MAPPING VEGETATION COVER AND DISTURBANCE BASED …

MAPPING VEGETATION COVER AND DISTURBANCE BASED ON UNMANNED

AIRCRAFT SYSTEMS (UASs) FOR A MILITARY INSTALLATION

by

Bibek Ban

B.S., Tribhuvan University, Nepal, 2014

A Thesis

Submitted in Partial Fulfillment of the Requirements for the

Master of Science Degree

Department of Geography and Environmental Resources

in the Graduate School

Southern Illinois University Carbondale

August 2021

THESIS APPROVAL



by

Bibek Ban

A Thesis Submitted in Partial

Fulfillment of the Requirements

for the Degree of

Master of Science

in the field of Geography and Environmental Resources

Approved by:

Dr. Guangxing Wang, Chair

Dr. Justin Schoof

Dr Ruopu Li

Graduate School

Southern Illinois University Carbondale

July 2, 2021

i

AN ABSTRACT OF THE THESIS OF

Bibek Ban, for the Master of Science degree in Geography and Environmental Resources

presented on June 28, 2021, at Southern Illinois University Carbondale.

TITLE: MAPPING VEGETATION COVER AND DISTURBANCE BASED ON

UNMANNED AIRCRAFT SYSTEMS (UASs) FOR A MILITARY INSTALLATION

MAJOR PROFESSOR: Dr. Guangxing Wang

Unmanned Aerial Vehicles (UAVs) can be used as a cost-effective alternative to map

military training induced vegetation disturbances and monitor their dynamics because of high

spatial resolution images provided at acceptable cost, great accuracy and flexibility of time to

collect the images. This study aims to develop a method to map vegetation cover change from

different military training activities using UAV imagery acquired at two different dates (before

and after the military training). Three flight boxes located in the Fort Riley (FR) were selected as

the study area. Vegetation cover data was collected in the field using Daubenmire frame for 1m,

5m and 10m sample plots. The UAV imagery was resampled to the spatial resolutions that match

the plot sizes. The UAV images were processed for their geometric and radiometric calibrations

and quality control. Eight vegetation indices (VIs) were calculated from UAV imagery and step-

wise regression was used to find the final model for each boxes. The results suggested that the

UAV images can be used to map vegetation cover and disturbance caused by military training

activities. Moreover, it was found that separately modelling the military training induced

vegetation disturbances for the training boxes led to greater accuracy than modelling the

vegetation disturbances by pooling the data together. The accuracy of modelling was also higher

ii

before the training than that after the training because the training activities led to higher spatial

variation of vegetation cover. In addition, the 5 m by 5 m spatial resolution images were more

capable in capturing spatial variation of the vegetation disturbances than those at 10 m by 10 m

spatial resolution, which implied that the 5 m by 5 m plots should be utilized for field data

collection. Finally, compared with the original UAV image bands, the VIs improved the

correlation with vegetation cover, and the Red Edge Modified Simple Ratio (REMSR),

Normalized Difference Vegetation Index (NDVI) and Green Normalized Difference Vegetation

Index (GNDVI) were more frequently selected in the final models than other vegetation indices

(VIs). Overall, this study enhanced the understanding of using UAV images to map vegetation

cover change from different military training activities

iii

ACKNOWLEDGMENTS

I am grateful to major Professor Dr. Guangxing Wang for believing me and providing

this opportunity to pursue my Master of Science in Geography and Environmental Resources. I

am thankful for his guidance, valuable suggestions, and encouragement during study time. I

would also like to express my sincere and heartily thanks to my committee members Dr. Justin

Schoof and Dr. Ruopu Li for their guidance and suggestions.

Finally, I am grateful to my Parents, my lovely sisters Biddhya and Binita for their love

and encouragement during study period.

iv

TABLE OF CONTENTS

CHAPTER PAGE

ABSTRACT……………………………………………………………………………………….i

ACKNOWLEDGEMENTS.……………………….………………………………......................ii

LIST OF TABLES……………………………………………………………………………….vii

LIST OF FIGURES………………………………………………………………………………x

CHAPTERS

CHAPTER 1 - INTRODUCTION………………………………………………………...1

1.1 Introduction……………………………………………………………………1

1.2 Research Statement…………………………………………………………...3

1.3 Objectives and Research Questions…………………………………………...4

CHAPTER 2 - LITERATURE REVIEW………………………………………………....5

2.1 Military activities induced disturbances at different levels…………………...5

2.2 Plot level studies………………………………………………………………6

2.3 Vehicle based studies………………………………………………………….7

2.4 Landscape level……………………………………………………………….7

2.5 Satellite images for military activity induced disturbance……………………8

2.6 Unmanned Aerial Vehicle for military activity induced disturbance……….10

CHAPTER 3 - MATERIAL AND METHODOLOGY…………………………………13

3.1 Study Area…………………………………………………………………...13

3.2 Datasets………………………………………………………………………16

3.2.1 Ground measurements……………………………………………..16

3.2.2 UAV images…………………………………………………….….19

v

3.3 UAV sensor…………………………………………………………………..20

3.4 Software Characteristics……………………………………………………..22

3.5 UAV image processing…………………………………………….………...22

3.6 Vegetation indices …………………………………………………………....27

3.7 Vegetation cover mapping……………………………………………….…..30

3.8 Model Selection and Accuracy assessment…………………………….……31

3.9 Cross Validation………………………………………………………..…….32

3.10 Vegetation ChangeMapping………………………………………………..33

CHAPTER 4 - RESULTS………………………………………………………………..34

4.1 UAV image processing………………………………………………………34

4.2 Vegetation indices …………………………………………………………...35

4.3 Field data…………………………………………………………………….40

4.4 Results from 5 m by 5 m field plots…………………………………………42

4.4.1 Vegetation indices…………………………………………………42

4.4.2 Correlation analysis………………………………………………..43

4.4.3. Vegetation cover models………………………………………….48

4.4.4 Mapping vegetation covers change………………………………..53

4.5 Results from 10 m by10 m field plots……………………………………….56

4.5.1 Vegetation indices…………………………………………………56

4.5.2 Correlation analysis………………………………………………..56

4.5.3 Vegetation cover models………………….………………………..61

4.5.3 Mapping vegetation covers change………………………………...65

CHAPTER 5 – DISCUSSION AND CONCLUSIONS………………………………….68

vi

5.1 Discussion …………………………………………………………...68

5.2 Conclusions…………………………………………………………..73

REFERENCES…………………………………………………………………………………..74

VITA…………………..………………………………………………………………………….80

vii

LIST OF TABLES

TABLE PAGE

Table 2.1: Performance comparison of UAVs with other platforms ............................................ 11

