

SOL_Z Depth from soil surface to bottom of layer

SOL_BD Moist bulk density

SOL_AWC Available water capacity of the soil layer

SOL_K Saturated hydraulic conductivity

SOL_CBN Organic carbon content

CLAY Clay content

SILT Silt content

SAND Sand content

ROCK Rock fragment content

SOL_ALB Moist soil albedo

USLE_K Soil erodibility (K) factor

The soil data for the catchment indicates that the area is composed of Soil Conservation

Service (SCS) hydrological group C. However, they differ in the soil unit categorisation.

Soil unit depicts the soil identifier of the dominant soil type based on texture. Soil unit 572

are mainly sandy clay soils found in the valley sections of steep areas. They are generally

fine textured skeletal soils of low chemical fertility. More than 52% of the catchment is

covered by soil unit 644 (clay loamy soils) mainly found in gently sloping uplands. These

are generally very deep, medium to fine textured soils of low chemical fertility. Due to the

clay content, they are sticky and plastic when wet, and forms casts that are firm when

moist and hard when dry. In contrast, the clay loamy soils found in the alluvial plains (688)

102

are generally mopanic. They are very deep, imperfectly to moderately well drained,

medium to fine textured and of moderate chemical fertility. Soils within the study area are

generally shallow. They contain considerable clay and colloids. They have a low to very

low rate of water infiltration when wet, which results in high run-off potential.

Meteorological data

The weather variables needed to represent the hydrological balance are precipitation, air

temperature, solar radiation, wind speed and relative humidity. Data were obtained from

the Department of Meteorology, Malawi from five weather stations within the study area

(Table 23). Daily values were sourced for precipitation, minimum/maximum temperature

wind speed and relative humidity, while solar radiation readings were available only as

monthly means.

Precipitation

Malawi lies in the eastern and south-eastern part of Africa. The rainy season generally

extends from October/November to April, reaching a peak between December and

February. Rainfall distribution during the rainy season is variable, depending on the

interplay between tropical and mid-latitude weather systems and convective variability.

There are variations in the amount of rain, its onset, duration and intensity during the wet

season. The annual average rainfall is 950 mm a-1

, with a standard deviation 274 mm a-1

.

Significant features of the inter-annual variability in southern African seasonal rains are

linked to the El Niño Southern Oscillation (ENSO). ENSO can manifest as either El Niño

or La Niña episodes, associated with warm and cool sea surface temperatures respectively

in the tropical Pacific. Three distinct seasonal patterns of rainfall can be identified in the

Shire River catchment, characterised by total annual rainfall and within season differences,

is illustrated in the seasonal plots for the years 1980 to 1989, and 2002. The patterns of

rainfall are definitive, rather than the annual average rainfall. The features of these patterns

are described below.

The first pattern, observed in Figure 31 for the years 1978/79, 1979/80, 1981/82, 1982/83,

1983/84 and 1986/87, is characterised by the early onset of rains (October), with dry spells

in January followed and low rain (below 300 mm) in February and March. Total rainfall

tended to be slightly below average (mean of these years was 935 mm a-1

), with 1980 and

1981 being the driest years (at 923 mm a-1

).

103

Jun Aug Oct Dec Feb Apr Jun0

200

400

600

800

Rainfall (mm)

1978/79

Jun Aug Oct Dec Feb Apr Jun

1979/80

Jun Aug Oct Dec Feb Apr Jun0

200

400

600

800

Rainfall (mm)

1981/82

Jun Aug Oct Dec Feb Apr Jun

1982/83

Jun Aug Oct Dec Feb Apr Jun0

200

400

600

800

Rainfall (mm)

1983/84

Jun Aug Oct Dec Feb Apr Jun

Balaka

Salima

Mangochi

Chancellor

Ntaja

1986/87

Figure 31: Rainfall variability in the Shire River catchment – first pattern (Data from Department of Meteorology, Malawi)

The second pattern is the inter-seasonal variability observed in Figure 32 for the years:

1980/81, 1984/85, 1988/89 and 2001/02. Inter-seasonal variability was associated with the

early onset of rains, all of which started in October except for the 2001/02 rainfall season.

All the stations recorded rains above 100 mm in January and February except for 1980/81,

which had dry spells in February below 100 mm. Total annual rainfall varied from average

to wet. Above average annual total rainfalls of > 1 100 mm were recorded for 1980/81 and

1988/89. The graphical representation makes it clear that rainfall continued up to the end

of April and beginning of May at all the stations.

104

Jun Aug Oct Dec Feb Apr Jun0

200

400

600

800

Balaka

Mangochi

Chancellor

Salima

Ntaja

Rainfall (mm)

Rainfall (mm)

1980/81

Jun Aug Oct Dec Feb Apr Jun

1984/85

Jun Aug Oct Dec Feb Apr Jun0

200

400

600

8001988/89

Jun Aug Oct Dec Feb Apr Jun

2001/02

Figure 32: Rainfall variability in the Shire River catchment – second pattern (Data from the Department of Meteorology, Malawi)

The third pattern is observed in Figure 33 for the years 1977/78, 1985/86 and 1987/88,

when the onset of rain was delayed until the second dekad of November at all stations

except Balaka, which experienced rain in October. Overall, all the stations recorded annual

rainfall above 800 mm a-1

, except Mangochi which had low rainfall (550 mm) in the

1985/86 rainfall season. High rainfall is observed in December and January, totalling a

mean monthly rainfall of < 300 mm. The rainfall season has demonstrated significant inter-

decadal variability associated with slightly low means in February (mean of 135 mm) and

above normal rains in March, with a mean of 348 mm.

105

Jun Aug Oct Dec Feb Apr Jun0

200

400

600

800Rainfall (mm)

1977/78

Jun Aug Oct Dec Feb Apr Jun

1985/86

Jun Aug Oct Dec Feb Apr Jun0

200

400

600

800

Balaka

Salima

Mangochi

Chancellor

Ntaja

Rainfall (mm)

1987/88

Figure 33: Rainfall variability in the Shire River catchment – third pattern (Data from the Department of Meteorology, Malawi)

Due to variability in the distribution and occurrence of rainfall, there are also possible

changes in the availability of water for plant growth, the regeneration of vegetation, river

flow and ground water recharge. Natural vegetation is highly sensitive to high or low

rainfall, which may in turn modify soil moisture and albedo.

The AVSWATX programme requires a set of rain gauges with representative distribution

over the catchment and daily values for precipitation quantities. It should be noted that

only five rainfall stations (Mangochi, Ntaja, Balaka, Salima and Chancellor College) were

used in this study. Although they are distributed across the catchment, the number is

smaller than would be ideal. The stations used met two basic criteria: location within the

study area (at representative points along the length of the catchment) and the existence of

sufficiently long observed records falling within the study period (1989-2002).

Nevertheless, the rainfall stations were considered to represent the main run-off

contributing sub-basins. The available meteorological stations are shown in Figure 34 and

the period of records within and surrounding the study area are shown in Table 23.

106

Salima

Chancellor College

Balaka

Mangochi

Ntaja

Liwonde

Salima

Chancellor College

Balaka

Mangochi

Ntaja

Liwonde

Figure 34: Weather stations and river gauging stations in the Shire River catchment

Table 23: Weather stations and available data

Station Altitude (m)

Rainfall Temp Relative Humidity

Windspeed Solar

radiation

Mangochi 480 1961 - 2006 1961 - 2004 1980 - 2000 1979 - 2003 1979 - 2003

Salima 520 1954 - 2006 1961 - 2005 no data no data no data

Balaka 660 1976 - 2006 1976 - 2006 no data no data no data

Ntaja 670 1971 - 2006 1985 - 2004 no data no data no data

Chancellor College

650 1975 - 2006 1982 - 2006 no data 1975 - 2000 no data

River discharge

Daily stage readings of the Shire River at two stations were obtained from the Hydrology

Department of the Ministry of Water Resources of Malawi (Table 24). There are only two

flow-gauging stations on the Shire River within the study area, one at the inlet to the valley

107

at Mangochi (1T1), and one at Liwonde (1B1), taken as the outlet (Figure 34). There are

no gauging stations anywhere on the tributaries within the catchment.

The streamflow at Mangochi includes the flow from the catchment and the main outflow

from Lake Malawi. This complicates the calculation of the net catchment flow from the

Liwonde catchment (the area being modelled), as the main streamflow is an order of

magnitude larger than the catchment generated flow. How these flows were processed to

derive the net catchment flow for Liwonde is described below.

Table 24: Daily river flow data

Station Area (km2) Period

Mangochi 12 650 1975 – 2001

Liwonde 13 020 1948 – 2002

The raw data as received from the Hydrology Department of the Ministry of Water

Resources in Malawi required critical analysis to check the reliability and quality of the

data. Figure 35 shows a time-series plot of inflow and outflow stream data recorded from

1976 to 2000 for the Shire River catchment as received.

108

Figure 35: Time series streamflow for the Shire River Mangochi (inflow) and Liwonde (outflow) for the period 1976 - 1981: data as received

There are regular data records for the period from 1976 to 1981, except for three days in

1978 for which there were no recorded values. The consistency and few data gaps in the

time series makes this period eligible for analysis.

The period between 1982 and 1983 depicts dry years in terms of rainfall, while from 1984

to 1988 the rains were within the normal average. There are, however, major

inconsistencies between inflow and outflow – outflows were much lower than inflows

from 1982 to 1983. The discrepancy between inflow and outflow is too large to be

attributed to any possible loss, extraction or consumption between the inflow and outflow

weirs. Although rainfall in the catchment was lower than the long-term average during

1982 and 1983, the reduction is not sufficient to explain the discrepancy. In the absence of

a natural physical cause, the discrepancies could be attributed to gauge malfunctioning or

errors in recording. Data from these years are unusable for this investigation.

The data recorded from 1984 to 1986 are consistently accurate, except for two days in

1984 where there are data gaps and outliers. Data from this period was usable for this

study.

109

Data records from 1987 to 1988 comprise a complete data set. However, imbalances

between inflow and outflow were questionable, with inflow much higher than the outflow.

These could be due to recording errors at the outlet gauging station, as rainfall seems to be

high during this period. Rainfall contributes to higher flows during the wet season, which

is not demonstrated in this case. Thus, these data could not be utilised for this study.

From 1992 to 2000, data collection appeared to be inconsistent, with numerous data gaps

in the time series, specifically for the outflow monitoring station. The outflow is uniform

with season, instead of showing the expected major seasonal variations of prior years. The

data set for this period was thus regarded as unreliable and of no use for this study.

Based on the above evaluation, usable subsets were selected from the Shire River

streamflow data, between 1977 and 1981 for calibration of the hydrological model and for

the land cover change application (Figure 36) and between 1984 and 1985, for validation

of the hydrological model. Data cleaning was done by interpolation for days for which data

were missing or misrepresented by unphysical spikes or drops.

0

200

400

600

800

1000

1200

Jan-77

May-77

Sep-77

Jan-78

May-78

Sep-78

Jan-79

May-79

Sep-79

Jan-80

May-80

Sep-80

Jan-81

May-81

Sep-81

Inflow outflow

Streamflow (m3 s-1)

0

200

400

600

800

1000

1200

Jan-77

May-77

Sep-77

Jan-78

May-78

Sep-78

Jan-79

May-79

Sep-79

Jan-80

May-80

Sep-80

Jan-81

May-81

Sep-81

Inflow outflow

Streamflow (m3 s-1)

Figure 36: Streamflow data for 1977 - 1981, data as received

110

The inflow and outflow for 1977 - 1981 shows a consistent trend except for a few days

where there are data gaps and outliers. There are also imbalances between the inflow and

outflow, which could be the result of several factors. The main water balance components

such as precipitation, evaporation, interflow and baseflow could explain some of the flow

imbalances between inflow and outflow. Lake Malombe, a large shallow lake, might

provide hydraulic resistance to streamflow inducing delays, especially during the dry

season when flows are generally low. In addition, this variation may also be caused by

unknown data-recording errors in both inflows and outflows, as data were recorded

manually. For example, from 13 to 18 July 1977 there was a drop in stream outflow by

more than 50 m3 s

-1 – probably due to instrument malfunctioning, as there is no plausible

natural explanation for this drop. Between 17 October and 6 November 1978, the outflow

data records were consistently below 220 m3 s

-1, while no inflow data were recorded from

1 to 6 November 1978. These inconsistencies could be due to instrument malfunctioning or

maintenance.

Another questionable period is the record from 13 to 18 May 1979, where three

consecutive days with identical values, followed by two days with the identical, but lower

readings, during a period of generally increasing streamflow, where recorded. The values

from 11 to 14 March 1980 are also questionable, as the inflow was consistently 800m3

s-1

.

There is also uncertainty in data recorded from 19 to 25 March 1980. In this case the

inflow was uniform over 6 days at 813 m3

s-1

, an improbable result suggesting that a value

had been copied to account for missing readings. Another human error recording probably

occurred on 26 July 1980: here the reading shows a singular high value of 932 m3

s-1

– a

non-physical jump of ~100 m3

s-1

, taking into account the previous day and the following

day records. A corrected value of 832 m3

s-1

was inserted in place of 932 m3

s-1

, assuming

that a recording error of a single digit had occurred. There is also uncertainty in the data

measurements from 16 to 30 April 1981, where all the inflow recorded were above

1 030 m3

s-1

, having increased suddenly from 940 m3 s

-1. Such distorted data sets could be

due to errors in recording or malfunctioning of the gauging instrument.

The identified misrepresented data constitutes 3.8% of the entire period (1977 to 1981).

The correction and interpolation of ~4% is considered small enough that these corrections

do not fundamentally alter the integrity of the streamflow records. After data cleaning by

the deletion of questionable records and interpolation, the calibration and validation results

111

presented are the best achievable. Data was cleaned further based on a weighted five-point

smoothing routine, using the mean and standard deviation. Figure 37 shows the smoothed

outflow data for the Shire River. This smoothing is justified by the fact that the time

constants of the response of a large catchment such as the Shire River catchment, is longer

than the one day measurement frequency at the gauging stations.

Hydrological characterisation of Shire River catchment based on hydrological variables

Following data smoothing, a hydrological characterisation was conducted to identify

contributing water sources and to understand the flow processes of the Shire River,

including specifically the interplay between inflows from Lake Malawi, buffer storage and

flow retardation within Lake Malombe, and normal catchment processes. The

characterisation analysis was carried out for five-year period between January 1977 and

December 1981. A number of approaches have been used to characterise the

ground/surface water interaction and the role it plays in streamflow of the Shire River. The

approaches include correlation/qualitative analysis of streamflow and precipitation.

0

200

400

600

800

1000

1200

Jan-77

May-77

Sep-77

Jan-78

May-78

Sep-78

Jan-79

May-79

Sep-79

Jan-80

May-80

Sep-80

Jan-81

May-81

Sep-81

Outflow (m3 s-1) Inflow (m3 s-1)

Stream flow (m3 s-1)

0

200

400

600

800

1000

1200

Jan-77

May-77

Sep-77

Jan-78

May-78

Sep-78

Jan-79

May-79

Sep-79

Jan-80

May-80

Sep-80

Jan-81

May-81

Sep-81

Outflow (m3 s-1) Inflow (m3 s-1)

Stream flow (m3 s-1)

Figure 37: Smoothed daily streamflow data from 1977 - 1981

The quantitative approach entailed correlations between (i) flow at the exit from Lake

Malawi into the Shire River at Mangochi gauging station (1T1), and (ii) flow discharges at

the outlet of the Shire catchment observed at Liwonde gauging station (1B1). Based on the

result obtained from the correlation between inflow and outflow characterisation, an

112

automated baseflow separation technique based on master recession curves, developed by

Arnold et al., [1995], was used to separate the observed stream outflow components into

baseflow and surface flow. This technique was developed and successfully used by Arnold

and Allen [1999] for estimating baseflow and annual ground water recharge from

streamflow hydrographs. Baseflow values obtained after the filter represent flow from

Lake Malawi, while surface flow comprises flow from the catchment (hereafter referred to

as catchment streamflow). Furthermore, a quantitative analysis was conducted between

precipitation data for gauging stations located within the catchment and catchment

streamflow.

Since the focus of this study was to evaluate the effects of derived quantitative land cover

changes on hydrological processes, the catchment streamflow, obtained from the above

inflow/outflow characterisation, was used for AVSWATX calibration and model

application. The catchment streamflow time series resulting from the above processing is

depicted in Figure 38. This data set is used for further modelling and evaluation of land

cover - hydrological interactions.

0

25

50

75

100

Jan-77

May-77

Sep-77

Jan-78

May-78

Sep-78

Jan-79

May-79

Sep-79

Jan-80

May-80

Sep-80

Jan-81

May-81

Sep-81

Catchment streamflow (m3s-1)

Catchment stream outflow (m3 s-1)

0

25

50

75

100

Jan-77

May-77

Sep-77

Jan-78

May-78

Sep-78

Jan-79

May-79

Sep-79

Jan-80

May-80

Sep-80

Jan-81

May-81

Sep-81

Catchment streamflow (m3s-1)

Catchment stream outflow (m3 s-1)

Figure 38: Smoothed catchment streamflow data, 1977 - 1981

113

Sensitivity analysis and auto-calibration

Sensitivity analysis is an instrument for the assessment of input parameters in relation to

their impact on model performance [Lenhart et al., 2002]. Model sensitivity is defined as

the change in model output per unit change to an input parameter. Sensitivity analysis may

be used to evaluate how model outputs vary over a range for a given input variable. Some

researchers have noted that sensitivity analysis and calibration are difficult with a large

number of parameters [Ndomba, 2007]. In general, parameter sensitivity analysis aids the

user in reducing the number of parameters that must be varied by identifying the critical

parameters and allowing insensitive parameters to be held fixed, thereby reducing the

complexity and computational time required for model calibration.

The theoretical background of the sensitivity analysis method that is implemented in SWAT

is called the Latin Hypercube One-Factor-At-a-Time (LH-OAT) design and was proposed

by Morris [1991]. The LH-OAT sensitivity analysis combines the strength of global and

local sensitivity methods of analysis. Prior to the calibration and validation process, a

sensitivity analysis based on the integration of LH-OAT was performed to reduce

uncertainty and provide parameter estimation guidance. There are more than sixty

parameters in the AVSWATX model. Some of these parameters vary by sub-basin, land

cover and soil type. This variation may increase the number of parameters substantially.