Table 3.1: Description of Micasense RedEdge-M sensor used to take multispectral images ...... 20

Table 3.2: Parameters of image bands from Micasense Rededge-M sensor ................................ 21

Table 4.1: The RMSE values of geometric images for the training boxes before and after the

military training. ............................................................................................................ 34

Table 4.2: Mean values of calculated vegetation indices from original images ........................... 36

Table 4.3: Descriptive statistics for each of military training boxes for 5 m by 5 m plot size ..... 41

Table 4.4: Descriptive statistics for each of military training boxes for 10 m by 10 m plot size . 42

Table 4.5: Mean values of calculated vegetation indices from 5 m resampled images ................ 42

Table 4.6: Correlation coefficient between vegetation cover and spectral variables obtained from

linear correlation analysis .............................................................................................. 44

Table 4.7: The matrix of correlation coefficients among the variables for box 1 before the


Table 4.8: The matrix of correlation coefficients among the variables for box 1 after the military

training. .......................................................................................................................... 45




training. .......................................................................................................................... 46



viii


training. .......................................................................................................................... 47

Table 4.13: Vegetation cover models based on stepwise linear regression and VIF and accuracy

assessment of predictions for each of three boxes using cross-validation ..................... 50

Table 4.14: Vegetation cover change statistics for each box derived from image differencing of

the maps before and after militrainig. ............................................................................ 54

Table 4.15: Mean values of calculated vegetation indices from 10 m resampled images ............ 56

Table 4.16: Correlation coefficient between vegetation cover and spectral variables obtained

from linear correlation analysis for 10 m resampled data.............................................. 57




training. .......................................................................................................................... 58




training. .......................................................................................................................... 59




training. .......................................................................................................................... 60


assessment of predictions for each of three boxes using cross-validation ..................... 62

ix

Table 4.24: Vegetation cover change statistics for each box derived from image differencing

from the before and after military training vegetation cover maps. .............................. 66

Table 5.1: Total time spent to process UAV images for six boxes .............................................. 69

x

LIST OF FIGURES

FIGURE PAGE

Figure 3.1: The map of Fort Riley Army Installation (left) shown by land cover classes with

three military training boxes outlined by black boundaries, and the zoomed in boxes

(right). ........................................................................................................................ 15

Figure 3.2: Spatial distribution of Ground Control Points used for geometric rectification of

UAV images .............................................................................................................. 16

Figure 3.3: Spatial distribution of Ground Truth Points used for accuracy assessment of

geometric rectification of UAV images .................................................................... 17

Figure 3.4: Spatial distribution and nested structure of Daubenmire frames or plots used to

collect vegetation cover data in the field. .................................................................. 18

Figure 3.5: Examples of five band UAV images acquired using Micasense Rededge-M camera

before pre-processing. ............................................................................................... 19

Figure 3.6: Micasense Rededge-M camera used for data collection. Source: Micasense (2017) 20

Figure 3.7: Spectral bands for the Micasense RedEdge-M camera and the spectral reflectance

curve of a green vegetation canopy (Micasense, 2017). ........................................... 21

Figure 3.8: Workflow for UAV image processing used in this study to produce orthomosaic

images and vegetation indices. .................................................................................. 23

Figure 3.9: Workflow to produce vegetation cover and disturbance map due to different military

training activities ....................................................................................................... 31

Figure 4.1: False color composite orthomosaic images of multispectral UAV images after

processing for three training boxes. .......................................................................... 35

Figure 4.2: The maps Normalized Difference Vegetation Index (NDVI) for each box before and

after military training, showing the training induced vegetation cover loss. ............ 36

xi

Figure 4.3: The maps of Green Normalized Difference Vegetation Index (GNDVI) for each box

before and after military training, showing the training induced vegetation cover

loss. ............................................................................................................................ 37

Figure 4.4: The maps of Red Edge Normalized Difference Vegetation Index (RENDVI) for each

box before and after military training, showing the training induced vegetation cover

loss. ............................................................................................................................ 37

Figure 4.5: The maps of Optimized Soil Adjusted Vegetation Index (OSAVI) for each box


loss. ............................................................................................................................ 38

Figure 4.6: The maps of Modified Soil Adjusted Vegetation Index (MSAVI2) for each box


loss. ............................................................................................................................ 38

Figure 4.7: The maps of Red Edge Modified Simple Ratio (REMSR) for each box before and

after military training, showing the training induced vegetation cover loss. ............ 39

Figure 4.8: The maps of Wide Dynamic Range Vegetation Index (WDRVI) for each box before

and after military training, showing the training induced vegetation cover loss. ..... 39

Figure 4.9: The maps of Atmospherically Resistant Vegetation Index (ARVI) for each box


loss. ............................................................................................................................ 40

Figure 4.10: The Residuals of predicted vegetation cover graphed against the fitted values for

box 1 before and after military training activities at the plot size of 5 m by 5 m. .... 51



xii



Figure 4.13: The Residuals of predicted vegetation cover graphed against the fitted values for the

pooled data from all boxes before and after military training activities at the plot size

of 5 m by 5 m ............................................................................................................ 52