However, a good number of other parameters are empirical or AVSWATX-specific. For

example, AVSWATX uses the SCS curve number method to estimate surface Curve

Number [USDA-SCS, 1972]. In the present study, twenty-seven parameters were included

in the analysis. The parameters selected have the greatest sensitivity to the hydrology of

the system and are related to land cover, run-off, groundwater penetration and soil

characteristics [Van Griensven, 2002].

Sensitivity analysis was performed for each sub-basin of the Shire River basin to reduce

uncertainty and provide parameter estimation guidance for the calibration steps. The

sensitivity analysis was carried out for a period of five years, including the calibration

period from 1977 to 1981. Parameter values as recommended by Van Griensven [2002]

were used as initial values for the analysis. Other researchers in the region, such as

Ndomba [2007], have successfully applied this approach.

During the sensitivity analysis, AVSWATX estimates the relative sensitivity (RS) of various

parameters. Since the parameters are of different types and vary in different magnitudes,

114

this assists in comparing the effects that different parameters have on the parameter

estimation process. Lenhart et al. [2002] categorized the relative sensitivity into four

classes. According to this classification, RS values between 0 - 0.05 are categorised as

small, while RS values between 0.05 - 0.2 are categorised as medium. On the higher side,

RS values greater than 0.2 - 1.0 and greater than one are classified as high or very high,

respectively. The relative sensitivity values found in the parameter estimation process were

categorized into these four classes. Sensitivity evaluations were carried out, producing nine

parameters for modelling. Identified parameters and relative sensitivities are presented in

Section 4.3.

Following the sensitivity analysis, auto-calibration was done for sensitive parameters only

to obtain the optimum values. The watershed model AVSWATX includes an option to

perform automatic calibration using the optimization algorithm. The automatic calibration

procedure in AVSWATX is based on the Shuffled Complex Evolution algorithm developed

at the University of Arizona (SCE-UA). SCE-UA is a global search algorithm that

minimizes a single objective function for up to 16 model parameters [Duan et al., 1992].

SCE-UA has been widely used in watershed model calibration and other areas of

hydrology such as soil erosion, sub-surface hydrology, remote sensing and land-surface

modelling. It has been found to be robust, effective and efficient. The SCE-UA has also

been applied with success to AVSWATX for hydrologic parameters [Eckhardt and Arnold,

2001] and hydrologic and water quality parameters [Van Griensven, 2002].

The auto-calibration tool, based on the SCE_UA algorithm available in AVSWATX, was

used to calibrate eight parameters in the model that govern streamflow in Shire River

catchment. The auto-calibration provides three methods of updating the parameters: (i)

replacement by a new value, (ii) adding fractionally to an initial value, and (iii) multiplying

an initial value by a factor. The second and third methods are references relative to an

initial value. This allows a lumped calibration of distributed parameters, which ensures that

the relative physical meaning is maintained (for example, the CN of forest is lower than the

CN of subsistence agriculture).

4.2.2 Model setup

The hydrological modelling using SWAT was based on the application of the Graphical

User Interface (GUI) of AVSWATX, which is embedded in ArcView [Di Luzio et al., 2001].

Tools are accessed through pull-down menus that are introduced in the various ArcView

115

GUI and custom dialogues. AVSWATX processes mapped land cover and soils data as well

as a Digital Elevation Model (DEM) to create a set of default model input files. Within

each sub-basin, Hydrological Response Units (HRUs) are created by AVSWATX. HRU

creation in AVSWATX requires land cover and soil threshold inputs [Di Luzio et al., 2001]

to define the level of spatial detail provided by the model. These thresholds are applied to

each sub-basin and function to control the size and number of HRUs created. Various GIS

data pre-processor modules were developed in the course of modelling the catchment

including catchment delineation, input map characterization and processing, stream and

outlet definition, the computation of geomorphic parameters and characterization of the

land cover and soil. Interactions between surface flow and subsurface flow in AVSWATX

are based on a linked surface to sub-surface flow model developed by Arnold et al. [1994].

The input data were prepared to the required format for input to the AVSWATX model. The

processing of spatially distributed data is achieved in a four-step approach.

The first step of the modelling process using AVSWATX is the catchment delineation,

which uses the DEM. Stream-catchment delineation identified the flow elements and

contributing areas (sub-basins) of the upper Shire River hydrological system. From the

DEM, a stream map was created using a minimum contributing source area (CSA)

threshold value of 8% to define lengths and numbers of stream channels. The threshold

value for the map represents the number of cells in the DEM that receives run-off from a

certain number of cells. All cells receiving run-off are classified as streams. Figure 39 is a

schematic representation of a grid-based catchment delineation showing flow from a DEM.

The stream map is used to locate outlet regions for individual sub-basins.

In the second step, the catchment is divided into model elements according to the land

cover properties of the area. This step requires a land cover classification system with

appropriate attribute data to be accepted by the program. The land cover layer generated in

this study using the FAO Land Cover Classification System (LCCS) classification (see

Chapter 2) was used. The LCCS classification system is not among the default schemes in

the AVSWATX program, but there is an option for user-defined land cover. This requires a

translation to be made in certain codes [Neitsch et al., 2005]. AVSWATX connects these codes

(Table 20) with the land cover/plant growth database through a look-up table.

116

Grid lines

Cell

Catchment outlet

Flow channel

Grid lines

Cell

Catchment outlet

Flow channel

Figure 39: Grid based discretisation and concept of flow path used in a cell

In the third step, the catchment is divided into model elements using a soil layer containing

soil properties of the catchment. All soil parameters are linked through look-up tables.

AVSWATX uses soil hydrological groups to define soil hydraulic properties. The drainage

of the soils is described according to the FAO classes [FAO/UNESCO, 2003] and used to

assign hydrological groups (in run-off generation perspective) based on permeability and

infiltration characteristics. The importance of assigning soil hydrological groups is that

they are used in AVSWATX to assign curve numbers (CN) for the land cover and soil

combinations, also known as HRUs, which are used in the equations to calculate the

partition between surface run-off and soil infiltration.

The catchment created in Step 1 is then intersected with land cover and soil data, and

parameters necessary for the hydrological model runs are estimated through a series of

look-up tables. This discretisation resulted in the definition of 13 sub-basins (Figure 40). In

AVSWATX, a catchment is delineated into sub-basins, which are then further subdivided

into HRUs. HRUs consist of homogeneous land cover and soil types. Based on two options

in AVSWATX, HRUs may represent either one sub-basin area with a dominant land cover

or soil type, or several homogeneous HRUs representing unique combinations. In this

117

study, sub-basin areas with a dominant land cover or soil type were used for the

simulations for computational efficiency. Each sub-basin delineated within AVSWATX is

simulated as a homogeneous area before estimates are summed for the basin.

The fourth stage is the incorporation of meteorological data. AVSWATX requires daily

meteorological data that either can be read from a measured data set or can be generated by

a weather generator model. In this study, a statistical weather generator file WXGEN

[Sharpley and Williams, 1990] was prepared to generate climatic data and fill in gaps in

the missing records from climatic data obtained. The weather generator was also used to

simulate daily values for variables from aggregated monthly solar radiation, as these were

only available as monthly values.

Figure 40: Sub-basins for the Shire River catchment

In to the case of precipitation, data can be put into the model from available measured

values. The distribution of precipitation measuring stations is not uniform across the

catchment, but is consistent enough, compared to the data availability in the country

overall. Most of the raw data vectors are incomplete to greater or lesser extents. However,

they are left as they are because AVSWATX uses a model developed by Nicks [1974] to

generate daily precipitation for simulations to interpolate missing data in the measured

118

records. This program fills data gaps or extends time series of daily data based on monthly

statistics. The monthly statistics, however, are based on long series of daily data. The

precipitation generator uses a first-order Markov chain model. Thus, input to the model

uses the monthly values for precipitation and the number of wet days per month. Given the

wet-dry state, the model determines stochastically whether precipitation occurs to obtain

the required daily inputs.

When a precipitation event occurs, the amount is determined by generating from a skewed

normal daily precipitation distribution or a modified exponential distribution. In this study,

a skewed normal daily precipitation distribution was opted for. Precipitation data were

retrieved by the AVSWATX input interface for the weather station nearest the centre of the

sub-basin. The details of the weather generator parameters and equations are given in the

SWAT model documentation [Neitsch et al., 2005].

4.2.3 Modelling the Shire River catchment

Calibration parameters

Simulation of the rainfall run-off model was based on previous experience and modelling

techniques published by various researchers [Ndomba, 2007; Ndomba et al., 2005; Van

Liew et al., 2005]. These techniques include a curve number method for calculating the

surface run-off [USDA-SCS, 1972], a first order Markov Chain Skewed Normal to

determine rainfall distribution, the Penman-Monteith method to compute potential

evapotranspiration, and the Muskingum routing method for flow routing water through

channel networks. Calibration was based on distributed unfilled rainfall data, whereby the

model uses a built-in weather generator to interpolate missing rainfall values.

The objective of a calibration procedure is to estimate the values for parameters that cannot

be assessed directly from field data. Parameter estimation is designed to reduce the

uncertainty of the modelling process. A typical approach is to select an initial estimate for

the parameters from within previously specified ranges. The parameter values are then

adjusted to optimise the agreement between model behaviour and that of the watershed.

The process of adjustment can be done manually or by computer-based automatic methods.

According to Refsgaard and Storm [1996], three types of calibration procedures can be

differentiated: trial-and-error manual parameter adjustment, automatic numerical parameter

optimization and a combination thereof. For hydrological model calibration, a combination

119

of manual and automatic procedures has been recommended [Gan, 1988]. Manual

calibration alone is tedious, time consuming, and requires an experienced researcher to

anticipate optimum parameter combinations. Thus, there have been several efforts towards

development of automated calibration methods. Automatic calibration, however, relies

heavily on the optimisation algorithm and the specified objective function. If not carefully

specified, the process might lead to a local optimisation point in the multidimensional

space that is unphysical or poorly matched to the overall catchment behaviour. Therefore,

in the calibration process for this study, a combination of manual and automatic parameter

estimation was adopted to optimise the nine selected parameters.

Available catchment streamflow data from January 1977 to December 1981 (5 years) were

used for model calibration, while data from January 1984 to December 1985 (2 years) were

used for model validation. Calibration was done at daily time intervals by comparing

modelled and measured streamflows at the outlet point of the catchment at the Liwonde

gauging station. Parameters through AVSWATX interface were changed in a semi-

distributed way (i.e. sub-basinwise). The optimisation was conducted on one or two

parameter(s) at a time depending on the computation resource required for a particular

simulation. The parameter ranges were selected with reference to relevant literature and

guidance from an experienced modeller (P.M. Ndomba, private communication).

Validation

To use any predictive watershed model for estimating the effectiveness of potential land

cover changes, the model must be first calibrated against measurements, and then be

validated (without further parameter adjustment) against an independent set of measured

data. This testing of a model on an independent data set is commonly referred to as model

validation. Model calibration determines the best or at least a reasonable, parameter set,

while validation ensures that the calibrated parameters set performs reasonably well under

an independent data set. Provided the model’s predictive capability is demonstrated as

reasonable in the calibration and validation phase, the model can be used with confidence

for future predictions in different land cover and management scenarios.

In this study, calibration and validation procedures presented in the SWAT user manual

were followed [Neitsch et al., 2005]. Calibration for water balance and streamflow was

done first for average annual conditions. Once the run was calibrated for annual conditions,

we shifted to the monthly records to fine-tune the calibration. Model efficiency expresses

120

the fraction of the measured streamflow variance that is reproduced by the model. Model

outputs were calibrated to fall within the average measured values and then model

performance statistics (ENS) (Equation 18) were evaluated by comparing simulated and

measured annual, monthly and daily catchment streamflows. Calibration and validation

were performed for the selected periods. The objective functions used to test the model

performance were Nash and Sutcliffe simulation efficiency (ENS) and the Index of

Volumetric Fit (IVF):

( ) ( )[ ]∑ ∑ −−−= 22/1 meanobssimobsNS QQQQE (18)

where ENS is coefficient of simulation efficiency, Qobs is the measured streamflow in each

model step (m3 s

-1) (in this case annual, monthly and daily), Qsim is the simulated

streamflow in each model step (m3 s

-1) (in this case annual, monthly and daily) and Qmean is

the mean measured streamflow in each model step during the evaluation period (m3 s

-1) (in

this case annual, monthly and daily).

Simulation results are good for values for values of ENS ≥ 0.75, while for values of ENS

between 0.75 and 0.36, the simulation results are acceptable. ENS values less than 0.36 are

unacceptable [Popov, 1979]. These values were considered adequate statistical values for

acceptable calibration for the purposes of this study. In addition to quantitative model

performance evaluation, simulated and measured flows were compared in graphical

displays to convey qualitative information such as trends and distribution patterns of flows.

4.2.4 Testing effects of land cover change

To test the assumption that land cover change has affected watershed streamflow, further

simulations were performed using the land cover classifications derived from two different

Landsat images in 1989 and 2002 respectively for the same rainfall regime: the 1977 to

1981 period. The 1989 land cover map was used for the calibration and validation runs of

the model. To isolate and evaluate the variability of streamflow due only to land cover

changes, the AVSWATX model was re-run using the 2002 land cover map, while all the

other input variables remained constant.

4.2.5 Scenario generation

Scenario analysis is a process of evaluating possible future events by considering

alternative possible outcomes. This analysis is designed to facilitate decision-making and

assessment through a complete consideration of possible outcomes and their implications.

121

The development of strategies for water resource planning and management and the

assessment of impacts of potential environmental change are often guided by the analysis

of multiple future scenarios.

The sensitivity the Shire River catchment hydrological system to future land cover changes

were analysed using AVSWATX model, by formulating a range of land cover change

scenarios. As a first test, two extreme limiting scenarios were considered – total

deforestation and total forestation. The modelling process was carried out to assess

influences on run-off components and total water yield in response to these bounding

conditions. Most likely changes to land covers are conversion from rangeland and forest to

agriculture in a land-degradation (pessimistic) scenario, and conversion from agricultural

land to rangeland and forest in a land-conservation (optimistic) scenario. Changes to all

three of these land cover types have influences on catchment hydrology. Urbanisation

equally affects hydrological processes. However, it was excluded from scenario analysis,

as it constitutes a small fraction of the total area, and in Malawi, still a small area of

change. For the second stage scenario modelling, a range of scenarios were considered, in

which fractional changes were made to these three land cover classes for the pessimistic

and optimistic trends. These sensitivity analyses were aimed at providing insights on the

proper behaviour of the model, and at generating plausible scenarios for guiding land use

policy and development strategies.

Each scenario was created by changing patches of selected land cover type one into target

land cover type two. Land cover classifications derived from the 2002 Landsat image were

used as the baseline. The type two land covers were defined by assigning changed areas to

the baseline classification, selected randomly by computer simulation within the selected

land cover type, and spread evenly over the target catchment. The total area affected in

each scenario was an assigned proportion of the original area of land cover type one.

Scenarios were developed for changes in the original areas of rangeland and forestry in

increments of 10%, 20%, and 40%. Scenario runs were conducted for these incremental

steps separately for each of the two types of land cover, for degradation and conservation

scenarios. Combined scenarios were run for the conservation scenario only, in which

agricultural land reverts to rangeland and forest, each changed by the medium and

maximum increments (20% and 40%) to evaluate combined effects on the hydrology.

122

A narrative description of the scenarios is provided below.

Business as usual

For the business as usual baseline scenario, the 2002 land cover grid (Chapter 2) was used.

Compared to the 1989 land cover map, the following gross changes are noted in the 2002

land cover map: Forests are reduced in the lower and upper escarpments within the

catchment in favour of transitional woodland-shrub; there is transformation of savanna

shrubland areas into cultivated and grazing land; and there are major increases in

grassland areas. The hydrological model was run with the 2002 land cover grid

unchanged, and the 1989 climatic data set. (Note that the streamflow data for 2002 was

unusable. Therefore, it was not possible to carry out a full validation of the AVSWATX

model using 2002 climatic records and the 2002 land cover classification. The AVSWATX

model, validated using the 1989 land cover image and earlier catchment streamflow, was

used instead.)

Land degradation scenario

The second scenario, termed land degradation, represents an unfavourable scenario,

including accelerated land cover change with extensive deforestation. In this case,

significant fractions of forest and savanna shrubland areas are transformed into the

category agricultural land generic, which includes subsistence agricultural areas,

transitional woodlands and sparsely vegetated areas. There is a significant decrease in the

area of the woodlands because of the increased use of land for agriculture to produce food

for the growing population. Although this scenario may imply increasing built-up areas, as

a fraction of the catchment this change is minor and so the size of the built-up area remains

constant. Three different extents of deforestation of forest deciduous and of rangeland

brush respectively were tested as shown in Table 25, indicating a range of increasingly

pessimistic scenarios up to 2020.

123

Table 25: Characteristics of 2002 land cover data and deforestation scenarios

Scenario Base line 1 2 3 4 5 6

Land use/Land cover Business as usual

Rangeland brush converted to Agricultural land generic

Forest deciduous converted to Agricultural land generic

Descrip-tion

Land cover 2002

10% 20% 40% 10% 20% 40%

Area (ha) 200 570 211 620 222 669 274 816 208 082 215 594 230 618 Agricultural land generic % 40.9 43.2 45.4 56.1 42.5 44 47.2

Area (ha) 110 495 99 445 88 396 66 297 110 495 110 495 110 495 Rangeland brush % 22.6 20.3 18.1 13.6 22.6 22.6 22.6

Area (ha) 75 119 75 119 75 119 45 071 67 607 60 095 45 071 Forest deciduous % 15.3 15.3 15.3 15.3 13.8 12.3 9.1

Land categories not changed in scenario analysis

Area (ha) 37 100 37 100 37 100 37 100 37 100 37 100 37 100 Water

% 7.6 7.6 7.6 7.6 7.6 7.6 7.6

Area (ha) 5 775 5 775 5 775 5 775 5 775 5 775 5 775 Residential medium density % 1.2 1.2 1.2 1.2 1.2 1.2 1.2

Area (ha) 24 412 24 412 24 412 24 412 24 412 24 412 24 412 Wetland

% 4.9 4.9 4.9 4.9 4.9 4.9 4.9

Area (ha) 36 629 36 629 36 629 36 629 36 629 36 629 36 629 Forest mixed % 7.5 7.5 7.5 7.5 7.5 7.5 7.5

Land conservation

The third scenario, termed land conservation, represents of the creation of a greener

environment through management and re-forestation. This scenario involves the

conversion of potentially vulnerable areas (subsistence agricultural land) into forest and

savanna woodlands. Different extents of forestation were investigated, as shown in Table

26, indicating an optimal scenario up to 2020.