Figure 4.14: Vegetation cover and change maps for box 1 from 5 m resolution resampled data. 54




box 1 before and after military training activities at the plot size of 10 m by 10 m. 63






overall pooled data before and after military training activities at the plot size of 10

m by 10 m. ................................................................................................................. 64

Figure 4.21: Vegetation cover and change maps for box 1 from 10 m resolution resampled data

................................................................................................................................... 66


................................................................................................................................... 67


................................................................................................................................... 67

1

1. CHAPTER 1

INTRODUCTION

1.1 Introduction

The US army manages land area of approximately 30 million acres at more than 5500

sites for different military activities like field maneuvers, combat vehicle operations, mortar and

artillery fire, and small firearms (Anderson et al. 1996; Howard et al. 2009; Rijal et al.2017;

Wang et al. 2007). ―Military off-road vehicular traffic is among human activities that are

responsible for the disturbance of ground and vegetation cover of landscapes which ultimately

increases the potential of rain-induced runoff and soil erosion‖ (Wang et al. 2007) and thus there

is increased concern for military land managers regarding land condition on military training

land (Wang et al. 2007). Military training activities are responsible for disturbing ground cover,

damaging plants and habitat, degrading water quality, land fragmentation, thereby degrading the

overall condition of land (Howard et al. 2009). The degradation of land increases if training

activities continue to occur and accumulate, whereas the land condition recovers if such activities

are halted, reduced, or rotated. Therefore, the state of training land condition is a function of

spatial intensity, temporal frequency of military activities, and natural recovery rates (Howard et

al. 2009). The mapping and assessment of disturbance both spatially and temporally of military

activities are crucial for the prediction of land conditions at military bases.

Remote sensing images have been used to map and monitor military activities. Remote

sensing images from finer spatial resolution to middle spatial resolution and coarser spatial

resolution are utilized (Hutchinson et al. 2015; Rijal et al. 2017; Wang et al. 2007, 2009, 2014).

The use of remote sensing images has its advantages and disadvantages. Fine spatial resolution

images are limited as the ground coverage is small and the cost is relatively high. Likewise,

2

coarse spatial resolution is limited because of its inability to provide detailed information due to

large pixel size. Thus, medium spatial resolution images have become popular because of their

free availability, relatively large and repetitive coverage, and long history.

Remote sensing-based methods provide the potential to map military training-induced

disturbances and monitor their dynamics. However, the availability of various remote sensing

images and mapping methods has made it difficult to select one kind of image and an accurate

method. For this reason, Construction Engineering Research Laboratory (CERL) proposed a

program entitled ―Demonstration and Validation of Unmanned Aerial Vehicles Monitoring

Methodologies Used for Natural Resource Assessment on Military Lands‖. The study used

unmanned Aerial Vehicles (UAVs) for mapping and monitoring military training induced

vegetation cover disturbance by comparison with other approaches for optimizing dynamic

monitoring of military land conditions.

The traditional way of field measurement to monitor the disturbances in land dynamics is

expensive, inefficient, and outdated. Range Training Land Assessment (RTLA) program which

was previously known as Integrated Training Area Management (ITAM) was one such approach

to measure the disturbances in military installations (Anderson et al. 1996; Rijal et al.2017).

Many studies are conducted using RTLA data to assess the land condition (Anderson et al.

2005b; Gertner et al. 2002, 2006; Howard et al. 2013; Senseman et al. 1996; Singer et al. 2012;

Wang et al. 2001, 2005, 2014). Wang et al. (2014) pointed to the need for remotely sensed data

and corresponding methods to generate the spatial distribution and patterns of land condition

because of the difficulties in the traditional approach and the high cost in field data collection.

3

1.2 Research Statement

Military vehicles used in training activities degrade the land condition. Resulting impacts

lead to increased soil erodibility from vegetation cover removal, loss of soil cohesion, soil

compaction, and rut creation. Maintaining healthy, continuous vegetation cover on Department

of Defense (DoD) lands is critical for realistic training and maintaining sustainable DoD lands.

DoD is facing challenges regarding cost-effective techniques of maximizing access to training

lands ensuring sustainable use for future operations. It is difficult to manage military land,

quantify disturbance due to military induced activities and to monitor different disturbance

mitigation efforts as they require a lot of resources and cost. The impacts are assessed by the

Range Training Land Assessment (RTLA) program which is designed at installation level.

The availability of various mapping methods made it difficult to choose a best method for

vegetation mapping. UAVs can be used as a cost-effective alternative to map the military

training-induced disturbances and monitor their dynamics because of their high spatial resolution

images, acceptable cost, great accuracy, and flexibility of acquisition time. Manned aircraft

systems equipped with different remote sensing sensors can provide necessary spatial resolution

data for accurate assessment of training disturbances, but it is not flexible to use the systems for

image collection. Moreover, UAVs can be relatively inexpensive compared to other remote

sensing systems.

UAV imaging as a remote sensing approach has become a popular environmental

monitoring tool. Especially the optical imagery, radar data, and LiDAR data are main foci of

researcher to reduce the uncertainties and spatial/temporal restrictions of ground level surveys.

The use of UAVs can be a solution to many natural and ground resources management problems.

There is also the potential of using UAV images for mapping vegetation cover and monitoring

vegetation disturbance caused by military training activities. Therefore, this study will focus on

4

using UAV data to map vegetation cover and assess disturbances caused by different military

training activities.

1.3 Objectives and Research Questions

The overall objective of this study is to develop a method to map vegetation cover change

from different military training activities using UAV imagery. The specific objectives are:

To develop vegetation cover model using UAV images to map and predict disturbances

caused by military training activities.

To evaluate the performance of different models at two image spatial resolutions

consistent with those of field observations.