124

Table 26: Characteristics of simulated land cover forestation scenarios

Scenario Base line

7 8 9 10 11 12

Busi-ness as usual

Savanna converted to Agri-

land

Agricultural land converted to Forest

Deciduous

Agri-land to both Savanna and Forest

Land use/Land cover

Descrip-tion

Land cover 2002

20% 40% 20% 40% 20% 40%

Area (ha) 200 570 160 456 120 342 160 456 120 342 120 40 114 Agricultural Land Generic % 40.9 32.7 24.6 32.7 24.6 24.6 8.2

Area (ha) 110 495 150 609 190 723 110 495 110 495 150 609

190 723 Rangeland brush % 22.6 30.7 38.9 22.6 22.6 30.7 38.9

Area (ha) 75 119 75 119 75 119 115 233 155 347 115 233

155 347 Forest Deciduous % 15.3 15.3 15.3 23.5 31.7 23.5 31.7

Land categories not changed in scenario analysis

Area (ha) 37 100 37 100 37 100 37 100 37 100 37 37 100 Water

% 7.6 7.6 7.6 7.6 7.6 7.6 7.6

Area (ha) 5 775 5 775 5 775 5 775 5 775 5 775

5 775 Residential medium density % 1.2 1.2 1.2 1.2 1.2 1.2 1.2

Area (ha) 24 412 24 412 24 412 24 412 24 412 24 412

24 412 Wetland

% 4.9 4.9 4.9 4.9 4.9 4.9 4.9

Area (ha) 36 629 36 629 36 629 36 629 36 629 36 629

36 629 Forest mixed

% 7.5 7.5 7.5 7.5 7.5 7.5 7.5

Absolute and relative changes in annual values of surface flow, baseflow and total water

yield were calculated for each scenario. Average annual and monthly outflows simulated at

the catchment outlet for the various change scenarios were then compared to the baseline

case.

4.3 Results and discussion

4.3.1 Hydrological characterization of Shire River catchment using hydrological

variables

Characterisation analysis in the upper Shire River has found that there is a strong positive

correlation (r2 = 98%) between measured river outflow from Lake Malawi measured at

Mangochi (Gauge No. 1T1) and the Shire River measured outflow, measured at Liwonde

(Gauge No. 1B1). This indicates that the inflow from Lake Malawi flowing into the Shire

River flows out from the river system at Liwonde, with minimal perturbation from the

Shire River catchment.

Although the statistical analysis suggests that the main source of inflow to Shire River is

Lake Malawi, a quantitative analysis suggests possible contributions from the catchment.

125

The hydrological variables used included daily rainfall and daily catchment streamflow

measured at the outlet (Liwonde). The results of the qualitative analysis showed that there

is an association between streamflow and the onset and offset of rainfall patterns,

suggesting that catchment rainfall does account for a substantial amount of groundwater

flow and surface flow into the river. As shown in Figure 41 at the onset of the rainy season

there is a steady increase in streamflow recorded at the outlet, while during the dry season

decreased flows have been recorded. At the onset of the rains, water losses by plant uptake

and evaporation decline because of milder temperatures. Natural streamflow increases as

rains produce more run-off. During the dry months of late April to late October,

evaporation and water used by plants decrease run-off and the amount of groundwater

available to support streamflow. Hence, streamflow declines. Precipitation is the most

significant climate variable affecting daily streamflow gains.

0

40

80

120

160

200

Jan-77

May-77

Sep-77

Jan-78

May-78

Sep-78

Jan-79

May-79

Sep-79

Jan-80

May-80

Sep-80

Jan-81

May-81

Sep-81

0

40

80

120

160

200

Catchment streamflow (m3 s-1) Rainfall (mm)

Catchment streamflow (m3 s-1)

Rainfall (mm)

0

40

80

120

160

200

Jan-77

May-77

Sep-77

Jan-78

May-78

Sep-78

Jan-79

May-79

Sep-79

Jan-80

May-80

Sep-80

Jan-81

May-81

Sep-81

0

40

80

120

160

200

Catchment streamflow (m3 s-1) Rainfall (mm)

Catchment streamflow (m3 s-1)

Rainfall (mm)

Figure 41: Time series plots of catchment streamflow and rainfall

The characterisation results indicate contributions to the Shire River flow from both Lake

Malawi and the Shire River catchment. However, to analyse the effects of catchment land

cover changes on surface run-off and seasonal river variations, water from Lake Malawi

flowing through the river system was excluded from the modelling process. The study only

modelled the catchment streamflow, which comprises surface run-off and baseflow from

the catchment.

126

4.3.2 Modelling of the Shire River catchment

Sensitivity analysis

For the Shire River catchment, sensitivity analyses showed that from the total of 28

parameters, only fourteen parameters revealed significant effects on the flow simulation.

Sensitivity analysis results are shown in Table 27, classified in terms of relative sensitivity,

as defined by Lenhart et al. [2002].

Table 27: Relative sensitivity values of the optimised parameters

Parameter Relative

sensitivity RS Category Ranking

SCS run-off curve number, CN2 3.820 very high 1

Soil evaporation compensation factor, ESCO 0.277 high 2

Soil available water capacity (mm WATER/mm soil), SOL_AWC

0.167 medium

3

Soil depth (mm), SOL_Z 0.053 medium 4

Maximum canopy storage (mm), CANMX 0.044 small 5

Saturated hydraulic conductivity (mm h-1), SOL_K 0.040 small 6

Surface run-off lag time (days), SURLAG 0.024 small 7

Average slope steepness (m/m), SLOPE 0.022 small 8

Baseflow alpha factor (days), ALPHA_BF 0.015 small 9

Moisture soil albedo, SOL_ALB 0.006 small 10

Channel effective hydraulic conductivity, CH_K2 0.006 small 11

Manning’s n value for main channel, CH_N 0.004 small 12

Average slope length, SLSUBBSN 0.003 small 13

Ground water “revap” coefficient, GW_REVAP 0.001 small 14

The sensitivity analysis identified the parameters SCS run-off curve number (CN2) and Soil

evaporation compensation factor (ESCO) as highly sensitive. Soil available water capacity

(SOL_AWC) and Soil depth (SOL_Z) were categorised as having medium sensitivity. The

rest of the parameters were found to have smaller relative sensitivities. It should be

mentioned that within the “small” relative sensitivity group, parameters that generally

govern the surface and sub-surface hydrological processes and have more physical

meaning to stream routing were selected for model optimisation. These parameters include

Surface run-off lag time (SURLAG), Saturated hydraulic conductivity (SOL_K), Baseflow

alpha factor (ALPHA_BF) and Ground water “revap” coefficient (GW_REVAP).

Five other low sensitivity parameters, not considered critical, were held fixed at default

values. These comprised Maximum canopy (CANMX), Average slope steepness (SLOPE),

127

Moisture soil albedo, (SOL_ALB), Channel effective hydraulic conductivity (CH_K2),

Manning’s n value for main channel (CH_N) and Average slope length (SLSUBBSN).

Auto-calibration

In the case of a new model application to an un-calibrated catchment, it is recommended

that the entire list of sensitive parameters be explored. Researchers working in well-

calibrated catchments have ended up using only a few of the most sensitive parameters

with confidence [Van Liew et al., 2005; Yapo et al., 1996]. However, for this study, the

eight sensitive parameters selected using the sensitivity analysis and based on physical

considerations, were optimised during calibration, following procedures incorporated in

the AVSWATX model.

Initial lower and upper default and final calibrated values are presented in Table 28. As

noted in the table, values for the SCS run-off curve number (CN2) and soil available water

capacity (SOL_AWC) are expressed as percentage change from the default value, and were

modified by multiplication of a relative change. High values of CN2, up to more than 80,

indicate greater contributions of precipitation to surface run-off than to subsurface flow.

The remaining parameters were calibrated by replacement of initial values.

Table 28: Parameter values calibrated in SWAT using the auto-calibration tool

Parameter Units Lower bound

Upper bound

Initial default value

Optimum value

Calibrated value

74 (RNGB) 7.4% 79

83 (AGRL) 4.2% 86

CN2*

-10 +10

73 (FRST) -7.5% 68

SOL_AWC* -50 +50 0.13 (688) 0.50% 0.07

mm WATER/mm

soil 0.10 (644) 0.90% 0.09

ESCO 0 1 0.950 0.90 0.900

Surlag days 0.5 10 4.0 0.70 0.70

GW_Revap 0.02 0.2 0.02 0.08 0.08

Alpha_BF days 0 1 1 0.007 0.007

Sol_z mm -50 50 300 300 300

Sol_k mm h-1 -50 50 38 38 38

CN2* and SOL_AWC* parameter values expressed as percent change from default values

128

Streamflow calibration, validation and model evaluation

Model simulation results are presented on an annual, monthly and daily basis for the

calibration and validation of the Shire River catchment streamflows. In the validation

process, the model was run without changing the input parameters, which were set during

the calibration process.

Calibration 1977-1981

Annual calibration

Measured catchment streamflow data for 1977-1981 were used for the calibration of the

AVSWATX model. Generally, the measured and simulated average annual volumes are

comparable (Table 29). The results show that the simulated average annual total water

yield was 138 mm, while the measured yield was 140 mm. Calibration results for surface

run-off and ground water components of the total water yield (expressed in mm) are shown

in Figure 42. Measured and calibrated mean annual average flow agree, as required. The

simulated annual average has a standard deviation of 17% (n = 5). The model

underestimates the mean annual flow by more than the one standard deviation for 1977.

Table 29: Average annual volumes obtained from calibration for 1977-1981

Total water yield (mm)

Surface flow (mm)

Baseflow (mm)

Measured# 139.6 82.1 59.4

Simulated 137.9 81.8 58.5

# Measured data as presented in catchment streamflow, Figure 41.

Annual validation

The results of the average annual model validation are shown in Figure 42. They show that

the simulated average annual total water yield 102 mm, while the measured yield was

157 mm. Annually, the model tended to underestimate water yield, with a standard

deviation of 16% (n = 2) over the validation years.

Monthly calibration

After the water-balance was calibrated for the annual simulation period, a seasonal

calibration and verification on a monthly basis was done. A time-series plot of monthly

catchment streamflow indicates an acceptable agreement between the measured and

simulated catchment monthly flows as indicated by the value of the Nash and Sutcliffe

efficiency criteria, ENS = 86% with standard deviation of 43% (n = 60) (Figure 43).

129

0

50

100

150

200

1977

1978

1979

1980

1981

1984

1985

Observed (mm) Simulated (mm)

Catchment streamflow (mm)

Calibration period Validation period

0

50

100

150

200

1977

1978

1979

1980

1981

1984

1985

Observed (mm) Simulated (mm)

Catchment streamflow (mm)

Calibration period Validation period

Figure 42: Comparison of measured and simulated average annual water yield (mm) by calibration and validation period

0

15

30

45

60

Jan-77

May-77

Sep-77

Jan-78

May-78

Sep-78

Jan-79

May-79

Sep-79

Jan-80

May-80

Sep-80

Jan-81

May-81

Sep-81

Catchment streamflow (m3 s-1)

Observed monthly catchment streamflow (m3 s-1)

Simulated monthly catchment streamflow (m3 s-1)

0

15

30

45

60

Jan-77

May-77

Sep-77

Jan-78

May-78

Sep-78

Jan-79

May-79

Sep-79

Jan-80

May-80

Sep-80

Jan-81

May-81

Sep-81

Catchment streamflow (m3 s-1)

Observed monthly catchment streamflow (m3 s-1)

Simulated monthly catchment streamflow (m3 s-1)

Figure 43: Comparison of monthly streamflows for calibration period, 1977 - 1981

130

Monthly validation

Monthly comparisons of the measured and simulated catchment streamflows for the

validation period (1984-1985) are presented in Figure 44. At the monthly time-step, the

calibrated model performed well when applied to the validation period. The statistical

evaluation of the simulated catchment streamflows yielded ENS = 64% with a standard

deviation of 63% (n = 24). Although some of the localised average storms (between July

and October 1984) were not well simulated, the model efficiency at monthly time-steps for

the Shire River catchment was considered acceptable. The model was therefore validated

for monthly predictions.

0

20

40

60

80

Jan-84

May-84

Sep-84

Jan-85

May-85

Sep-85

Catchment streamflow (m3s-1)

Observed monthly catchment streamflow (m3 s-1)

Simulated monthly catchment streamflow (m3 s-1)

0

20

40

60

80

Jan-84

May-84

Sep-84

Jan-85

May-85

Sep-85

Catchment streamflow (m3s-1)

Observed monthly catchment streamflow (m3 s-1)

Simulated monthly catchment streamflow (m3 s-1)

Figure 44: Comparison of monthly catchment streamflows for validation period, 1984 - 1985

Daily calibration

Model runs were conducted on a daily basis to compare the simulation output with the

measured daily catchment streamflow. A time-series plot of daily measured and simulated

catchment streamflow for the calibration period indicates agreement between the main

periods of enhanced and low flow (Figure 45). The daily calibration was considered

131

acceptable with a good agreement between observed and simulated flows as shown by the

Nash-Sutcliffe simulation efficiency where ENS = 42%.

Catchment streamflow (m3s-1)

0

20

40

60

80

100

Jan-77

May-77

Sep-77

Jan-78

May-78

Sep-78

Jan-79

May-79

Sep-79

Jan-80

May-80

Sep-80

Jan-81

May-81

Sep-81

Observed daily catchment streamflow (m3 s-1)

Simulated daily catchment streamflow (m3 s-1)

Catchment streamflow (m3s-1)

0

20

40

60

80

100

Jan-77

May-77

Sep-77

Jan-78

May-78

Sep-78

Jan-79

May-79

Sep-79

Jan-80

May-80

Sep-80

Jan-81

May-81

Sep-81

Observed daily catchment streamflow (m3 s-1)

Simulated daily catchment streamflow (m3 s-1)

Figure 45: Comparison of daily catchment streamflows for calibration period 1977 - 1981

It can be observed from Figure 45 that the model failed to capture some of the peaks in the

rainy season (i.e. between November and March) during 1977 and 1978. The simulated

flows under predict measurements by up to a factor of two. The dry season flow in the

latter years (1979, 1980 and 1981) has a uniform baseline much higher than the measured

flows.

The model inefficiencies were due to its failure to capture some of the low flows especially

during the long dry seasons between May and October. Possible causes of the uniform

simulations include:

• Poor representation of reliable spatial precipitation data because of an insufficient

number rainfall stations, including certain sub-catchments with no stations at all. The

available raingauge (Salima) is upstream, while Chancellor College is downstream of

sub-basins 4, 12 and 13 (Figure 34).

132

• The model was calibrated on the assumption that the soil comprised lumped clay-

loam. This soil generally has a low infiltration capacity. In reality, the basin consists

of a variety of soil units, which have not yet been mapped in sufficient detail or

precision.

• From a hydrological perspective, the occurrence of groundwater resources within the

Shire rift valley area is associated with low yielding weathered Precambrian

basement gneiss complex formations [Malawi Government, 2001]. The long

residence time of the groundwater may imply the existence of low hydraulic

gradients with subsequent low flow velocities, or low transmissivities due to the

nature of the aquifer material. Both theories seem to apply to the Shire River

catchment, as it is located in the Shire plain, which is characterised by flat

topography and clay-loam soils [Bath, 1980]. However, trends in groundwater levels

are due to many different factors and this requires more research, as data gathered by

the Department of Water Resources in Malawi on groundwater levels are

inconclusive.

• Possible error in the manipulation of the streamflow data from the inflow and

outflow measuring weirs to separate the Lake Malawi outflow form the catchment

contributions. As the latter flows are small compared to the Lake Malawi flow,

uncertainties are magnified in the adjusted set of measured streamflows.

Overall, the model was able to reproduce the main features of the streamflow in terms of

magnitude and seasonality. From the results, the model gives an adequate representation of

the water balance and outflow hydrographs at the basin outlet.

Daily validation

Validation runs were also performed to compare the daily measured (1984 – 1985) and

simulated catchment streamflows (Figure 46). The statistical evaluation of the simulated

catchment streamflow yielded ENS = 36% for the daily predictions. This model prediction

for the daily Shire River catchment streamflow is barely acceptable. The simulation flows

were able to follow the measured pattern but failed to simulate several of the highest flow

peaks in February, March and October 1984, and in February and March 1985. The model

gave a smooth outflow from May to November 1985, and appeared to be unresponsive to

minor fluctuations driving the measured flow. Uniform prediction of the catchment

streamflow in this period could be attributed to poor representation of rainfall data.

133

0

50

100

150

Jan-84

May-84

Sep-84

Jan-85

May-85

Sep-85

Catchment streamflow (m3 s-1)

Observed daily catchment streamflow (m3 s-1)

Simulated daily catchment streamflow (m3 s-1)

0

50

100

150

Jan-84

May-84

Sep-84

Jan-85

May-85

Sep-85

Catchment streamflow (m3 s-1)

Observed daily catchment streamflow (m3 s-1)

Simulated daily catchment streamflow (m3 s-1)

Figure 46: Comparison of daily catchment streamflow for validation period: 1984 - 1985

Results of land cover change test

In this section, results are presented for the test of substituting the 2002 land cover

classifications into the model, in place of the 1989 land cover, keeping climate input data

sets and all other parameters constant. Climatic data for the period 1977 to 1981 were used

for this comparison.

General patterns and trends of land cover change in the Shire River catchment between

1989 and 2002 suggest transition towards the degradation of woodlands and an increase in

subsistence agricultural land. The land cover mapping showed that 23% of the land was

covered by agricultural land in 1989. Subsistence agricultural area has increased by 18%,

occupying 41% of the study area in 2002. The increase in subsistence agricultural land

corresponds with the simultaneous decrease in vegetated areas. Summaries of the changes

in land cover distribution, showing the percentage of land cover for 1989 and 2002, are

presented in Chapter 2. Simulation runs were conducted on an annual, monthly and daily

basis to compare the modelling outputs using the 1989 and the 2002 land covers.

134

A comparison of the multi-year average annual streamflows generated using 1989 and

2002 land covers respectively is presented in Table 30, while Figure 47 presents mean

annual catchment streamflows. The 1989 land cover yielded mean annual water yield of

140 mm compared to 207 mm from the 2002 land cover. The maximum annual flow yield

for the 1989 land cover was 170 mm while in 2002 it was 318 mm (for the 1978

meteorological year). The total annual volume of water simulated increased by 47% with

the 2002 land cover.