To fulfill the objectives of this research we have to answer following research questions:

1. How to use UAV images for mapping vegetation cover loss due to military training

activities?

2. How to develop a vegetation cover model using different spectral variables?

3. How do the different resolutions of UAV imagery impact the performance of vegetation

cover models?

5

2. CHAPTER 2

LITERATURE REVIEW

2.1 Military activities induced disturbances at different levels

Many relevant studies were conducted on this subject in the United States and can be

divided into four groups: sample plot level, vehicle-based level, landscape level, and multiple

spatial and temporal scales (Anderson et al. 2005). Specific vehicles and their small-scale impact

regime are assessed in plot level studies (Abele et al. 1984; Althoff et al. 2005; Leis et al. 2005;

Wilson, 1988). The impact caused by vehicle characteristics and the use of various kind of

vehicles on land condition have been assessed on studies that focus on vehicle-based research

(Althoff et al. 2005; Jones et al. 2005). The spatial pattern and the distribution of impact caused

by off-road vehicle use have been studied on landscape level studies (Anderson et al. 2005;

Gertner et al. 2002; Wang et al. 2007). To study the impact of off-road vehicle use on land

condition are scale relevant, thus there is necessary that such studies should be focused on

multiple temporal and spatial level (Howard et al. 2013). To assess the impact of spatial and

temporal dynamics of vehicle use in military installations, sample plot data along with ancillary

data are combined for spatial interpolation (Wang et al. 2007).

Studies that quantify the impact of military training on land can be summarized into

groups that focus on soil properties, soil erosion, vegetation and soil disturbance, and

biodiversity (Althoff et al. 2005, 2007; Anderson et al. 2005; Barron et al. 2012; Delaney et al.

2011). Especially, many reports focus on the impacts of military training disturbances on

degradation of installation land conditions through predicting the spatial distribution of soil

erosion and monitoring its dynamics based on a revised universal soil loss equation (RUSLE)

6

and its integration with remotely sensed data (Anderson et al. 2005; Gertner et al. 2002; Johnson

et al. 2011; Rijal et al. 2017; Wang et al. 2007, 2009, 2014).

2.2 Plot level studies

Plot studies are focused at discrete spatial scales to assess the impact of military vehicles.

These studies focus on the impact assessment due to specific vehicle under controlled condition

and to study the ability of site to sustain and recover from vehicular damage (Anderson et al.

2005). The studies are typically done on most damaging vehicles, most common vehicle used in

military training, or new weapon systems being used at an installment. Dependent variables

selected in these studies are based on soil and vegetation impact in the tracked area or those areas

that are adjacent to the track (Anderson et al. 2005).

Althoff et al. (2005) assessed the impact of M1A2 military vehicle on the vegetation and

soil. The study used the soil moisture, soil type and tracking regime as controlled variables. This

study design has some advantages like the impact regimes and conditions of sites are well known

and different statistical methods can be used to assess the cause-and-effect relationship. The

dependent variables can be assessed simultaneously as well (Althoff et al. 2005; Anderson et al.

2005). Similarly in the study from Palazzo et al. (2005), author used the existing controlled and

repeated plot study to monitor the impact due to vehicle use and to assess the relationship among

vehicular damage and recovery of different grass. To quantify vehicular impact on multiple

sites, Hansen et al. (2005) used the observational design. They used the single track of known

vehicle to assess the impact. The study from Anderson et al (2005) used the data from existing

source from inventory and monitoring programs to monitor the impact due to vehicle activities.

7

2.3 Vehicle based studies

Only one or few vehicle types are include in plot level studies, and a single vehicle type

is not used in military training. The vehicle-based studies are focused on multiple off road

vehicular use in military activities. Military vehicles are designed so as to achieve military

mission with limited thought about environmental impacts (Anderson et al. 2005). These designs

impact the performance of vehicle off-road and the sites. Vehicle based studies focus on the

impact of off-road vehicle based on their properties and design characteristics.

Jones et al. (2005) studied the soil rutting as impact of military vehicle. The researcher

evaluated the degradation of terrain characteristics with design class of military vehicle.

Similarly, Raper (2005) assessed the impact of vehicle types and design configuration on

different soil properties which includes compaction and rut formation and concluded that it may

not be possible to eliminate soil compaction from vehicle traffic but can be controlled and

reduced through proper management. Since these studies account for characteristics of vehicle

and site interaction model, they are very helpful for assessing the impact from wide range of

vehicle types.

2.4 Landscape level

―To fully assess impacts at the landscape level, one must understand when and where

vehicles can potentially be used, where and how vehicles are actually used, and how vehicles

may be used in the future‖ (Anderson et al. 2005). It is good to know about when, where and

how vehicles are used, since it is crucial in assessment of impact as different sites have different

susceptibility to damage. Shoop et al. (2005) assessed the potential impact of new vehicles on

8

training land from a vehicle mobility analysis. To map maneuver area, the study used site

characteristics and properties of vehicle.

Different approaches are used in different studies to quantify the spatial distribution of

site. Singer et al. (2012) used global positioning system (GPS) to mark actual pattern of vehicle

use. This data along with impact data was used to estimate unit impacts. Based on actual on-site

vehicle use, Ayers et al. (2005) monitored the newly formed roads that were unmapped. This

method has a great advantage as this approach indicates the spatial patterns of actual use rather

than potential use.

2.5 Satellite images for military activity induced disturbance

In recent years much research is done to assess the impact of military training activities

on land condition, using remote sensing data. The use of military tracked and wheeled vehicle in

installation of military bases has made these studies mainly focused on disturbances on ground

and vegetation canopy cover (Rijal et al. 2017). The dynamics of installation land conditions

are driven by the changes in ground and vegetation canopy cover (Rijal et al. 2017). The use of

Advance Very High-Resolution Radiometer (AVHRR) data to monitor the change in vegetation

cover caused by use of military vehicle in Fort Bliss, New Mexico, was done by Minor et al.