Table 30: Parameters obtained from annual simulations for 1989 and 2002 land cover

Item 1989 land cover data 2002 land cover data

Mean annual water yield 140 mm 207 mm

Mean annual surface flow 82 mm 155 mm

Maximum annual mean flow (1978) 170 mm 318 mm

Simulated annual catchment streamflow, 1989 land cover (mm)

Simulated annual catchment streamflow, 2002 land cover (mm)

0

100

200

300

400

1977

1978

1979

1980

1981

Catchment annual streamflow (mm)

Simulated annual catchment streamflow, 1989 land cover (mm)

Simulated annual catchment streamflow, 2002 land cover (mm)

0

100

200

300

400

1977

1978

1979

1980

1981

Catchment annual streamflow (mm)

Figure 47: Simulated annual catchment streamflow for 1989 and 2002 land cover

Average annual catchment streamflows are directly related to land cover type, soil

characteristics and annual precipitation. In the study area, subsistence agricultural areas

have increased between 1989 and 2002, with most of the increase occurring in previously

135

vegetated areas of savanna and forest. Agricultural land has the highest potential for runoff

because the land is kept bare at the onset of the rainy season. AVSWATX simulates surface

run-off volumes and peak run-off rates for each HRU. Surface run-off is estimated with a

modification of the SCS curve number method [USDA-SCS, 1986]. In this study,

subsistence agricultural land has a curve number of 86 compared to 79 and 68 for savanna

and forest. The curve number is a dimensionless parameter indicating the runoff response

of a drainage basin. Changing land cover results in a different run-off curve number, which

could result in changes in rainfall run-off responses. High curve numbers signify high

surface run-off and low infiltration.

The total annual precipitation for 1977 to 1981 varies between 800 and 1500 mm with a

standard deviation of 23% from the long-term mean of 1090 mm. Annual variation of

rainfall from 1976 to 1981 from the long-term mean is shown in Figure 48. 1978 was an

excessively wet year (variance 47% above average), preceded by moderately high rainfall

in 1976 and a year of average rainfall in 1977. However, 1981 was a significantly dry year.

-40

-30

-20

-10

0

10

20

30

40

50

1976

1977

1978

1979

1980

1981Variance %

-40

-30

-20

-10

0

10

20

30

40

50

1976

1977

1978

1979

1980

1981Variance %

Figure 48: Rainfall variability between 1977 and 1981, referenced against long-term mean (1976 – 2002)

The catchment streamflow response analysis of the study area indicated that high rainfall

controls much of the increase in streamflow in the Shire River catchment. Taking the year

136

1976 as a lead in the period of intensive study, the recorded above average rainfall (Figure

48) is assumed to have contributed to saturated soil moisture storage. Any increments of

rainfall on already saturated soils contribute to increases in surface run-off, hence the

increase in streamflow in 1977 and 1978 (Figure 47). This trend is also related to 2002

land cover, which is dominated by subsistence agriculture, having recorded the highest

average annual catchment streamflow of greater than 300 mm in 1978, a year of 47%

above average rainfall.

The rainfall regime in 1979 is characterised by the onset of rainfall in mid-October,

average monthly rainfall below 200 mm in December, dry spells in January and average

monthly rainfall of less than 250 mm in March and April. This particular year corresponds

to decreasing total average annual streamflow of 135 mm for the 2002 land cover

compared to 112 mm for the 1989 land cover. In 1980 there was a slight increase in the

mean annual rainfall received (1 039 mm). However, two rainfall stations (Mangochi and

Balaka) had the lowest total rainfall (~550 mm and ~700 mm respectively). In this year,

there was an increase in catchment streamflow: 1989 streamflow is 23 mm lower than that

in 2002. This could be attributed to change in land cover from savanna to subsistence

agriculture and the increase in precipitation.

Another noticeable observation in the rainfall pattern is the decrease in simulated annual

catchment streamflow in 1981. The year was characterised by slightly below normal

precipitation with dry spells in January and low rains in February and March. Average

annual rainfall was ~750 mm. The study finds that the decrease in precipitation combined

with drying of the soils explains a gradual reduction in annual catchment streamflow (1979

– 1981) for the 2002 land cover. In contrast, for the 1989 land cover with higher forest

cover, the relative decrease is much smaller, possibly due to greater soil moisture retention

and increased infiltration.

Furthermore, the soil data for the study area indicates that the area is composed of the Soil

Conservation Service (SCS) hydrological group C. These soils are fine to very fine

textured and are generally shallow. They have a low to very low rate of water infiltration

when wet, which results in high run-off potential. Streamflow patterns in this catchment

are strongly influenced by the seasonal cycle of rainwater. Annual streamflows occur in

response to timing and degree of precipitation and the land cover characteristics.

137

To understand the flow processes during different seasons under different land cover

conditions, the average monthly streamflows were plotted for the wet and dry season and

compared. In the Shire River catchment, there are two seasons - wet weather occurs from

November to March and dry weather events occur between April and October. This two-

season climate creates significant differences in streamflow. Seasonal variations predicted

from the two land cover classifications (1989 and 2002) are presented in Figure 49.

From the modeling process, HRUs from the 1989 land cover are dominated by savanna,

while HRUs from the 2002 land cover are dominated by subsistence agriculture. The

average monthly streamflow shows differences between the two simulations (Figure 49).

For the 1989 land cover average monthly streamflow was in the range of 14 m3

s-1

to

31 m3

s-1

, while that of 2002 land cover data was between 16 m3

s-1

and 57 m3 s

-1. The

minimum average monthly catchment streamflow simulated for 1989 is 19 m3

s-1

, while

that of 2002 is 30 m3

s-1

. In general, storm events from the 2002 land cover yield both high

surface run-off rates and high total water volumes. The majority of peak flows occur

during the months of November to March, which is the rainy season in the study area.

0

20

40

60

80

100

Jan Mar May Jul Sep Nov Jan Mar May Jul Sep Nov

0

20

40

60

80

100

l min max h avg

1989 2002

Average monthly streamflow (m3s-1)

0

20

40

60

80

100

Jan Mar May Jul Sep Nov Jan Mar May Jul Sep Nov

0

20

40

60

80

100

l min max h avg

1989 2002

Average monthly streamflow (m3s-1)

Figure 49: Monthly mean, standard deviations and maxima of daily simulated catchment streamflows for 1989 and 2002 land cover simulations

138

The dominance of savanna landscapes in 1989 yielded relatively low average monthly

catchment streamflows. In comparison, the 2002 land cover, which is dominated by

agricultural land yielded high average monthly catchment streamflows.

A comparison of the multi-year daily catchment streamflows generated using 1989 and

2002 land covers respectively is presented in Table 31 and Figure 50

Table 31: Parameters obtained from daily simulations for 1989 and 2002 land cover

Item 1989 land cover data 2002 land cover data

Maximum daily flow 83 m3 s-1 154 m

3 s-1

Minimum daily flow 1 m3 s-1 3 m

3 s-1

Average daily flow 19 m3 s-1 30 m

3 s-1

Simulated daily catchment streamflow,1989 land cover (m3 s-1)

Simulated daily catchment streamflow, 2002 land cover (m3 s-1)

Catchment streamflow (m3s-1)

0

40

80

120

160

200

Jan-77

May-77

Sep-77

Jan-78

May-78

Sep-78

Jan-79

May-79

Sep-79

Jan-80

May-80

Sep-80

Jan-81

May-81

Sep-81

Simulated daily catchment streamflow,1989 land cover (m3 s-1)

Simulated daily catchment streamflow, 2002 land cover (m3 s-1)

Catchment streamflow (m3s-1)

0

40

80

120

160

200

Jan-77

May-77

Sep-77

Jan-78

May-78

Sep-78

Jan-79

May-79

Sep-79

Jan-80

May-80

Sep-80

Jan-81

May-81

Sep-81

Figure 50: Comparison of simulated daily catchment streamflows for 1989 and 2002 land cover data

The simulation results demonstrate that there are differences in the daily peak flows

between the 1989 land cover and 2002 land cover. The hydrograph generated for the 1989

land cover produced the highest daily peak flow of 83 m3

s-1

, whereas the 2002 land cover

produced the highest daily peak of 154 m3

s-1

(Figure 50). However, an anomaly was

observed during the dry season of 1979 to 1981. The model failed to capture the dry season

139

variations for both the 1989 and 2002 land cover, which simulated a uniform baseline.

Possible reasons have been outlined in Section 4.3.2 as these model inefficiencies were

also observed during model calibration within this study. The daily catchment streamflow

analysis for each year of the study period indicates high flow peaks at the beginning of the

rainy season, implying that rainfall on saturated soils causes much of the flow. Saturated

soils are efficient at producing more streamflow. High surface flow signifies low

infiltration during the rainy season and consequently diminishes ground water

replenishment during the dry season.

Total streamflow is composed of surface run-off and baseflow (lateral flow and shallow

ground water discharge to streams). Comparisons were also done to evaluate differences in

surface flow and baseflow from different land cover types classified from the 1989 and the

2002 land cover data sets. Between 1989 and 2002, dominant land cover changes were

observed in sub-basins 2, 4, 5, 6, 7, 9, 11, 12 and 13, while sub-basins 1, 3, 8 and 10 did

not change (Figure 40). Table 32 and Table 33 show differences in surface run-off,

baseflow and the percentage changes.

Table 32: Surface run-off simulated from 1989 and 2002 land cover

Simulated average annual surface run-off (mm)

Change (mm)

Percentage change (%) Sub-basin

1989 2002

1 51 51 0 0

2 42 149 107 255

3 Reservoir 0 0 0

4 51 175 124 243

5 39 147 108 277

6 52 175 123 237

7 43 4 -39 -91

8 52 52 0 0

9 42 149 107 255

10 149 149 0 72

11 88 235 147 167

12 11 267 256 2327

13 88 235 147 167

Entire Shire River catchment

82 155 73 89

The highest annual surface run-off of 267 mm was generated in densely populated

settlement areas in sub-basin 12. The dominant land cover in sub-basin 12 was forest in

140

1989 and had changed to subsistence agriculture in 2002. Thus, surface run-off increased

by 2 327% from 11 mm to 267 mm, while baseflow decreased by -48% from 190 mm to

98 mm. Subsistence agricultural land is characterised by scattered grasslands, used for

communal grazing during the dry season. At the onset of the rainy season, most of these

fields are bare soil that have been prepared for planting, or denuded by grazing. Most of

the run-off generated in the cultivated and grazing lands constitutes storm flow, especially

at the beginning of the rainy season between November and January.

Table 33: Baseflow simulated from 1989 and 2002 land cover

Simulated average annual baseflow (mm)

Change (mm)

Percentage change (%) Sub-basin

1989 2002

1 93 93 0 0

2 66 30 -36 -55

3 Reservoir 0 0 0

4 90 55 -35 -39

5 83 57 -26 -46

6 100 63 -37 -37

7 70 84 14 20

8 99 99 0 0

9 65 29 -36 -55

10 29 29 0 0

11 101 52 -49 -49

12 190 98 -92 -48

13 101 58 -43 -43

Entire Shire River catchment

59 74 15 25

Similar observations have been noted in sub-basins 2, 4, 5, 6, 9, 11 and 13 where dominant

land cover has changed from savanna shrubs (RNGB) to subsistence agriculture (AGRL).

Under the simulated conditions of a fixed climatic scenario, with only the measured land

cover changes inserted, simulated surface run-off increased by 255%, 243%, 277%, 237%,

255%, 167% and 167% in sub-basins 2, 4, 5, 6, 9, 11 and 13 respectively. Proportionately,

baseflow decreased by -55%, -39%, -46%, -37%, -55%, -49% and -43% respectively. The

high variability between sub-basins indicates the need to segment catchments into these

smaller units, to understand the full extent of land cover changes.

Large portions of the original woodlands have been cleared. This is due to expansion of

agricultural land for food production, fuelwood and construction. The majority of

141

subsistence farmers in Malawi practice traditional methods of cultivation. Soil and water

conservation technologies are not practised, and generally the adoption rate for most land

husbandry technologies is low [Malawi Government, 2001]. Hence, when agricultural land

becomes impoverished, it is common practise to clear fresh forest or savanna.

With less plant cover, more rainfall runs off the surface rather than infiltrates into the

ground. If the run-off from a storm is greater, the chance of the flow exceeding the stream

capacity and causing flooding increases. In addition, water that runs off does not have a

chance to recharge groundwater. Groundwater flows slowly into streams, usually over a

period of months, providing steady baseflow (flow in streams in times without rainfall).

This sequence of events is a cause of concern as it makes a major difference to stream

characteristics and health, causing increased streambank erosion due to higher peak flows

during the rainy season and periods of very low flow due to the decreased baseflow during

the dry season.

It is notable that the dominant land cover in sub-basin 7 changed from savanna shrubs

(RNGB) to forest (FRSD) in 2002. Sub-basin 7 is located within the valley of the Shire

River, which represents the smallest sub-basin (79 km2). The change to dominant forest for

the 2002 land cover yielded the minimum annual surface run-off (dropping from 43 mm to

4 mm), proportionate with high baseflow (rising from 70 mm to 84 mm) compared to the

other sub-basins. Changes in land cover may have a significant effect on available water

resources and could determine the amount of water that flows in rivers. Forested areas

have good infiltration, a large baseflow component and small storm flow. Forests absorb

water, retain it, release it slowly, and have low erosion rates [Calder, 2000].

Comparable studies have been done to assess the ability of the SWAT model to analyse the

effects of land cover changes on streamflow. Tadele et al. [2007] investigated land cover

dynamics and its impacts on streamflow at Hare River watershed, southern Rift Valley

Lakes Basin, Ethiopia. In order to asses the variability of streamflow due to the land cover

dynamics from 1975 to 2004, the AVSWATX model was run using two land cover maps

(1992 and 2004), while all the other input variables were similar for both simulations. It

was identified that mean monthly discharge for wet months had increased by 13% while in

the dry season they decreased by up to 31% during the 1992 – 2004 period due to land

cover change.

142

A critical aspect of this study has been to establish the links between extensive land cover

change and hydrological responses. Simulations were performed by substituting 1989 land

cover with 2002 land cover classifications while keeping climate constant. Large portions

of the Shire River catchment have been transformed from savanna and forest to subsistence

agricultural land between 1989 and 2002. Simulation results revealed that streamflow

patterns in this catchment have been strongly influenced by the land cover changes that

have occurred over the relatively short period of thirteen years.

Further, it has been demonstrated that subsistence agricultural areas generated higher

surface run-off than savanna and forest areas. Areas dominated by subsistence agriculture

are generally bare at the onset of the rainy season, and hence less effective in reducing run-

off. Higher surface run-off and less infiltration are not conducive of sustainability for rural

agriculture as they erode fertile surface soils and make areas more vulnerable to periods of

drought. The use of small-scale irrigation agriculture to combat cyclical food insecurity

may be hampered by the lack of perennial water flows.

Additionally, water in the Shire River is almost totally allocated to hydro-electric power

stations in the mid-catchment, irrigation schemes in the lower-catchment and domestic

purposes [Malawi Government, 2001]. Water shortages during drought periods may

therefore increase water contestation among different users.

This study has also established that increasing forest and savanna area yields less surface

run-off compared to that of areas under subsistence agriculture. There is also an increase in

baseflow, which contributes to ground water replenishment. Thus, forest areas are the

preferred land cover type with regard to the sustainability of the water balance for water

resources management.

4.3.3 Scenario outcomes

Simulations were performed using three scenarios, which were compared to the baseline

case. The baseline case was taken as the simulation using measured climate (1977 to 1981)

and the 2002 land cover image. The scenarios are the case of bounding scenarios (which

were used to test the sensitivity of AVSWATX in simulating changes in land cover

characteristics), land degradation and land conservation specified in Table 25 and Table

26 respectively. The hydrological responses at the outlet thus simulated were compared to

the baseline case in terms of average annual catchment streamflows, surface run-off and

baseflows.

143

Bounding scenarios

Changes in land cover cause changes in curve number, the internal model parameter that

determines the partitioning of precipitation into surface flow and percolation. Evaluation of

model sensitivity, through the trial of extreme land cover scenarios, revealed the maximum

degree and patterns of interaction between changes in land cover and modelled

hydrological responses.

The results of the two extreme scenarios are compared to the baseline case (Table 34).

Total deforestation of the entire land surface of the catchment generated a total water yield

of 463 mm a-1

, compared to 397 mm a-1

for the reference scenario. In the case of total

deforestation, there is increase in surface flow of 437 mm a-1

compared to 322 mm a-1

, and

a decrease in baseflow of 26 mm a-1

in comparison with 74 mm a-1

for the reference

scenario. Bare areas have a strong effect by promoting rapid run-off and thereby reducing

percolation. Ground storage is reduced and surface direct evaporation enhanced.

For total forestation, total water yield reduced to 308 mm a-1

. In the case of total land cover

change to forest, there is a decrease in surface flow from 322 mm a-1

to 195 mm a-1

and

increase in baseflow from 74 mm a-1

to 113 mm a-1

. Forests absorb most of the

precipitation hence there is increased interception, percolation and evapotranspiration,

rather than prompt streamflow.

Results of this sensitivity analysis show the maximum changes that could be expected from

extreme changes in land cover, and hence bounding conditions for the further scenario

modelling. The model behaves as expected, with bare land increasing prompt run-off and

reducing percolation, and conversely reducing run-off and total yield for full forestation.

Table 34: Simulation results from bounding cases scenarios

Total water yield (mm) Surface flow (mm) Baseflow (mm)

2002 land cover 397 322 74

Total change to bare soil 463 437 26

Total change to forest 308 195 113

Land degradation

Simulation runs were performed using land cover from scenarios termed land degradation,

which represents an unfavourable scenario with accelerated land cover change with

144

extensive deforestation. Table 35 lists the annual values of catchment streamflow,

baseflow and surface flow for each case as simulated from the land degradation scenarios.

Table 35: Simulation results from land degradation scenarios

A graphical representation of the baseflow results is shown in Figure 51 while Figure 52

shows the surface flow results.