(1999). The authors used the time series of Normalized Difference Vegetation Index (NDVI)

from AVHRR. Similarly in the study to assess the impact on environment due to use of vehicle

traffic at different spatial level in Fort Benning, Georgia, Dale et al. (2005) used Landsat data

and regional simulation model. Using Landsat and RTLA data at Fort Bliss, Texas; and Yakima

Training Center, Washington, Tweddale (2001) and Senseman et al. (1996) respectively

monitored the changes in the vegetation cover due to training activities and vehicle traffic.

9

Anderson et al. (2005) used the Landsat and RTLA data to monitor and predict the vegetation

cover degradation by military training activities and concluded that vegetation cove factor which

is define as the function of canopy cover, ground cover and minimum rain drop height in revised

universal soil loss equation (RUSLE) was correlated with soil erosion.

Remote sensing data have been widely used to develop methodologies for land degradation

modeling and mapping. Wang et al. (2009) used Landsat data along with spatial conditional co-

simulation approach to monitor and estimates military activities on land condition. Singer et al.

(2012) used global information system (GIS) with spatial multi-criteria decision analysis to

model global and local indicators and monitor the environmental dynamics at Fort Riley (FR).

The researchers used integrated approach and used variables like soil erosion, water quality,

landscape fragmentation and noise. The same military installation site was chosen by Wang et al.

(2014) to study the cumulative impact raised from different intensity of military training

activities, burning, and haying.

There are global and local modeling methods used to model the relationships of field

measurements of military training induced disturbances with remote sensing image derived

variables. It is further used to generate spatially explicit estimates of military training induced

disturbance (Howard et al. 2013; Rijal et al. 2017; Wang et al. 2007, 2009, 2014). The global

modeling methods lack the ability to characterize local spatial variability and their assumption of

normal data distribution lead to underestimations and overestimations for small and large values

respectively. The local modeling methods like geographically weighted regression (GWR) uses

local variability and spatial autocorrelation of variables and can offer a great improvement of

estimates but their use is more complicated (Wang et al. 2007, 2009, 2014).

10

2.6 Unmanned Aerial Vehicle for military activity induced disturbance

―Let them fly and they will create a new market‖ (Colomina et al. 2014). UAS has

different names like UAV, ―aerial robot‖ and ―drone‖. The term ―UAS‖ was adopted by

Department of Defense (DOD), United States. A UAV is a component of an UAS consisting of

additionally a ground-based controller and a system of communications with the UAV

(Colomina et al. 2014). UASs could be used for precision agriculture, vegetation mapping and

monitoring, etc. That is, UAS technologies can basically consist of three different components

UAV, ground control station, and communication data link (Colomina et al. 2014). There are

also other components like autopilots, navigation sensors, and wireless system. Everaerts et al.

(2004) used stratospheric UAS to gather high resolution aerial photos for near real-time large

scale mapping. The study shows the comparative analysis of different systems like airborne

platforms, satellite platforms, low-altitude, and high altitude UAS. Ground control stations are

very important in UAS as they monitor and command the aircraft. The outcomes of the payload

sensor are observed with ground control points.

Processing of data and measurements is critical in any remote sensing systems (Colomina

et al. 2014). The acquisition of data from UASs might look simple but the data does not translate

into image processing software. More complicated methods are required for data processing like

radiometric and geometric rectification of UAV images and creation of orthomosaic images. Qin

et al (2013) described the challenges of processing, orientating, and calibrating UAS datasets.

UASs are characterized by low cost and light navigation system. Orientation system

which depends upon requirements of orientation includes a mapping grade or geodetic-grade

sensors (Colomina et al. 2014). The navigation system in mapping grade case cannot be used as

aerial control and thus direct sensor orientation or integrated sensor orientation cannot be used

(Colomina et al. 2014). There are three camera calibration strategies. The first is recommended

11

by Remondino et al. (2011) where the camera is calibrated shortly before or after the flight

mission. The second is to apply self-calibration like as in drone. Third is the combination of both

as described by Cramer (2013). Whatever may be the calibration strategy, the sensor orientation

is based on auto detection and image measurement of points.

The mapping products of UASs include digital surface model (DSM) and orthophotos.

After the image orientation, calibration, developing DSM and orthophotos is very common. But

due to limited research, there is still a challenge for obtaining high quality orthophotos after pre-

processing of UAS images (Colomina, 2014). Rosnell et al. (2012) compared the results from

post processing of images obtained from two different cameras set up with two software systems.

The authors realized the need for new mapping methods as the conventional processing approach

did not fully capture to match low altitude and large three-dimensional models. Haala et al.

(2013) did a similar study from two cameras using SURE software.

Using UAV images from UASs has many advantages when compared to traditional

remote sensing data. The major advantage is the high spatial resolutions of the images with low

cost and high temporal resolutions due to its flexibility on operation. Table 2.1 shows the

performance comparison of UASs with other widely used remote sensing systems (Liu et al.

2016).

Table 2.1: Performance comparison of UAVs with other platforms

Attributes UAVs Manned

Aircrafts Satellites

Endurance Low Medium High

Payload capacity Low High Medium

Cost Medium Low Medium

Images without cloud

coverage High Medium Low

Maneuverability High Medium Low

Deployability High Medium Low

Autonomy requirement Medium Low Medium

12

Several studies have focused on the potential of use of UAVs for vegetation mapping.

Most of the studies for vegetation mapping are focused on crop fields where the scan areas are

limited and well known (Salami et al. 2014). Salami et al. (2014) mentioned that there have been

reports related to wildlands such as forested lands. In recent years, use of UAVs for vegetation

mapping has become popular. For example, Quanlong et al (2015) used UAV images for urban

vegetation mapping. The studies by Bertacchi et al. (2019), Meng et al. (2017), Arnold et al.