0

20

40

60

80

10% RNGB

converted to

AGRL

20% RNGB

converted to

AGRL

40% RNGB

converted to

AGRL

Baseflow (mm)

0

20

40

60

80

0

20

40

60

80

10% FRSD

converted to

AGRL

20% FRSD

converted to

AGRL

40%FRSD

converted to

AGRL

Baseflow (mm)

0

10

20

30

40

Baseflow (mm) Total area converted 2002 Baseflow

Total area changed (103 ha)

Total area changed (103ha)

0

20

40

60

80

10% RNGB

converted to

AGRL

20% RNGB

converted to

AGRL

40% RNGB

converted to

AGRL

Baseflow (mm)

0

20

40

60

80

0

20

40

60

80

10% FRSD

converted to

AGRL

20% FRSD

converted to

AGRL

40%FRSD

converted to

AGRL

Baseflow (mm)

0

10

20

30

40

Baseflow (mm) Total area converted 2002 Baseflow

Total area changed (103 ha)

Total area changed (103ha)

Figure 51: Baseflow simulation results obtained from land degradation scenarios

Decreasing land under savanna and forest cover, differences in total annual catchment

streamflows, baseflows and surface flows were observed. Under the existing land cover

Scenario Land cover change

Total water yield (mm)

% Change

Surface flow (mm)

% Change

Baseflow (mm)

% Change

2002 baseline case 397 322 74

1 10% from RNGB

397 0 323 0 74 0

2 20% from RNGB

397 0 326 0 70 1.2

3 40% from RNGB

402 1.4 334 1.9 68 3.7

4 10% from FRSD

394 -0.7 321 -1.0 72 -0.3

5 20% from FRSD

396 -0.3 325 -0.3 70 0.9

6 40% from FRSD

397 0.2 327 0.3 70 1.5

145

(reference scenario, 2002), the annual catchment streamflow observed was 397 mm, with

baseflow making up to 19% and the remaining 81% from surface flow. In terms of

absolute mean annual changes between the reference state and the scenarios, decreasing

forest areas by different magnitudes increases surface flow and decreases baseflow and

total water yield at the outlet.

A 10% decrease in the savanna areas increases total water yield from 397 mm a-1

to

402 mm a-1

. With a 10% decrease in savanna, baseflow remains at 74 mm a-1

while surface

flow was 323 mm a-1

with minor difference from the reference scenario at 322 mm a-1

.

However, decreasing savanna areas by 20% produces an increase in total water yield and

surface flow but a decrease in baseflow. Total water yield increased from 397 mm a-1

to

402 mm a-1

while surface flow increased from 322 mm a-1

to 334 mm a-1

. A minor

decrease in baseflow has been observed once savanna was decreased by 20%. The

conversion of savanna into subsistence agriculture does not cause considerable changes in

the macro scale basin. The reason for this may be the similarity of the characteristics of

both land cover forms under tropical natural conditions. Savanna vegetation has scattered

trees and is deciduous during the dry season. Thus, a considerable amount of land is

exposed, similar to subsistence agricultural land.

0

100

200

300

400

10% RNGB

converted to

AGRL

20% RNGB

converted to

AGRL

40% RNGB

converted to

AGRL

Surface flow (mm)

0

20

40

60

80

Surface flow (mm) Total area converted 2002 Surfaceflow

0

100

200

300

400

10% FRSD

converted to

AGRL

20% FRSD

converted to

AGRL

40%FRSD

converted to

AGRL

Surface flow (mm)

0

20

40

60

80

Total area changed (103ha)

Total area changed (103ha)

0

100

200

300

400

10% RNGB

converted to

AGRL

20% RNGB

converted to

AGRL

40% RNGB

converted to

AGRL

Surface flow (mm)

0

20

40

60

80

Surface flow (mm) Total area converted 2002 Surfaceflow

0

100

200

300

400

10% FRSD

converted to

AGRL

20% FRSD

converted to

AGRL

40%FRSD

converted to

AGRL

Surface flow (mm)

0

20

40

60

80

Total area changed (103ha)

Total area changed (103ha)

Figure 52: Surface flow simulation results obtained from land degradation scenarios

A 40% decrease in forest yielded total water yield of 397 mm a-1

, not different from the

reference scenario. Slight decreases in baseflow ranging from 70 mm a-1

to 72 mm a-1

have

been observed with an increase in the magnitude of deforestation. For example, a 10%

decrease in forest results in a decrease in baseflow from 74 mm a-1

to 72 mm a-1

, while a

146

40% decrease in forest produces a decrease in baseflow 70 mm a-1

. Surface flows were

observed to be increasing as forestland is reduced. For example, a 40% decrease in forest

increases surface flow from 322 mm a-1

to 327 mm a-1

.

The relative effect of woodland (savanna and forest) reduction (expressed as percentages)

on the annual water balance is compared to that of subsistence agriculture. The 10%

decrease in forest reduces total water yield by 0.7%, whereas a 20% reduction of forest

leads to a decline of water yield by 0.3%. Overall, the strongest relative impact can be

observed in the amount of surface run-off. Decreasing land under savanna cover by 20%

increases surface flow by 4% while decreasing forest by 40% increases surface flow by

2%. This is due to the increase in land under subsistence farming. The absence of trees and

shrubs implies a minimum in surface evapotranspiration and, consequently, a maximum in

run-off. The soil is less protected against raindrop impact under agriculture since after

harvesting and shortly after sowing, when the plants do not cover the soil completely.

Land conservation scenario

Simulation runs were also performed for land cover scenarios termed land conservation,

which represents the creation of a greener environment through management and re-

forestation. Table 36 lists the annual volumes of catchment streamflow, baseflow and

surface flow for each scenario as simulated from land conservation scenarios. The

conversion of subsistence agricultural land into woodlands (savanna and forests) leads to

changes in the water balance.

Table 36: Simulation results obtained from land conservation scenarios

Scenario Land cover change

Total water yield (mm)

% Change

Surface flow (mm)

% Change

Baseflow (mm)

% Change

2002 baseline 397 322 74

7 20% to RNGB 373 -6 284 -8 89 -12

8 40% to RNGB 332 -16 217 -22 115 -33

9 20% to FRSD 398 0.5 298 0.2 99 -8

10 40% to FRSD 397 0.2 173 0.6 225 -46

11 20% both RNGB and FRSD

371 -12 192 -16 158 -41

12 40% both RNGB and FRSD

350 -6 161 -9 210 -50

147

A 40% increase in forest increases total water yield from 397 mm a-1

to 398 mm a-1

.

However, the proportion of water infiltrating increases from 74 mm a-1

to 225 mm a-1

while surface flow decreased from 322 mm a-1

to 173 mm a-1

. Slight increases in baseflow

have been observed with a 20% increase in forest, when baseflow increases from 74 mm a-

1 to 99 mm a

-1. Surface flows were observed to decrease as savanna expands. For example,

a 40% increase in savanna, results in a drop in total water yield from 397 mm a-1

to

332 mm a-1

. This reduction is associated with an increase in baseflow from 74 mm a-1

to

115 mm a-1

. Increasing the areas of forest and savanna decreases water yield and surface

flow, and increases baseflow. For instance, an increase of 20% decreases total water yield

to 371 mm a-1

while a 40% increase decreases water yield to 350 mm a-1

compared to

397 mm a-1

from the reference scenario. The reduction of the mean annual flow results in a

decreasing surface flow during the rainy season with a simultaneous increase in baseflow.

Graphical representations of the changes for baseflow are shown in Figure 53, while for

surface flow they are presented in Figure 53 while for surface flow they are presented in

Figure 54.

0

50

100

150

200

250

20% AGRL converted to

FRSD

40% AGRL converted to

FRSD

Baseflow (mm)

-100

-80

-60

-40

-20

0

Baseflow (mm) 2002 Baseflow Total area converted

0

50

100

150

20% AGRL converted to

RNGB

40% AGRL converted to

RNGB

Baseflow (mm)

-150

-100

-50

0

Total area changed (103ha)

Total area changed (103ha)

0

50

100

150

200

250

20% AGRL converted to

FRSD

40% AGRL converted to

FRSD

Baseflow (mm)

-100

-80

-60

-40

-20

0

Baseflow (mm) 2002 Baseflow Total area converted

0

50

100

150

20% AGRL converted to

RNGB

40% AGRL converted to

RNGB

Baseflow (mm)

-150

-100

-50

0

Total area changed (103ha)

Total area changed (103ha)

Figure 53: Baseflow simulation results from land conservation scenarios

148

Surface flow (mm) Total area converted 2002 Surfaceflow

0

100

200

300

400

20% AGRL converted to

FRSD

40% AGRL converted to

FRSD

Surface flow (mm)

-100

-50

0

0

100

200

300

400

20% AGRL converted to

RNGB

40% AGRL converted to

RNGB

Surface flow (mm)

-100

-50

0

Total area changed (103ha)

Total area changed (103 ha)

Surface flow (mm) Total area converted 2002 Surfaceflow

0

100

200

300

400

20% AGRL converted to

FRSD

40% AGRL converted to

FRSD

Surface flow (mm)

-100

-50

0

0

100

200

300

400

20% AGRL converted to

RNGB

40% AGRL converted to

RNGB

Surface flow (mm)

-100

-50

0

Total area changed (103ha)

Total area changed (103 ha)

Figure 54: Surface flow simulation results from land conservation scenarios

Increasing woodland (both savanna and forest) has a profound effect on the annual water

balance when compared to the reference scenario (Figure 55). The 20% increase in the

forest increases total water yield by 0.5% whereas an additional 20% increase of forest

leads to a decline in water yield by 0.2%. This could be due to the small amount of land

initially under forest.

0

100

200

300

400

20% AGRL converted to

FRSD and RNGB

40% AGRL converted to

FRSD and RNGB

Surface flow (mm)

-200

-150

-100

-50

0

Surface flow (mm) Total area converted 2002 Surfaceflow

0

50

100

150

200

250

20% AGRL converted to

FRSD and RNGB

40% AGRL converted to

FRSD and RNGB

Baseflow (mm)

-250

-200

-150

-100

-50

0

Baseflow (mm) 2002 Baseflow Total area converted

Total area converted (103ha)

Total area converted (103ha)

0

100

200

300

400

20% AGRL converted to

FRSD and RNGB

40% AGRL converted to

FRSD and RNGB

Surface flow (mm)

-200

-150

-100

-50

0

Surface flow (mm) Total area converted 2002 Surfaceflow

0

50

100

150

200

250

20% AGRL converted to

FRSD and RNGB

40% AGRL converted to

FRSD and RNGB

Baseflow (mm)

-250

-200

-150

-100

-50

0

Baseflow (mm) 2002 Baseflow Total area converted

Total area converted (103ha)

Total area converted (103ha)

Figure 55: Baseflow and surface flow simulation results from land conservation scenarios

Generally, the strongest relative impact can be observed in the amount of surface flow.

Increasing land under savanna cover by 40% decreases surface flow by 33%, while

increasing the savanna area by 20% decreases surface flow by 12%. Differences in surface

flow and baseflow were also observed with increases of both forest and savanna areas by

149

40% and 20% respectively. Reductions in surface flow of 50% and 40% were observed

respectively. This is due to the decrease in land under subsistence farming. Baseflow

increases and surface flow decreases are due to the higher interception of forests and

savanna woodlands in comparison to maize and legumes. Observational and experimental

studies have established that forests consume more water during evapotransipiration than

any other forms of land cover [Dingman, 2008]. Large reductions of the river discharge

and therefore a considerable increase of water retention in the catchment would occur only

in the case of forestation of large areas. Therefore, depending on the magnitude of

percentage change in forest cover, forestation would decrease average and dry-season

streamflow.

Seasonal variability

Streamflow and baseflow vary significantly during the year because of seasonal weather

changes, coupled to variations in land cover characteristics. The seasonal water balance

analyses suggest that when the river discharge increases in the Shire River in November-

March, the amount of maximum catchment streamflow decreases with increase in

reforested areas (Figure 56). In the scenario where only savanna replaces subsistence

agricultural land by 20% (scenario 7), the maximum catchment streamflow is 259 m3 s

-1.

An increase of 40% yielded a maximum catchment streamflow of 216 m3 s

-1 compared to a

maximum catchment streamflow of 281 m3 s

-1 for the 2002 land cover. On average, when

there is a 20% savanna increase, monthly catchment streamflow ranges from 13 m3 s

-1 to

151 m3 s

-1, while a 40% savanna increase yields monthly catchment streamflow averages

ranging between 17 m3 s

-1 and 124 m

3 s

-1. Average monthly flows for the reference

scenario range from 12 m3 s

-1 to 166 m

3 s

-1. Overall, the seasonal pattern of streamflow

responds to the precipitation pattern, with greatest streamflow and baseflow occurring in

the November to March period.

150

Average monthly streamflow (m3s-1)

0

50

100

150

200

250

300

Sc7

Jan

Sc7

Mar

Sc7

May

Sc7

Jul

Sc7

Sep

Sc7

Nov

Sc8

Jan

Sc8

Mar

Sc8

May

Sc8

Jul

Sc8

Sep

Sc8

Nov

0

50

100

150

200

250

300

l min max h Reference avg

Average monthly streamflow (m3s-1)

0

50

100

150

200

250

300

Sc7

Jan

Sc7

Mar

Sc7

May

Sc7

Jul

Sc7

Sep

Sc7

Nov

Sc8

Jan

Sc8

Mar

Sc8

May

Sc8

Jul

Sc8

Sep

Sc8

Nov

0

50

100

150

200

250

300

l min max h Reference avg

Figure 56: Monthly mean, standard deviations and maxima of daily simulated catchment streamflows for 2002 land cover scenario 7 and 8 simulations

The model was used also to evaluate the potential effects of increasing forest areas in

regions initially covered by subsistence agriculture. The seasonal water balance analyses

shows that when land under forest vegetation increases, dry season flow increases and wet

season peak flow decreases compared to the reference land cover scenario (Figure 57). The

maximum catchment streamflow decreases and there is an increase in the scale of

reforestation. When forest areas were increased by 20% (scenario 9), catchment

streamflow yielded a mean monthly maximum catchment streamflow of 276 m3 s

-1 in

February, while a 40% forest increase (scenario 10) yielded 230 m3 s

-1. On average, with a

20% forest increase, monthly catchment streamflow averages range from 16 m3 s

-1 to

159 m3 s

-1, while a 40% forest increase yields monthly catchment streamflow averages

between 31 m3 s

-1 and 123 m

3 s

-1. Average monthly flows for the reference scenario range

from 12 m3 s

-1 to 166 m

3 s

-1. A comparison of the water balance components over the two

seasons for the different reforestation magnitudes indicates that the contribution of

baseflow during the dry season is high, while maximum catchment streamflow is

considerably lower than the reference scenario.

151

0

50

100

150

200

250

300

Sc9

Jan

Sc9

Mar

Sc9

May

Sc9

Jul

Sc9

Sep

Sc9

Nov

Sc10

Jan

Sc10

Mar

Sc10

May

Sc10

Jul

Sc10

Sep

Sc10

Nov

0

50

100

150

200

250

300

l min max h Reference avg

Average monthly streamflow (m3s-1)

0

50

100

150

200

250

300

Sc9

Jan

Sc9

Mar

Sc9

May

Sc9

Jul

Sc9

Sep

Sc9

Nov

Sc10

Jan

Sc10

Mar

Sc10

May

Sc10

Jul

Sc10

Sep

Sc10

Nov

0

50

100

150

200

250

300

l min max h Reference avg

Average monthly streamflow (m3s-1)

Figure 57: Monthly mean, standard deviations and maxima of daily simulated catchment streamflows for 2002 land cover, scenario 9 and 10 simulations

Impact simulations were carried out for the analyses of the current situation (reference

scenario) considering land cover changes towards land conservation. Scenario 11 and 12

correspond to conversion of potentially vulnerable areas (subsistence agricultural land) into

forest and savanna woodlands. Figure 58 shows the simulation results of the seasonal

variations under 20% and 40% increases in both forest and savanna land cover scenarios.

During the low flow period from April to October, dry season flow increases compared to

the reference land cover scenario when there is large scale reforestation. On average, with

a 20% increase in forest and savanna, monthly averages range from 24 m3 s

-1 to 118 m

3 s

-1,

while a 40% forest and savanna increase yields monthly flow averages between 32 m3 s

-1

and 111 m3 s

-1. Average monthly flows for the reference scenario range from 12 m

3 s

-1 to

166 m3 s

-1. Seasonal differences are also detected in terms of maximum flows during the

wet season. The reference land cover yields a maximum flow of 281 m3 s

-1 while

reforesting land by 20% and 40% yields flows of 206 m3 s

-1 and 193 m

3 s

-1 respectively.

Large-scale reforestation can significantly reduce average annual water flow in rivers and

affect the seasonal distribution flow.

152

0

50

100

150

200

250

Sc11

Jan

Sc11

Mar

Sc11

May

Sc11

Jul

Sc11

Sep

Sc11

Nov

Sc12

Jan

Sc12

Mar

Sc12

May

Sc12

Jul

Sc12

Sep

Sc12

Nov

0

50

100

150

200

250

l min max h avg Reference

Average monthly streamflow (m3s-1)

0

50

100

150

200

250

Sc11

Jan

Sc11

Mar

Sc11

May

Sc11

Jul

Sc11

Sep

Sc11

Nov

Sc12

Jan

Sc12

Mar

Sc12

May

Sc12

Jul

Sc12

Sep

Sc12

Nov

0

50

100

150

200

250

l min max h avg Reference

Average monthly streamflow (m3s-1)

Figure 58: Monthly mean, standard deviations and maxima of daily simulated catchment streamflows for 2002 land cover, scenario 11 and 12 simulations

Similar studies have been performed with the SWAT model to simulate the hydrological

behaviour of catchment areas using hypothetical land cover scenarios [Eckhardt et al.,

2003; Forher et al., 2001; Huisman et al., 2004] For example, simulations were run to

analyse the effects of the hydrological response of a catchment to different land use options

(Fohrer, 2001). In this study, SWAT was applied for the Dietzhölze watershed in Germany,

with two land use scenarios using the 1992 land cover data as a reference scenario. The

model results showed that the decrease of forest and corresponding increase in grassland

increased the peak flow rate (from 72 mm a-1

to 126 mm a-1

), and thus increases the risk of

flooding. The expansion of grassland and cropland in formerly forested areas increased

baseflow by 2%. A minor increase in run-off was observed due to the additional cropland

proportion. Baseflow decreased by 8% and total streamflow was reduced by 14 mm a-1

in

comparison with the actual 1994 land use.

Comprehensive hydrological models like AVSWATX can thus give valuable information,

which can be incorporated in catchment management studies. The hydrological processes

(surface flow, baseflow and infiltration) are largely driven by the nature and density of

land cover and the type of land cover over a catchment. In this study, the applicability of

AVSWATX in different contexts of catchment management has been explored. The capacity

to predict the effects of future land cover changes is very important for future use and

management strategies in the Shire River catchment.

153

4.4 Conclusion

The spatial semi-distributed hydrological model, AVSWATX, in combination with GIS, has

been applied to the upper Shire River basin, the largest surface water resource in Malawi.

The model was used to simulate water balance and river flows using digital elevation; soil

and land cover data; and five years (1977 to 1981) of observed daily precipitation,

temperature, wind speed and relative humidity data. The model was successful in

reproducing streamflow within the limits of observational and modelling errors. The study

showed that surface-water model parameters are sensitive and have physical meaning,

especially the CN2, ESCO and SOL_AWC. From the results obtained, it has been

concluded that the model has relatively high confidence and gives a good representation of

the water balance and outflow hydrographs at the basin outlet. The model performance

(1977-1981) using the Nash-Sutcliffe efficiency for reproducing catchment streamflow is

86% and 42% for monthly and daily calibrations respectively.

The fact that the optimal evaluation results have not been achieved could be a result of the

limited global input parameters and the complexity of optimisation of some of these

parameters. The lack of high resolution DEM, soil maps and a dense network of

precipitation and weather input data stations for the Shire River basin compromised the

simulations especially for the daily runs. There was generally an underestimation of

baseflows, especially during the dry months, which could be attributed to limited detailed

spatial rainfall data availability. Low yielding weathered Precambrian basement gneiss

complex formations within the Shire rift valley region could also have affected ability of

the model to simulate the lowest baseflows. In spite of these limitations, the model

captured the dynamic of flow generation well, with surface run-off dominating during the

rainy season and shallow aquifer contributing during the dry months. In addition, the

AVSWATX model provided insight in the main flow processes during the year. The model

appears to be suitable for application to large tropical river basins with ungauged stations.

The use of ArcView GIS with the AVSWATX model enabled the performance of quicker

hydrological analyses, especially for large basins, using the semi-distributed model.

In the present work, the influences of hypothetical land cover change characteristics upon

the hydrological processes that precipitation undergoes on delivery to the land surface were

also investigated. All of the flow parameters discussed – streamflows, baseflow and

surface flow are affected by the amount of plant cover, which is influenced by density and

spatial distribution. The results of the analysis have highlighted the sensitivity of surface

154

flow and baseflow in response to the suggested decrease in woodlands. Consequently,

seasonal variations in river discharge have been associated with decreased infiltration

during the wet season, which can affect the streamflow during the dry season by inhibiting

ground water recharge. Evaluation of effects of forestation on flow variability has

demonstrated that land covered with forest decreases surface flow and increases ground

water recharge.

Hydrological processes are an integrated indicator of catchment conditions, and changes in

land cover may affect the overall health and functioning of a catchment. An understanding

of temporal changes and trends in streamflow and the proportion of surface flow and

baseflow is critical for directing efforts in managing land cover and improving agricultural

practices. Exploring land cover change scenarios provides a wider applicability for

assessing the effects of land cover or land management change and hydrological responses.

If policies for land use management are to be established, methods must be available to

demonstrate that change has occurred, and what the nature and source of the change

involve. This method of evaluating of the effects of land management on water availability

can be used when planning for sustainable land and water resources management.

155

Chapter 5

5 CONCLUSION AND RECOMMENDATIONS

This final chapter summarises the main research findings and recommends future

investigations to advance the field of physically based hydrological modelling for the

management of water resources in Malawi.

5.1 Conclusion

The Shire River provides an important water resource for the socio-economic development

of Malawi. Currently, there is concern about the growing competition for water resources

within the catchment. With increasing human activities, it is important to understand

interactions between hydrological regimes and associated land use, and land cover change

in the catchment. Such interventions can be achieved by integrating land use planning and

water resources management. Therefore, a comprehensive assessment of the spatial and

temporal distribution of land cover change, as well as impacts on the land - hydrological

processes are required to resolve present problems and avoid potential crises in future

water resource allocations. At national level, results of this study can be used to improve

land use management to achieve sustainable water utilisation in the Shire River catchment.

This study was based on the hypothesis that unsustainable changes in land cover due to

human activities are significantly altering aggregate catchment conditions, giving rise to

long-term, potentially irreversible changes in river flow characteristics. The study was

aimed at answering the question: What are the effects of significant land cover changes

over the past two decades on river flow characteristics that are important for water

resources, environmental functioning and hydrological processes within the Shire River

catchment? Land cover changes associated with growing human populations and expected

changes in climatic conditions are likely to accelerate alterations in hydrological

phenomena and processes on various scales. Consequently, these changes significantly

influence the quantity and quality of water resources for nature and human society. This

aim was further pursued in the context of developing integrated land use planning and

water resources management in Malawi.

The basis of this research comprised multi-temporal classification of Landsat satellite

imagery (1989 and 2002) to provide a recent perspective of land cover types and changes

within the Shire River catchment of Malawi. Results from this study indicate successfully

156

the synergy between Landsat data, vector data and detailed ground information in mapping

land cover features. It was found that by integrating contextual information and ancillary

data, discrimination of heterogeneous land cover classes dominant in savanna woodlands

was improved. Eight land cover classes, comprising cultivated or grazing lands,

grasslands, savanna shrubs, marshes, woody open, woody closed, built-up areas and fresh

water, were mapped for the Shire River catchment, with an overall accuracy of 87%. The

land cover maps generated have compatible digital formats, hence they can be applied

easily to a variety of future GIS applications. Additional themes can be incorporated as

more resource information becomes available, or as new management needs are identified.

To map the land cover variables, a hierarchical legend structure determined by the Food

and Agriculture Organisation (FAO) Land Cover Classification System (LCCS)

(FAO/LCCS) was used [Di Gregorio and Jansen, 2005]. LCCS was used as a basis for the

classification to achieve legend harmonisation within Africa and on a global scale.

Flexibility of the hierarchical system allowed incorporation of digital elevation objects, soil

and underlying geological features as well as other available geographical datasets. By

integrating contextual information and ancillary data, classification accuracy was

improved. The new FAO/LCCS classification is also internally consistent, allowing

scalability and updatability that can be used at different scales and different levels of detail

to discriminate land cover features. In addition to compatibility with the FAO/LCCS, the

derived land cover maps have provided recent and improved classification accuracy, added

thematic detail compared to the most recent existing 1992 land cover maps for Malawi.

Accurate and up-to-date land cover change information is necessary to provide an

improved spatial representation of land cover data for hydrological parameterisation and

modelling, including other applications, to reflect actual land cover conditions.

Results from the mapping were used to analyse land cover change between 1989 and 2002.

Performing change detection analysis for the Shire River catchment revealed distinct

patterns in land cover change regarding disturbance and fragmentation of the landscape.

Carrying out this study in Malawi, therefore, has provided valuable information to evaluate

land cover dynamics and the percentage of cover change in the catchment. The change

analysis showed an increase of 4 622 ha (7.9%) in cultivated or grazing lands emanating

from declining natural vegetated areas. The main driving factors are subsistence

agricultural expansion and demand for wood resources. As agriculture continues to play a

157

dominant role in land cover conversion and degradation from Brachystegia woodlands,

more open and drier vegetation formations will continue to evolve. To this end, it is

apparent that the rapid increase in cultivated areas within the catchment will not only

decrease the amount of forests and vegetated areas but also increase run-off potential. This

process will diminish the overall quality and quantity of water resources. The mapping and

change detection exercise has not only characterised the landscape but will develop our

understanding of how to manage it more effectively. Distinguishing and quantifying where

potentially risky land cover changes occur is critical to the initiation of regular monitoring

of resources and the environment in general. Especially in Africa and Malawi in particular,

rigorous and standardised land cover change data have been missing prior to this study,

with which to parameterise predictive hydrological models and for other applications.

National policies and new upcoming environmental directives require a more frequent

update of the extent, rate and direction of such land cover changes.

Fundamental to land cover changes is their impact on the surface hydrology within

catchments of large rivers. In this study, the AVSWATX hydrological model was used to

simulate water balance and streamflows of the Shire River catchment. Furthermore,

AVSWATX hydrological model was used to evaluate effects of derived land cover changes

on hydrological processes. Model calibration was performed by using digital elevation

data, soil and land cover data, and five years (1977 to 1981) of observed daily

precipitation, temperature, wind speed and relative humidity data. The model was

successful, within limits of observational and modelling errors, in reproducing streamflow.

From the results, it has been concluded that the model has a relatively high confidence and

gives an adequate representation of the water balance and outflow hydrographs at the basin

outlet. The model performance (1977-1981) using the Nash-Sutcliffe efficiency for

reproducing streamflows was 86% and 42% for monthly and daily calibrations

respectively.

In this study, the applicability of AVSWATX hydrological model in different contexts of

catchment management was revealed. In particular, the possibility of predicting impacts of

land cover changes for deciding on future uses of the Shire River catchment was

demonstrated. Land cover change scenarios were generated in which fractional changes

were made to land cover, pessimistic – changing forest to subsistence agriculture, and

optimistic – partial restoration of subsistence agriculture to savanna and forest. Simulation

158

results for the Shire River catchment indicated that increasing subsistence agricultural

areas and simultaneous declines of woodland resulted in increased annual and event

surface flow volumes. Parameters potentially sensitive to negative impacts, in this study

being: any increases in surface run-off, increases in storm flow and/or declines in

groundwater percolation, have been identified. A scenario in which savanna areas were

decreased by 20% indicated annual average increases of total water yield, but decreased

baseflow. Total water yield increased from 397 mm a-1

to 402 mm a-1

while surface flow

increased from 322 mm a-1

to 334 mm a-1

. Under a future “greening” scenario, the model

results indicate that improvements to the sustainability of catchment hydrology can be

expected. The notable changes observed in this scenario include decreases in total water

yield and surface flow, and a corresponding increase of percolation into the groundwater

table. In a scenario of increasing forested areas, catchment streamflow decreased from

322 mm a-1

for the reference case, to 276 mm a-1

for a 20% forest area increase, and to

230 mm a-1

for a 40% forest area increase. Land cover characteristics are one of the major

factors affecting hydrological processes of catchment. Scenario analyses such as this one

improve our ability to make informed decisions and policies regarding land and water

resource management. The results on the simulation of total water yield, surface flow and

baseflow have demonstrated that catchment management (rehabilitation) is imperative to

reduce surface run-off, increase infiltration and therefore, sustainability of water resources.

The incorporation of remote sensing, Geographical Information Systems (GIS) and

ArcView Soil and Water Assessment Tool eXtendable (AVSWATX) model provides a

powerful tool for assessing the impacts of land management on river flow patterns and

irrigation water availability. Remote sensing has the ability of viewing and repetitive

coverage, which provides useful information on land cover dynamics. GIS is an efficient

tool for presentation of input data as required by hydrological models. Therefore, using

remotely sensed data, GIS and AVSWATX to simulate the run-off process and total water

yield is more advantageous when the study area is large as was the case in this research.

Comprehensive hydrological models like AVSWATX can thus give valuable information,

which can be incorporated in catchment management studies in Malawi. Hydrological

effects of changes in land cover are difficult to discern in the case of large-basins that have

a variety of land cover classes and vegetation in various stages of regeneration. In addition,

spatial and temporal rainfall variations may exist across decades. Observational studies of

159

the effects of land cover conversions on the hydrology of large river basins are scarce,

especially in the tropics. The lack of studies of very large and persistent land cover

conversions suggests that if appropriate land cover, precipitation and discharge data were

available, it would be possible to determine whether the impact of land cover change

across very large catchments is similar to that observed in smaller catchments

Furthermore, this study has integrated land cover and water resources management within

the framework of Integrated Development Planning. Integrated development planning

involves balancing the integrity of landscape and hydrological characteristics to achieve

sustainable long-term management. This approach will enhance understanding of the

ongoing land cover changes, processes and perhaps predict future trends in the upper Shire

River catchment. Within such an approach, proper policies of water resources, water

demand management, water availability for small-scale irrigation and sustainable

agricultural practices can then be developed.

5.2 Recommendations

Land cover mapping and change detection studies are valuable especially for water

quantity and water quality predictions, and assessing the hydrological effects of such

changes. However, much research remains to be done to improve upon the results of land

cover mapping and change detection. Land cover studies could be improved by using

images acquired from different seasons of the year and additional years to avoid the

snapshot situation. Inadequacy of streamflow records and a limited number of gauging

stations became the most serious limitation in efforts to model the Shire River catchment.

The challenge for further studies is the need for quality assurance and prompt quality

control on routine hydrological and climatic data in government agencies to obtain long-

term ecological records (LTER). More probable would be the applicability of research

findings in the context of sustainability and governance of natural resources in Malawi and

more broadly.

To ensure that the next attempt to synthesize land cover data and hydrological variables at

a local and global scale avoids the shortcomings and pitfalls identified in the current

exercise, some of the priorities for future observations and research are outlined. It is

anticipated, that with all these improvements, future research could provide a robust

hydrological model that could be used with greater confidence in planning and decision-

making.

160

In this study, some of the land cover classes presented particular problems for mapping

from satellite imagery. From a remote sensing point of view, discriminating open mixed

deciduous forest with an area cleared under cultivation or logging activities proved to be

difficult. During the dry season when there is little chlorophyll in the vegetation, grazing

causes exposure of soil between remaining vegetation. This resulted in similarity between

spectral values, making it difficult to distinguish the two classes. Future studies in such

kind of environments must consider wet and dry season images so that features that may

not be captured in one season are captured in the other. In this case, other sources of

satellite images are recommended such as Synthetic Aperture Radar (SAR) may be useful

since they are able to penetrate through cloud cover during the rainy season. In addition,

multi-temporal data merging could also help to discriminate certain land cover classes.

In the process of land cover mapping, an important observation was also made considering

the amount of cleared vegetated areas. It is possible that some of the smaller cleared areas

of less than a couple of hectares were perhaps not detected as subsistence agricultural, and

assigned to other land cover class. This potential misclassification is a result of inherent

limitations of the spatial resolution of Landsat images (30 m x 30 m pixels). In addition,

spectral resolution is also characteristic of inherent limitations by assuming that every land

class incorporated in this study has a constant and consistent spectral signature. Future

studies should consider incorporating natural processes affecting landscape change such as

fire and natural vegetative succession, which could also account for vegetation loss. As

advancements in technology are made, there are refinements of spatial, spectral, temporal,

and radiometric resolution of sensors, and more efficient image processing methods. For

example, hyperspectral images that provides resolution sufficient to identify and

distinguish spectrally similar but distinct materials. These sensors provide the potential for

more accurate and detailed information extraction than is possible with earlier multi-

spectral images, such as Landsat. Recently launched hyperspectral space-borne sensors

include Hyperion on NASA’s EO-1 satellite, the CHRIS sensor on the European Space

Agency’s PROBA satellite and the FTHSI sensor on the U.S. Air Force Research Lab’s

MightySat II satellite. These developments will allow for more accurate change detection.

Keeping track of these changes is important to our understanding of the Earth as a system

i.e. knowledge of why and where the changes occur.

FAO/LCCS land cover classification system was launched in 2005. The new classification

system is internally consistent, allows scalability and mappability that can be used at

161

different scales and levels of detail to discriminate land cover features. Based on successful

application of the FAO/LCCS classification system in the study, I recommend wider

adoption of this classification system in African land cover mapping. However, to capture

the heterogeneity and detail of landscapes requires intense and rigorous fieldwork. This

makes validation of land cover maps difficult with limited financial resources.

Modelling efforts were hampered by lack of consistent and continuous streamflow data.

The study recommends the need for dual measurements and additional measuring gauging

stations to improve long-term streamflow data collection in the study region for

hydrological modelling. There is need to explore automatic recording at gauging stations to

minimise the potential for incorporation of invalid data. In addition, to provide streamflow

information to meet national needs, the information obtained from stream gauges needs to

be consistent, obtained using standard techniques and technology, and be subject to

standardised quality assurance and quality control procedures. Each stream gauge site

should be required to enter and maintain basic metadata (i.e. documentation) about the site,

stations, and station variables in the database. After the metadata has been entered for each

station, the data should undergo several levels of quality control analysis to identify

questionable values. Values flagged as invalid (e.g. consistently similar values or

unexceptionally high fluctuations of one or two days) are removed from data sets prior to

archiving or distribution.

Systematic, consistent measurements of streamflow and climatic data must be administered

by well-trained and competent staff for effective management of data sets. This study

recommends formation of long-term collaboration between the University of Malawi, and

Departments responsible for water resources and meteorology (Department of Water and

the Department of Meteorology). The goal of this partnership is to provide opportunities to

increase the supply and retention of graduates from the University of Malawi for positions

with the Malawi Government and international partners. The establishment and

development of such relationships will improve professional skills and satisfy the goals of

both the government and the University for socio-economic development.

This is the first study of its kind in Malawi, such that other studies are required to

complement findings in this study with other techniques. One option would be to

investigate the use of remote sensing (River and Lake Altimetry Radar) to take lake and

river flow measurements to constrain uncertainties from stream gauging stations. From

162

such water level measurements, river discharge can be calculated, hence providing a more

reliable streamflow data set compared to the (seriously flawed) manually recorded data

used in the this study.

Accurate rainfall data for the catchment are very critical for streamflow prediction.

Numerous river basins in the world are characterised by limited measurements of key

hydrological parameters such as precipitation, especially in southern Africa, Malawi

inclusive. Therefore, this study recommends use of spatially distributed rainfall from other

sources such as meteorological satellites may produce an improvement over the data-

scarce areas. Radar-sensed rainfall can be expected to have better spatial accuracy than a

sparse network of rain-gauge stations. Some of the relevant space-borne instruments are

the Advanced Very High Resolution Radiometer (AVHRR) series on board the National

Oceanic and Atmospheric Administration (NOAA), METEOSAT Second Generation

(MSG) with its Spinning Enhanced Visible and Infrared Imager (SEVIRI) and the

Precipitation Radar (PR) on Tropical Rainfall Measuring Mission (TRMM).

For long-term investigations, the interactions of climate change and land cover become

essential to assess the effects related to hydrology. Therefore, future work on the impact of

land cover change should be extended by the consideration of anticipated climate change,

its modification of hydrological processes and the feedback on land cover.