(2013), and Jafri et al. (2007) focused on the use of UAV images and their derived vegetation

indices (VIs) to map vegetation cover.

13

3. CHAPTER 3

MATERIAL AND METHODOLOGY

3.1 Study Area

Fort Riley (FR) Military Installation was selected for military training induced vegetation

disturbance mapping using UAV images (Fig. 3.1). FR was established in 1853 as a military

deploying station (Singer et al. 2012) and is located in Geary, Riley, and Clay counties of

northeastern Kansas, USA. Currently, the FR installation occupies an area of 41,154 ha, of which

69.2% (28,725 ha) is used for training and the remaining 30.8 percent (12,429 ha) is occupied by

the impact region, the Douthit gunnery complex, and the cantonment areas, where house offices,

base buildings, and vehicle repair facilities are located .The training land is divided into four fire

areas namely gunnery ranges, small arms and impact area, and training areas for maneuvers. A

total of 103 training areas are located in an installation where maneuvers and combat operations

take place, and the impact areas like on the central east part of the army installation.

Fort Riley is described by a tallgrass prairie and dominated by grassland (32,200 ha),

shrubland (6000 ha), and forests (1600 ha) (Althoff et al. 2005; Bailey 1976). Prevailing

grassland species incorporate big bluestem (Andropogon gerardii), switchgrass (Panicum

virgatum), Indiangrass (Sorghastrum nutans), composite dropseed (Sporobolus compositus),

and little bluestem (Schizachyrium scorparium) (Althoff et al. 2005, Koch et al. 2012).

Prevailing shrubland species incorporate Buckbrush or Coralberry (Symphoricarpos

orbiculatas), smooth sumac (Rhus glabra), leadplant (Amorpha canescans), harsh leaved

dogwood (Cornus drummondii), and white sagebrush (Artemisia ludoviciana) which are found

along the edges of the forest and disengaged patches in meadows (Althoff et al. 2005, Koch et al.

2012). Prevailing forest species incorporate chinquapin oak (Quercus muhlenbergii), pod oak

14

(Quercus macrocarpa), American elm (Ulmus History of the U.S), hackberry (Celtis

occidentalis), and dark pecan (Jugland nigra) which are found along riparian marshes (Althoff et

al. 2005, Koch et al. 2012). The eastern bit of the establishment is overwhelmed by highly

productive, warm-season grasses blended in with yearly and perennial forbs and contain hidden

substrate of shale and limestone substrate, though, the western part has less relief and more

profound soils geography with plant species going through succession back to the native prairie

from cultivation (Quist et al. 2003). The vast majority of the 15 streams situated in FR are

perennial that begins and contained inside the limit of the Installation (Quist et al. 2003).

Different military training activities take place in FR, that includes combat vehicle

operations, live-fire exercises, field maneuvers, artillery, and tank firing exercises, mortar, small

arms fire, etc. (US Army 1994). In addition to these activities, burning (natural and prescribed),

and haying are common in the study area that has a significant impact on land and vegetation

condition. In this study, military training induced vegetation disturbance mapping was examined

in three military training boxes shown in Fig. 3.1 located on central region of FR. The map was

developed using the data derived from National Land Cover Database (NLCD), 2016. It was

noticed that all training boxes are herbaceous and shrubs, and only of six land classes as defined

by NLCD are present in the study boxes.

15

Figure 3.1: The map of Fort Riley Army Installation (left) shown by land cover classes with

three military training boxes outlined by black boundaries, and the zoomed in boxes (right).

16

3.2 Datasets

The datasets used in this study were divided into ground measurements and images.

3.2.1 Ground measurements

3.2.1.1 Ground Control Points

A total of 93 ground control points (GCPs) were distributed over the study area. Out of

collected ground control points, 35 of them were outside the boxes, which can be used for

satellite image geometric rectification. Ground control points are used for geometric calibration

of images in this study.

Figure 3.2: Spatial distribution of Ground Control Points used for geometric rectification of

UAV images

17

3.2.1.2 Ground Truth Points (GTPs)

A total of 42 ground truth points (GTPs) were collected on structures or features that will

not change over time like the inlet and outlet of a culvert, a flagpole, etc. Ground truth points are

used for geometric calibration of images. GTPs were not used in this research but will be used in

future for accuracy assessment of geometric calibration.

3.2.1.3 Daubenmire Frame plots

A total of 36 10 m × 10m Daubenmire frames or plots were designed to measure

vegetation cover disturbance due to military training activities (Fig. 3.4). Each of the 10 m × 10

Figure 3.3: Spatial distribution of Ground Truth Points used for accuracy assessment of geometric

rectification of UAV images

18

m plots was divided into four 5 m × 5 m sub-samples (quadrats) in which five 1 m × 1m sub-

plots were located. Vegetation cover data were collected from a total of three hundreds twenty

four 1 m × 1 m sub-plots in July (28th, 29th, and 30th) and August (18th, 19th, and 24th),

respectively. Every Training Area or box got 10-20 plots and 1/5 of the plots were designed for

undisturbed, 1/5 bare ground, and 3/5 in low, medium, and high disturbance within each of 10 m

x 10 m plots. We have the vegetation cover datasets for 1 m, 5 m, and 10 m plot sizes. Every 5 m

and 10 m plot the vegetation cover data are the averages of 1 m sub-plot data. In this study, we

only utilized the datasets for 5 m and 10 m plot sizes for vegetation cover mapping.

Figure 3.4: Spatial distribution and nested structure of Daubenmire frames or plots used

to collect vegetation cover data in the field.