An active management strategy aimed at the conservation and regeneration of the natural

vegetation is recommended to improve the distribution of water throughout the entire Shire

River catchment during both dry and wet periods. Degraded lands should be afforested and

more intense monitoring of the on-going agroforestry and forestation efforts. It is also

recommended to start reclamation of degraded savanna lands and forests in the catchment,

particularly in the southern and eastern hilly areas. These play an important role in the

generation of run-off due to low soil permeability. This is expected to significantly

increase baseflow and decrease storm flow events. Environmental management planning

should explicitly deal with protection of natural forests on steeply sloping sections of the

landscape.

It is estimated that 89% of the population in Malawi is rural and depends on subsistence

agriculture and natural resources for their livelihoods [National Statistical Office, 2000].

This study recommends exploration of possibilities of establishing other economic

163

activities in the area and rural areas in general. This has been outlined in the Millennium

Development Goals and the “Malawi Vision 2020” document to empower rural

communities. Such activities could be mineral exploration, small-scale rural business

adventures and a general industrialisation to reduce the overdependence on subsistence

agriculture. The University of Malawi could therefore provide expertise in terms of

management and training of service providers. This could be a step in the right direction

for the sustainability of natural resources in the country.

Small-scale irrigation projects should be encouraged especially in areas where water is

available. Communities must be provided with facilities for water harvesting for dry

season farming, while minimising surface run-off, and unlimited expansion onto virgin

savanna and forested areas.

More intense extension education is required for rural communities on land and water

conservation to minimise soil loss. Proper land management practices must be enforced,

for example, in agroforestry, contouring on steep sloping areas to increase water retention.

The decrease in soil quality due to present subsistence agricultural practices also has vital

implications for the sustainability of the landscape and water resources. It calls for

measures to restore soil fertility. The use of improved fallows, composting and the practice

of agroforestry might be useful.

5.3 Concluding remarks

The study has shown that the condition of water resources in the Shire River catchment,

Malawi, has been affected adversely by rapid changes in land cover over the last two

decades. This research has contributed to narrowing an important knowledge gap, by

establishing a land use planning and water resources management nexus in a quantitative

manner. It is essential to combine landscape change analyses with hydrological parameters

for improved understanding of the processes of land cover change and hydrological

variables. This helps in linking patterns to processes, and in designing policy interventions

aimed at reducing the unfavorable effects of land cover change on hydrology. Such

developments are needed for sustainable water utilisation and food security if the

Millennium Development Goals (MDGs) are to be realised in Malawi and other

developing countries.

164

Lack of studies showing land cover changes at local levels is one of the major problems

frustrating policy makers in their attempt to adopt sustainable development efforts. While

focus has been on global land cover change, there is lack of empirical studies on land cover

changes. Combining spatial data and modern tools of remote sensing has provided insight

into the scale of land cover change from a landscape perspective. Analyses of land cover

dynamics at landscape scale grasp the complexity of events. This research has contributed

to a better understanding of the amount of environmental degradation on landscape spatial

scales, occurring on short (decadal) time scales. The results from this research are

particularly relevant for Malawi, and in other developing countries that have similar

conditions to Malawi, and similar lack of historical environmental monitoring records.

The present study has adopted the hierarchical legend structure determined by the Food

and Agriculture Organisation (FAO) Land Cover Classification System (LCCS) to label

land cover variables. LCCS is a relatively new classification system, which was launched

in 2005. No land cover mapping study to date has been conducted in Malawi using LCCS

as a coding classification system. This study serves as one of the initial studies to test the

applicability of LCCS in southern Africa. The adoption of LCCS carried out in this

research may thus be an important step towards a rigorous update and translation of the

existing land cover maps for Malawi and the SADC region. This then leads to the

compatibility of land cover maps generated in Malawi, other parts of Africa and the world

at large for regional and international applications.

The research has also contributed to the sustainability land cover change and hydrology

literature by providing evidence of strong linkages between land cover, land use change

and streamflow response. This has implications for food, water and (fuelwood) energy

security on rural and peri-urban Africa. The application of modern tools of remote sensing

combined with advanced numerical hydrological models (such as AVSWATX) have been

demonstrated to be in developing tools for monitoring and managing land cover and

hydrology in data scarce regions. These research results provide a platform for selecting

variables and identifying functional relationships in dynamic, process-based hydrological

models. Research processes developed in this study can be applied in cost-effective ways

by environmental managers in land-use management and hydrological decision-support

systems, and for policy formulation. Thus, the results from this research are particularly

relevant for Malawi and in other developing countries in formulating, implementing and

monitoring strategies for sustainable development.

165

References

Abbot, J., and K. Homewood (1999), A history of change: causes of miombo woodland

decline in a protected area in Malawi, Journal of Applied Ecology, 36 422-433.

Abbot, P. G., C. F. Khori, and J. D. Lowore (1995), Management of Miombo by local

communities, edited by F. R. I. Malawi, Zomba, Malawi.

Acreman, M., J. Almagro, J. Alvarez, F. Bouraoui, R. Bradford, J. Bromley, B. Croke, S.

Crooks, J. Cruces, J. Dolz, M. Dunbar, T. Estrela, P.Fernandez-Carrasco, J. Fornes, G.

Gustard, R. Haverkamp, de l.H.A., N. Hernandez-Mora, R. Llamas, C.L.Martinez, J.

Papamasorakis, R. Ragab, M.Sanchez, I. Vardavas, and T.Webb (2000), Groundwater

and River Resources Programme on a European Scale (GRAPES), Institute of

Hydrology, Wallingford, UK.

Alwashe, M. A., and A. Y. Bokhari (1993), Monitoring vegetation changes in Al Madinah,

Saudi Arabia, using Thematic Mapper data, International Journal of Remote Sensing,

14, 191-197.

Andersen, J., J. C. Refsgaard, and K. H. Jensen (2001), Distributed hydrological modeling

of the Senegal River Basin – model construction and validation, Journal of Hydrology,

247, 200-214.

Anderson, J. R., E. E. Hardy, J. T. Roach, and R. E. Witmer (1976), A land use and land

cover classification system for use with remote sensor data, edited, U.S. Geological

Survey, USGS, Washington D.C.

Anyamba, A., C. O. Justice, C. J. Tucker, and R. Mahoney (2003), Seasonal to interannual

variability of vegetation and fires at SAFARI 2000 sites from advanced very high

resolution radiometer time series data, Journal of Geophysical Research, 108.

Arnold, C. L., and C. J. Gibbons (1996), Impervious surface coverage: the emergence of a

key environmental indicator, Journal of American Planning Association, 62, 243-258.

Arnold, J. G., and P. M. Allen (1999), Automated methods for estimating baseflow and

ground water recharge from streamflow records Journal of the American Water

Resources Association, 35, 411-424.

Arnold, J. G., P. M. Allen, R. Muttiah, and G. Bernhardt (1995), Automated base flow

separation and recession analysis techniques, Ground Water, 33, 1010-1018.

Arnold, J. G., and N. Fohrer (2005), SWAT 2000, current capabilities and research

opportunities in applied watershed modeling, Hydrological Processes, 19, 563-572.

Arnold, J. G., R. Srinivasan, R. S. Muttiah, and J. R. Williams (1998), Large area

hydrologic modeling and assessment part I: model development, Journal of American

Water Resources Association, 34, 73-89.

Arnold, J. G., J. R. Williams, A. D. Nicks, and N. B. Sammons (1990), SWRRB, a basin

scale simulation model for soil and water resources management, Texas A&M

University Press, Texas.

Arnold, J. G., J. R. Williams, R. Srinivasan, K. W. King, and R. H. Griggs (1994), SWAT:

Soil Water Assessment Tool, U. S. Department of Agriculture, Agricultural Research

Service, Grassland, Soil and Water Research Laboratory, Temple, TX.

166

Augenstein, E., D. Stow, and A. Hope (1991), Evaluation of SPOT HRV-XS data for help

resource inventories, Photogrammetric Engineering and Remote Sensing, 57, 501-509.

Bastable, H. G., W. J. Shuttleworth, R. L. G. Dallarosa, G. Fisch, and C. A. Nobre (1993),

Observation of climate, albedo and surface radiation over cleared and undisturbed

Amazonian forest, International Journal of Climatology, 13, 783-796.

Bath, A. H. (1980), Hydrochemistry in groundwater Development: Report on an advisory

visit to Malawi, Institution of Geological Sciences.

Benedini, E., F. Cammillozzi, and A. Martinelli (2003), Model SWAT application in the

Alban Hills (Central Italy), paper presented at 2nd International SWAT conference,

TWRI Technical Report 266, Bari, Italy.

Biging, G., D. Colby, and R. G. Congalton (1998), Sampling systems for change detection

accuracy assessment in Remote sensing change detection: Environmental monitoring

methods and applications, edited by R. S. Lunetta and C. D. Eldvidge, Ann Arbor

Press, Chelsea, MI.

Birhanu, B. Z. (2005), Application of a GIS based SWAT model in simulating the

available water resources in a Pangani basin subcatchment, University of Dar es

Salaam, Dar es Salaam.

Boardman, J., and D. T. Favis-Mortlock (1993), Climate change and soil erosion in Britain,

Journal of Geographical, 159, 179-183.

Booth, D. B. (1991), Urbanization and the natural drainage system-impacts, solutions and

prognoses, Northwest Environmental Journal, 7, 93-118.

Bormann, H. (2005), Regional hydrological modeling in Benin (West Africa): Uncertainty

issues versus scenarios of expected future environmental change, Physics and

Chemistry of the Earth, 30, 472-484

Bosch, J. M. (1979), Treatment effects on annual and dry period stream flow at Cathedral

Peak, Journal of South African Forest, 108, 29-38.

Bouten, W., P. J. F. Smart, and E. D. Water (1991), Microwave transmission, a new tool in

forest hydrological research, Journal of Hydrology, 124 119-130.

Bradley, P. N., and K. McNamara (1993), Living with trees: Policies for forestry

management in Zimbabwe, World Bank Technical Paper, Washington DC.

Brooks, K. N., P. F. Ffolliot, H. M. Gregersen, and J. L. Thames (1991), Hydrology and

the Management of Watersheds, Iowa State Univeristy Press, Iowa.

Bruinjzeel, L. A. (2001), Hydrology of tropical montane cloud forests: a Reassessment

Land Use and Water resources research (LUWRR)

Burges, S. J., M. S. Wigmosta, and J. M. Meena (1998), Hydrologic effects of land-use

change in a zero-order catchment, Journal of Hydrologic Engineering, 3, 86-97.

Busche, H., C. Hiepe, and B. Diekkruger (2005), Modelling the effects of Land use and

climate change on hydrology and soil erosion in a sub-humid African catchment, paper

presented at Proceedings of the 3rd International SWAT conference, EAWAG-Zurich,

Switzerland, 11th –15th July 2005.

Calder, I. R. (1992), Hydrological Effects of Land Use Change, in Handbook of

Hydrology, edited by D. R. Maidment, pp. 13.11-13.15, McGraw-Hill, New York.

Calder, I. R. (1998), Water resources and land use issues, International Water Management

Institute, Colombo, Sri Lanka.

Calder, I. R. (2000), Land use impacts on water resources, Background paper No. 1, Land-

water linkage in rural watersheds.

167

Calder, I. R. (2002), Forests and hydrological services: Reconciling science and public

perceptions, Land use and water resources research 2, 2.1-2.12.

Cao, W., W. B. Bowden, T. Davie, and A. Fenemor (2003), Application of SWAT in a

large mountanous with high spatial variability, paper presented at 2nd International

SWAT conference, TWRI Technical Report 266, Bari, Italy, July 1-4.

Cheng, S., and R. Wang (2002), An approach for evaluating the hydrological effects of

urbanisation and its application, Hydrological Processes, 16, 1403-1418.

Chow, V. T., D. R. Maidment, and L. W. Mays (Eds.) (1988), Applied hydrology,

McGraw-Hill, New York.

Cihlar, J. (2000), Land cover mapping of large areas from satellites: status and research

priorities, International Journal of Remote Sensing, 21, 1093-1114.

Cihlar, J., and L. J. M. Jansen (2001), From land Cover to Land Use: A Methodology for

Efficient Land Use mapping over large areas, The Professional Geographer, 53, 275-

289.

Congalton, R. G. (1991), A Review of Assessing the Accuracy of Classifications of

Remotely Sensed Data, Remote Sensing of Environment, 37, 35-46.

Congalton, R. G., and K. Green (1999), Assessing the Accuracy of Remotely Sensed Data:

Principles and Practices, Lewis Publishers, New York.

Consultative Group on International Agricultural Research Digital Elevation Model data,

edited.

Coppin, P., I. Jonckheere, K. Nackaerts, B. Muys, and E. Lambin (2004), Digital change

detection methods in ecosystem monitoring: a review, International Journal of Remote

Sensing, 25, 1565-1596.

Corves, C., and C. J. Place (1994), Mapping the reliability of satellite-derived landcover

maps-an example from the Central Brazilian Amazon Basin, International Journal of

Remote Sensing, 15, 1283-1294.

Crapper, P., and K. Hynson (1983), Change detection using Landsat photographic imagery,

Remote Sensing of the Environment, 13, 291-300.

Culf, A. D., G. Fisch, and M. G. Hodnett (1995), The Albedo of Amazonian Forest and

Ranchland, Journal of Climate, 8, 1544-1554.

Dagnachew, L., C. Vallet-Coulomb, and F. Gasse (2003), Hydrological response of a

catchment to climate and land use changes in Tropical Africa: case study South Central

Ethiopia, Journal of Hydrology, 275, 67-85.

Dale, P. E. R., A. L. Chandica, and M. Evans (1996), Using image subtraction and

classification to evaluate change in subtropical intertidal wetlands, International

Journal of Remote Sensing, 17, 703-719.

Dauze, S. E. K., F. D. Vescovi, G. Menz, and P. L. G. Vlek (2001), Land use/cover of the

Volta basin of Ghana, 2000, paper presented at Status conference GLOWA, Munich,

6th -8th May, 2002.

De Jong, S. M. (1994), Derivation of vegetative variables from a Landsat TM image for

modeling soil erosion Earth Surface Processes and Landforms 19, 165-178.

De Roo, A., M. Odijk, G. Schmuck, E. Koster, and A. Lucieer (2001), Assessing the

effects of land use changes on floods in the Meuse and Oder catchment, Physics and

Chemistry of the Earth (B), 26, 593-599.

DeFries R., and K. Eshleman (2004), Land use change and hydrological processes: a major

focus for the future (invited commentary), Hydrological Processes, 18, 2183-2186.

168

Desanker, P. V., P. G. Frost, C. O. Justice, and R. J. Scholes (1997), The Miombo

Network: Framework fro a terrestrial transect study of land use and land cover change

in the Miombo ecosystem of central Africa, International Geosphere- Biosphere

Programme, Stolkholm.

Dewees, P. A. (1994), Social and Economic aspects of Miombo woodland management in

southern Africa: Options and Opportunities for research, Center for International

Forestry Research, Bogor, Indonesia.

Di Gregorio, A., and L. J. M. Jansen (2000), Land cover classification system LCCS:

Classification concepts and use manual; FAO Environment and Natural Resources

Service, FAO Land and Water Development Division, Rome.

Di Gregorio, A., and L. J. M. Jansen (2005), Land cover classification system LCCS:

Classification concepts and use manual; FAO Environment and Natural Resources

Service, FAO Land and Water Development Division, Rome.

Di, K., R. Ma, and R. Li (2003), Geometric processing of IKONOS Geo stereo imagery for

coastal mapping applications, Photogrammetric Engineering and Remote Sensing, 69,

873-879.

Di Luzio, M., R. Srinivasan, J. G. Arnol, and S. Neitsch (2001), Arcview Interface for

SWAT 2000 User’s Guide.

Dingman, S. L. (2002), Physical Hydrology, 2nd ed., Prentice Hall.

Dingman, S. L. (2008), Physical hydrology, 2nd ed., Waveland press,inc., Long Grove,

Illinois.

Dinicola, R. S. (1990), Characterization and simulation of rainfall-runoff relations for

headwater basins in western King and Snohomish Counties, 52 pp, Geological Survey

Water-Resources Investigations Report, Washington: U.S.

Dolman, A. J., and A. Verhagen (Eds.) (2003), Land Use and Global Environment

Change, in Global Environmental Change and Land Use, Kluwer Academic

Publishers, Dordrecht, Netherlands.

Duan, Q., S. Sorooshian, and V. K. Gupta (1992), Effective and efficient global

optimisation for conceptual rainfall-runoff models, Water Resources Research, 28,

1015-1031.

Eckhardt, D. W., J. P. Verdin, and G. R. Lyford (1990), Automated update of an irrigation

lands GIS using SPOT HRV imagery, Photogrammetric Engineering and Remote

Sensing, 56, 1515–1522.

Eckhardt, K., and J. G. Arnold (2001), Automatic calibration of a distributed catchment

model, Journal of Hydrology 251, 103-109.

Eckhardt, K., L. Breuerb, and H.-G. Frede (2003), Parameter uncertainty and the

significance of simulated land use change effects, Journal of Hydrology, 273, 164-176.

Evans, R. (1990), Water erosion in British farmers’ fields–some causes, impacts,

predictions, Progress in Physical Geography, 14, 199-219.

Famiglietti, J. S., and E. F. Wood (1994), Multi-scale modeling of spatially variable water

and energy balance processes, Water Resources Research, 30, 3061-3078.

FAO/UNESCO (2003), Digital Soil Map of the World and Derived Soil Properties,

Version 3.6. CD-ROM, Information Division, FAO, Viale delle Terme di Caracalla,

00100 Rome, Italy.

Food and Agriculture Organisation (2005), Land Cover Classification System:

Classification concepts and User manual, 190 pp, FAO, Rome.

169

Foody, G. M. (2001), Monitoring the Magnitude of Land-Cover Change around the

Southern Limits of the Sahara, Photogrammetric Engineering and Remote Sensing, 67,

841-847.

Forher, N., S. Haverkamp, K. Eckhardt, and H. Frede (2001), Hydrological response to

land use changes on the catchment scale, Physics and Chemistry of the Earth (B), 26,

577-582.

Franklin, J. F. (1992), Scientific basis for new perspectives in forests and streams, in

Watershed Management, edited by R. J. Naiman, Springer-Verlag, New York.

Frost, P. G. H. (Ed.) (1996), The Ecology of Miombo Woodlands in The Miombo in

Transition: Woodlands and Welfare in Africa Centre for International Forestry

Research, Bogor, Indonesia.

Fung, L., and E. LeDrew (1987), Application of principal components analysis to change

detection, Photogrammetric Engineering and Remote Sensing, 53, 1649–1658.

Fung, T. (1990), An assessment of TM imagery for land-cover change detection, IEEE

Transactions on Geoscience and Remote Sensing, 28, 681- 692.

Gan, T. Y. (1988), Application of scientific modelling of hydrological response from

hypothetical small catchments to assess a complex conceptual rainfall runoff model,

Department of Civil Engineering, University of Washington, Seattle, Washington.

Geist, H. J., and E. F. Lambin (2002), Proximate causes and underlying forces of tropical

deforestation, BioScience 52, 143-150.

Giambelluca, T. W., M. A. Nullet, A. D. Zieglet, and L.Tran (2000), Latent and sensible

heat energy flux over deforested land surfaces in the eastern Amazon and northern

Thailand, Singapore Journal of Tropical Geography, 21, 107-130.

Global Land Cover Facility (2005), Landsat Imagery, edited.

Global Water Partnership (2005), Sustatinable management of water resources, edited.

Green, K., D. Kempka, and L. Lackey (1994), Using remote sensing to detect and monitor

land-cover and land-use change, Photogrammetric Engineering and Remote Sensing,

60, 331–337.

Green, R., and S. Nanthambwe (1992), Land resources appraisal of the agricultural

development divisions, MOA/UNDP/FAO, MLW/85/011, Lilongwe.

Green, W. H., and G. A. Ampt (1911), Studies on soil physics, 1.The flow of air and water

through soils, Journal of Agricultural Science, 4, 11-24.

Hall, R. L., and I. R. Calder (1993), Drop size modification by forest canopies-

measurements using a disdrometer, Journal of Geographical Research, 90, 465-470.

Hargreaves, G. H., and Z. A. Samani (1985), Reference crop evapotranspiration from

temperature, Applied Engineering in Agriculture, 1.

Headstreams Project (2004), Headstreams, sustainable management and development of

SADC countries, edited, Okavango Research Center, Maun, Botswana.

Hollis, G. E. (1975), The effects of urbanization on floods of different recurrence intervals,

Water Resources Research, 11, 431-435.

Howarth, P. J., and G. M. Wickware (1981), Procedures for change detection using

Landsat digital data, International Journal of Remote Sensing, 2, 277–291.

Hudak, A., and C. A. Wessman (2000), Deforestation in Mwanza District, Malawi, from

1981 to 1992, as determined from Landsat MSS imagery, Journal of Applied

Geography, 20, 155-175.

170

Huisman, J. A., L. Breuer, and H.-G. Frede (2004), Sensitivity of simulated hydrological

fluxes towards changes in soil properties in response to land use change, Physics and

Chemistry of the Earth, 29, 749-758.

Hunt, C. E. (2004), Thirsty planet: Strategies for sustainable water management, Zed

books, London.

Hutcheson, M. A. (1998), Africa south of the Sahara: Physical and social geography,

Malawi, 27th Edition ed., Europe Publications Limited, London.

IPCC (2001), Climate Change 2001: Impacts, Adaptation and Vulnerability, Contribution

of the Working Group II to the Third Assessment Report of Intergovernmental Panel on

Climate Change, Cambridge University Press, Cambridge.

Jansen, L. J. M., and A. Di Gregorio (2002), Parametric land cover and land use

classification as tools for environmental change detection, Agriculture, Ecosystems and

Environment, 91, 89-100.

Jennings, D. B., and S. T. Jarnagin (2002), Changes in Anthropogenic Impervious

Surfaces, Precipitation and Daily Streamflow Discharge: A Historical Perspective in a

Mid-Atlantic Subwatershed, Landscape Ecology, 17, 471-489.

Jensen, J. R. (1986), Introductory digital image processing, Prentice Hall, New Jersey.

Jensen, J. R. (1995), Introductory Digital Image Processing: a remote sensing perspective,

Prentice-Hall, New Jersey.

Jensen, J. R. (1996), Introductory digital image processing: A remote sensing perspective,

2nd ed., Upper Saddle River, NJ, Prentice Hall.

Jensen, J. R. (2005), Introductory digital image processing: A remote sensing perspective,

Prentice Hall, USA, USA.

Jensen, J. R., S. Narumalani, O. Weatherbee, and J. R. Mackey (1993), Measurement of

seasonal and yearly cattail and waterlily changes using multi date SPOT panchromatic

data, Photogrammetric Engineering and Remote Sensing, 59, 519–525.

Jipp, P. H., D. C. Nepstad, D. K. Cassel, and C. R. d. Carvalho (1998), Deep soil moisture

storage and transpiration in forests and pastures of seasonally-dry Amazonia, Climatic

Change, 39, 395–412.

Justice, C. O., J. R. G. Townshend, E. F. Vermote, E. Masouka, R. E. Wolfe, N. Saleous,

D. P. Roy, and J. T. Morisette (2002), An Overview of MODIS Land data processing

and products status, Remote Sensing of the Environment, 83, 3-15.

Kafakoma, R. (1996), Environmental rehabilitation of the refugee impacted areas, Malawi,

paper presented at World Bank/UNEP Africa forestry forum, Unpublished, Nairobi,

Kenya.

Kalipeni, E. (1996), Demographic Response to Environmental Pressure in Malawi,

Population and Environment, 17, 285-308.

Kaluwa, P. W. R., F. M. Mtambo, and R. Fachi (1997), The country situation report on

water resources in Malawi, UNDP/SADC Water Initiative, Lilongwe, Malawi.

Kauth, R. J., and G. S. Thomas (1976), The Tasseled Cap – A graphic description of the

spectral-temporal development of agricultural crops as seen by Landsat, paper

presented at Proceedings symposium on machine processing remotely sensed data,

Indiana.4B 41-50, Purdue University.

Kepner, W. G., D. J. Semmens, S. D. Basset, D. A. Mouat, and D. C. Goodrich (2004),

Scenario analysis for the San Pedro River, analyzing hydrological consequences for a

future environment, Environmental Modeling and Assessment, 94, 115-127.

171

Kidd, C. H. R. (1983), A water resources evaluation of Lake Malawi and Shire River,

World Meteorological Organisation, Geneva.

Lahmer, W., B. Pfutzer, and A. Becker (2001), Assessment of land use and climate change

impacts on the mesoscale, Physics and Chemistry of the Earth (B), 26, 565-575.

Lambin, E. F., and H. A. Strahler (1994), Indicators of land-cover change for change-

vector analysis in multitemporal space at coarse spatial scales, International Journal of

Remote Sensing, 15, 2099-2119.

Lambin, E. F. a. H. G. (2006), Land-use and land-cover change: local processes and

global impacts, 204 pp., Springer, Berlin, Heidelberg, Germany.

Landmann, T. (2003), Estimating fire severity from remote sensing information for

experimental fires in the Kruger National Park, South African Journal of Science, 99,

357-360.

Larkin, E. D. (2002), Radiation balance of over forested and agricultural sites in northern

Thailand, University of Hawai’i Manoa.

Lenhart, T., K. Eckhardt, N. Fohrer, and H. G. Frede (2002), Comparison of two different

approaches of sensitivity analysis, Physics and Chemistry of the Earth, 27, 645-654.

Leopold, L. B. (1968), The hydrologic effects of urban land use: Hydrology for urban land

planning — A guidebook of the hydrologic effects of urban land use:U.S. Geological

Survey Circular 554, 18 pp.

Lillesand, T. M., R. W. Kiefer, and J. W. Chipman (2004), Remote sensing and image

interpretation, 5th ed., John Wiley and sons, USA.

Liu, Y. B., and F. D. Smedt (2004), WetSpa extension, a GIS based hydrologic model for

flood prediction and watershed management, 1-108 pp, Vrije Universiteit Brussel,

Belgium.

Lorup, J. K., J. C. Refsgaard, and D. Mazvimavi (1998), Assessing the effect of land use

on catchment run-off by combined use of statistical tests and hydrological modelling:

Case studies from Zimbabwe, Journal of Hydrology, 205, 147-163.

Loveland, T. R., B. C. Reed, J. F. Brown, D. O. Ohlen, Z. Zhu, L. Yang, and J. W.

Merchant (2000), Development of a global land cover characteristics database and

IGBP DISCover from 1km AVHRR data, International Journal of Remote Sensing, 21,

1303-1330.

Low, A. B., and A. G. Rebelo (1996), Vegetation of South Africa, Lesotho and Swaziland,

Department of Environmental Affairs and Tourism, Pretoria, South Africa.

Maidment, D. R. (1993), Handbook of hydrology, McGraw-Hall, New York.

Malawi Government (1986), The national atlas of Malawi, edited, Government Press,

Zomba.

Malawi Government (1998a), Malawi Vision 2020: National Long-term Development

perspective for Malawi.

Malawi Government (1998b), State of the Environment Report for Malawi, Likuni press,

Lilongwe, Malawi.

Malawi Government (1999), Water Resources Management and Policy Strategies,

Lilongwe, Malawi.

Malawi Government (2001), State of the Environment Report for Malawi, Likuni press,

Lilongwe, Malawi.

172

Malawi Government and Satellitbild (1993), Forest Resources Mapping and Biomass

Assessment for Malawi., edited by D. o. F. Ministry of Forestry and Natural Resources,

Satellitbild, a subsidiary of Swedish Space Corporation, Lilongwe, Malawi and Kiruna,

Sweden.

Mather, P. M. (1993), Computer processing of remotely sensed images, 352 pp., John

Wiley.

Meijerink, A. M. J., D. E. Brouwer, C. M. Mannaerts, and C. Valenzuela (1994),

Introduction to the use of Geographical Information Systems for practical hydrology,

ITC, Enschede.

Meyer, W. B. (1995), Past and present land use and land cover in the United States of

America, Consequencies, 1, 25-33.

Meyer, W. B., and B. L. Turner (1994), Changes in Land Use and Land Cover: A Global

Perspective, Cambridge University Press, New York.

Microimages (2005), TNTmips map image processing system, version 7.0, 206 South 13th

Street, 11th Floor-sharp Tower, Lincoln, NE 68508-20N, USA.

Mikkola, K. (1996), A remote sensing analysis of vegetation damage around smelters in

the Kola Peninsula, Russia, International Journal of Remote Sensing, 17, 3675–3690.

Miller, S. N., D. P. Guertin, and D. C. Goodrich (2003), Deriving stream channel

morphology using GIS-based watershed analysis. Chapter 5 in GIS for Water

Resources and Watershed Management, edited by J. G. Lyon, pp. 53-61, Taylor and

Francis, New York.

Miller, S. N., M. Hernandez, R. C. Miller, D. C. Goodrich, W. G. Kepner, D. L. Heggem,

M. L. Mehaffey, F. K. Devonald, and P. Miller (2002), Integrating landscape

assessment and hydrologic modelling in land cover change analysis, Journal of

American Water Resources Association, 38, 915-929.

Milner, A. K. (1988), Change detection analysis using Landsat imagery: a review of

methodology, paper presented at Proceedings of IGARSS ’88 Symposium, Paris: ESA

Publications Division, Edinburgh, Scotland, 13–16 September 1988.

Molicova, H., M. Grimaldi, M. Bonell, and P. Hubert (1997), Using topmodel towards

identifying and modeling the hydrological patterns within a headwater, humid, tropical

catchment, Hydrological Processes, 11, 1169-1196.

Monteith, J. L. (1965), Evaporation and Environment, Symposium Soc Exploration

Biology, 19, 205-234.

Morris, B. (1995), Woodland village: Reflections on the "animal estate" in rural Malawi,

Royal Anthropological Institute, 1.

Morris, M. D. (1991), Factorial sampling plans for preliminary computational experiments,

Technometrics, 33, 161-174.

Mouat, D. A., G. G. Mahin, and J. Lancaster (1993), Remote sensing techniques in the

analysis of change detection, Geocarto International, 2, 39-50.

Moussa, R., M. Voltz, and P. Andrieux (2002), Effect of the spatial organization of

agricultural management on the hydrological behaviour of a farmed catchment during

flood events, Hydrological Processes, 16, 393-412.

Moyo, S., P. O’keefe, and M. Sill (1993), The southern African environment: Profiles of

the SADC countries, Earthsan publications Ltd, London.

Muchoney, D. M., and B. N. Haack (1994), Change detection for monitoring forest

defoliation, Photogrammetric Engineering and Remote Sensing, 60, 1243–1251.

173

Mulungu, D., and S. Munishi (2007), Simiyu River catchment parameterization using

SWAT model, Physics and Chemistry of the Earth, 32, 1032-1039.

Mulwafu, W., C. Chipeta, G. Chavula, A. Ferguson, B. Nkhoma, and G. Chilima (2003),

Water demand management in Malawi: problems and prospects for its promotion,

Physics and Chemistry of the Earth 28, 787–796.

Munyati, C. (2000), Wetland change detection on the Kafue Flats, Zambia, by remote

sensing in Eastern Zambia, 1972-1989, International Journal of Remote Sensing, 21,

21, 301-322.

NASA Goddard Space Flight Centre (2002), Landsat 7 Science Data Users Handbook,

edited.

National Statistical Office (1991), Malawi Population and Housing Census 1987:

Summary of final results, edited by N. S. Office, Government printer, Zomba.

National Statistical Office (2000), Malawi Population and Housing Census 1998:

Summary of final results, edited by N. S. Office, Government printer, Zomba.

Ndomba, P. M. (2007), Modelling of Erosion Processes and Reservoir Sedimentation in

the Pangani River Basin, Upstream of Nyumba ya Mungu Reservoir, University of Dar

es Salaam, Dar es Salaam.

Ndomba, P. M., F. Mtalo, and A. Killingtveit (2005), The Suitability of SWAT Model in

Sediment Yield Modelling for Ungauged Catchments. A Case of Simiyu

Subcatchment, Tanzania, paper presented at Proceedings of the 3rd International

SWAT conference, EAWAG-Zurich, Switzerland, 11th –15th July 2005.

Nedkov, S., and M. Nikolova (2006), Modelling flood hazard in Yantra River basin, in

SWAT edited.

Neitsch, S. L., J. G. Arnold, J. R. Kiniry, and J. R. Williams (2005), SWAT theoretical

documentation version 2005

Newson, M. (1992), Land, Water and Development: River Basin Systems and Sustainable

Management, Routledge, London.

Nicks, A. D. (1974), Stochastic generation of the occurrence, pattern and location of

maximum amount of daily rainfall, paper presented at Statistical Hydrology, Aug.-Sept.

1971 Tucson, AZ. USDA Misc. Publ., U.S. Gov. Print. Office, Washington, DC.

Ominde, S. H., and C. Juma (1991), Introduction in A change in weather: African

perspectives on climate change, edited by S. H. Ominde and C. Juma, pp. 3-12, ACTS

Press, Nairobi.

Orr, B., B. Eiswerth, T. Finan, and L. Malembo (1998), Malawi: Public Lands Utilisation

Study (PLUS), Final Report University of Arizona Office of Arid Lands Studies and

the Forestry Research Institute of Malawi, Lilongwe.

Palamuleni, L. G., T. Landmann, and H. J. Annegarn (2006), Land cover mapping for the

Shire River catchment in Malawi using Landsat satellite data, paper presented at 6th

African Association of Remote Sensing of the Environment (AARSE), Cairo, Egypt,

30 Octocer - 2nd November.

Palamuleni, L. G., T. Landmann, and H. J. Annegarn (2007), Mapping rural savanna

woodlands, a comparison of maximum likelihood and fuzzy classifiers, paper presented

at International Geoscience and Remote Sensing Symposium (IGARSS’07), Official

IEEE Proceedings, Peer-reviewed, Barcelona, Spain, 23 -27 July.

174

Palamuleni, L. G., T. Landmann, and H. J. Annegarn (2008), An assessment of land cover

change using multi-temporal Landsat imagery for the Shire River catchment, Malawi,

paper presented at 7th African Association of Remote Sensing of the Environment

(AARSE) Conference, Accra, Ghana, 27 - 30 October.

Pappagallo, G., A. Lo Porto, and A. Leone (2003), Use of the SWAT Model for evaluation

of anthropic impacts on water resources quality and availability in the Celone Creek

Basin (Apulia - Italy), paper presented at 2nd International SWAT conference, TWRI

technical Report 266, Bari, Italy.

Perrin, C., C. Michel, and V. Andreassian (2001), Does a large number of parameters

enhance model performance? Comparative assessment of common catchment model

structures on 429 catchments, Journal of Hydrology, 242, 275-301.

Place, F., and K. Otsuka (2001), Population, land tenure and natural resources in Malawi,

Journal of Environmental Economics and Management, 41, 13-32.

Popov, E. G. (1979), Gidrometeoizdat (in Russian), Gidrologicheskie Prognozy

(Hydrological Forecasts).

Prakash, A., and R. P. Gupta (1998), Land use mapping and change detection in a coal

mining area: A case study in the Jharia coalfield, India, International Journal of

Remote Sensing, 19, 391-410.

Price, P. K., D. A. Pyke, and L. Mendes (1992), Shrub die back in a semi arid ecosystem:

the integration of remote sensing and geographic information systems for detecting

vegetation change, Photogrammetric Engineering and Remote Sensing, 58, 455–463.

Priestley, C. H. B., and R. J. Taylor (1972), On the assessment of surface heat flux and

evaporation using large-scale parameters, Monitoring Weather Review, 100, 81-92.

Refsgaard, J. C., and B. Storm (1996), Construction, calibration and validation of

hydrological models, in Distributed Hydrological Modelling, edited by M. B. Abbott

and J. C. Refsgaard, pp. 41-54, Kluwer Academic Publishers.

Riebsame, W. E., W. J. Parton, K. A. Galvin, I. C. Bulke, L. Bohren, R. Young, and E.

Knop (1994), Integated modelling of land use and cover change, Bioscience, 44, 350-

356.

Robinson, M. (1990), Impact of improved drainage on river flows, Institute of Hydrology

Oxon, UK.

Rockström, J., J. Barron, and P. Fox (2002), Rainwater management for increased

productivity among small-holder farmers in drought prone environments, Physics and

Chemistry of the Earth, 27, 949-959.

Rohr, P. C. (2003), A hydrological study concerning the southern slopes of Mt.

Kilimanjaro, Tanzania., Norway, Trondheim.

Roughgarden, J., S. Running, and P. Matson (1991), What does remote sensing do for

ecology, Ecology, 76, 1918-1922.

Rutchy, K., and L. Vilchek (1999), Air Photointerpretation and Satellite Imagery Analysis

Techniques for mapping Cattail Coverage in a Northern Everglades Impoundment,

Photogrammetric Engineering and Remote Sensing, 65.