19

3.2.2 UAV images

The UAV images for each box were taken at two different times. The first UAV flight

mission took place on July 27th

before military training activities and the second flight mission

took place on August 17th

for boxes 1 and 2 and 18th

for box 3 after training activities to map

vegetation cover before and after disturbances. Three types of UAV images were acquired

during flight missions. The multispectral images were taken using Micasense Rededge-M camera

and RGB image acquired using Sony ILCE-7RM3 and in addition, Lidar data were collected

using Geodetics GEO-MMS package w/Velodyne VLP-16 LiDAR. In this study, only

multispectral images were used for vegetation cover mapping (Fig. 3.5).

Figure 3.5: Examples of five band UAV images acquired using Micasense Rededge-M camera

before pre-processing.

20

3.3 UAV sensor

The Micasense RedEdge-M was used to collect UAV data. The specification of sensor is

provided in Table 3.1 and Fig. 3.6. HFOV is defined as horizontal field of view.

Table 3.1: Description of Micasense RedEdge-M sensor used to take multispectral images

Camera Type Multispectral

Name Micasense RedEdge-M

Focal length (mm) 5.5

HFOV(o) 47.2

Spectral Bands 5 (Narrowband: Blue, Green, Red,

Red Edge, Near IR)

Size dimensions (mm) 9.4 cm x 6.3 cm x 4.6 cm (3.7‖ x 2.5‖

x 1.8‖)

Image Format RAW/TIFF

Weight (kg) 0.17 (including DLS)

Figure 3.6: Micasense Rededge-M camera used for data collection. Source: Micasense (2017)

The multispectral camera used is a five-band Micasense Rededge-M. It can acquire Red

(B1), Green (B2), Blue (B3), Near Infrared (B4) and Red Edge (B5) bands simultaneously

(Table 3.2 and Fig. 3.7). The ground Sample distance is 8.2 cm/pixel per band at 120 m above

21

ground level (AGL). The camera is powered by UAV batteries and provides geo-tagged images

(Micasense 2017).

Table 3.2: Parameters of image bands from Micasense Rededge-M sensor

Band number Band name Center

Wavelength

(nm)

Bandwidth

FWHM(nm)

Range (nm)

1 Blue (B1) 475 20 465-485

2 Green (B2) 560 20 550-570

3 Red (B3) 668 10 663-673

4 NIR (B4) 840 40 820-860

5 Red Edge (B5) 717 10 712-722

The Downwelling Light Sensor (DLS) is a 5-band light sensor connected to the

Micasense Rededge camera and its global positioning systems (GPS). It is mounted on the top of

the UAV, facing the sky. It contains a light diffuser that provides an irradiance reference value

for each spectral band in W/m2 /nm. It also has a 16 magnetometer, providing the measured

orientation of the DLS in degrees (irradiance yaw, pitch, and roll). The information is available

in the metadata of the images captured by the camera (Micasense 2017). Spectral response curve

for Micasense RedEdge-M camera is given in figure 3.7. The colors of the lines correspond to

the camera bands. The brown line is a standard reflectance profile of a green vegetation canopy.

Figure 3.7: Spectral bands for the Micasense RedEdge-M camera and the spectral reflectance

curve of a green vegetation canopy (Micasense, 2017).

22

3.4 Software Characteristics

There are different software packages available that can perform Structure-From-Motion

(SFM). The most commonly used software to process Micasense RedEdge imagery is Pixel 4D,

Agisoft Metashape Pro and Atlas. For multispectral image processing and vegetation indices

calculation in this study, Agisoft Metashape Pro was used.

Agisoft Metashape is a stand-alone software product developed by Russian company

Agisoft LLC located in St. Petersburg. It performs ―photogrammetric processing of digital

images and generates 3D spatial data to be used in GIS application, cultural heritage

documentation, and visual effects production as well for indirect measurement of objects of

various scales‖ (Agisoft LLC 2020).

3.5 UAV image processing

The methodological framework of UAV image processing adopted in this research to

produce orthomosaic images and calculate VIs is shown in Fig. 3.8. Flight planning and image

acquisition are necessary before image processing. The areas of flight were delineated and flight

planning was done to cover the areas. The UAV images are taken during flight for each training

box. The obtained images were preprocessed and the pre-processing included determination,

radiometric calibration, geometric correction, orthomosaic image creation, etc.

23

Figure 3.8: Workflow for UAV image processing used in this study to produce orthomosaic

images and vegetation indices.

3.5.1 Determination of image Quality

Poor quality images e.g. vague images can influence the alignment of UAV images,

which impacts the quality of the final products. Thus, it is better to estimate the quality of the

images before further processing. The image quality parameter is calculated based on the

sharpness level of the most focused part of an image (Agisoft manual 2019). Image sharpness is

defined as the clarity with which details are rendered in an image. Agisoft metashape (2019)

recommended the image quality values less than 0.5 should be disabled and exclude from further

processing. After experiment, in this study, all images greater than 0.65 units were selected.

24

3.5.2 Radiometric calibration

Radiometric calibration is an important process in multispectral imagery. Radiometric

calibration considers directional estimations (like the position of the sensor and the sun),

irradiance estimations (utilizing tools like light sensors or reflectance panels), as well the gain

and exposure information from the camera. Considering these variables the transformation of

raw digital numbers, to sensor reflectance or irradiance, and afterward to surface reflectance

values (Micasense, 2021).

The main part of the radiometric workflow is to collect irradiance data for each flight

which is done by taking images of a calibrated reflectance panel pre and post-flight. The

calibrated images are loaded separately while uploading images in software, which are later used

for radiometric calibration of all images. With a unit of W/m2/sr/nm, RedEdge radiometric

calibration converts the raw pixel values of an image into spectral radiance. It compensates for

sensor black-level, the sensitivity of the sensor, sensor gain and exposure settings, and lens

vignette effects (Micasense 2021). The radiometric calibration is done using software, where the

software automatically adds the calibrated images to a separate folder. We need to load the panel

data and perform the radiometric calibration using panel data and sun sensor. The formula for

computing spectral radiance L from pixel value p, used in RedEdge camera radiometric model

adopted from Micasense is:

1

2 3

( , )* * BL

e e

a p pL V x y

g t a y a t y

(3.1)

Where,

p is the normalized raw pixel value

PBL is the normalized black level value (can be found in metadata tags)

25

a1 a2, a3 are the radiometric calibration coefficients

V(x, y) is the vignette polynomial function for pixel location (x, y).

te is the image exposure time

g is the sensor gain setting (can be found in metadata tags)

x, y are the pixel column and row number, respectively

L is the spectral radiance in W/m2/sr/nm

3.5.3 Feature extraction across the photos (Alignment)

Image alignment is simply known as finding the spatial mapping where an element in one

image is matched with elements in the second image. The points that are stable under viewpoint

and lightning variations are detected and descriptors are generated for each point based on the

local neighborhood. These descriptors are used to detect any correspondence across images and

align them. Image alignment is done using Agisoft which is similar to SIFT (Scale-invariant

feature transform) in other software, although it uses a different algorithm for a higher quality

image (Tagle et al. 2017; Agisoft LLC 2016).

3.5.4 Geometric Calibration

To reduce the difference between the locations of objects in images and their locations on

the ground surface, geometric calibration is performed. A most common way to improve

geometric accuracy is by geo-referencing images with ground control points (GCPs). The GCPs

are loaded in the images, and each image that contains ground control points is selected. If the

26

GCPs are not in their positions, we can place the points to their locations by dragging the points.

Here to ensure maximum geometric accuracy, it is always important to optimize cameras after

adding or editing GCPs (Metashape 2021).

3.5.5 Optimization of Camera Alignment

Camera optimization is an adjustment procedure on aligned images simultaneously

refining interior and exterior camera orientation parameters and triangulated tie point

coordinates. The camera optimization is done using Agisoft where adjustment is performed using

all measurements and corresponding accuracies (Metashape 2021).

3.5.6 Tie point error reduction

After the images are optimized, it is now recommended to clean the sparse point cloud

from re-projection errors which represent the residuals of image coordinates determined by

matching algorithms (Mayer et al. 2018). Three different criteria namely reconstruction

uncertainty, projection accuracy, and re-projection error are used to remove unnecessary tie

points. The cleaning process is conducted by utilizing the Gradual Selection tool.

3.5.7 Solving for camera intrinsic and extrinsic orientation parameters (Sparse point cloud

creation)

The set of matching image data points in a three-dimensional coordinate system is known

as a sparse point cloud. The generation of sparse point cloud uses an algorithm to find camera

locations and refines them using bundle adjustment to generate sparse cloud (Tagle et al. 2017;

Agisoft LLC 2016).

27

3.5.8 Dense surface reconstruction (Dense point cloud generation)

Dense point cloud generation is another important step in UAV image processing.

Different processing methods are available while the fast method utilizes a multi-view approach

to generate a 3-dimensional (3D) dense point cloud (Tagle et al. 2017; Agisoft LLC 2016).

3.5.9 Digital Elevation Model (DEM)

DEM is generated using a dense point cloud that is generated earlier. DEM can include

both the terrain and objects above ground like trees or buildings known as a digital surface model

or only using the landscape of territory known as digital terrain model (Metashape 2021).

3.5.10 Orthomosaic Creation

The creation of orthomosaic images is important as the orthomosaic images are used as a

base layer for further post-processing of images and vectorization. It is the process where

photogrammetrically rectified images are stitched together from a collection of images. Images

are projected according to their external orientation/internal orientation data on a surface of the

user’s choice like DEM or mesh to create orthomosaic (Metashape, 2021). Here DEM is used to

create the orthomosaic of images. The use of software made it simple to create orthomosaic for a

very large dataset.

3.6 Vegetation indices

Once the orthomosaic images have been created, the next step in UAV image processing

is to calculate VIs. For multispectral imagery, the calculated VIs can reveal the dynamics of

vegetation cover. All the VIs have been calculated using Metashape and extracted. A total of

28

eight widely used vegetation indices were calculated in this study. The selection of vegetation

indices is based on the available bands in multispectral images and their ability of capturing

vegetation canopy structures and their dynamics.

3.6.1 Normalized Difference Vegetation Index (NDVI)

NDVI is used as a measure of healthy green vegetation. It is more robust as it uses

highest absorption and reflectance region of chlorophyll and it uses the combination of

normalized difference formulation (Rouse et al. 1974).

( Re )

( Re )

NIR dNDVI

NIR d

(3.2)

3.6.2 Green Normalized Difference Vegetation Index (GNDVI)

The index uses green spectrum as opposed to red spectrum in NDVI. This is more

sensitive to chlorophyll than NDVI (Gitelson et al. 1998)

( )

( )

NIR GreenGNDVI

NIR Green

(3.3)

3.6.3 Modified Soil Adjusted Vegetation Index (MSAVI2)

This vegetation index is the improved version as it reduces soil noise and increases the

dynamic range of vegetation based on inductive method. It does not use a constant slope (L)

value to indicate healthy green vegetation (Qi et al. 1994)

22* 1 (2* 1) 8( Re )2

2

NIR NIR NIR dMSAVI

(3.4)

29

3.6.4 Red Edge Normalized Difference Vegetation Index (RENDVI)

This is different from NDVI as it uses Red Edge band as opposed to main absorption and

reflectance peaks. It is focused on sensitivity of red edge to small change in foliage content

(Sims et al. 2002).

Re

Re

NIR dedgeRENDVI

NIR dedge

(3.5)

3.6.5 Optimized Soil Adjusted Vegetation Index (OSAVI)

This index provides greater soil variation for low cover, while increased sensitivity for

vegetation cover greater than 50% (Rondeaux et al. 1996). This uses standard value of 0.16 for

canopy adjustment factor.

( Re )

( Re 0.16)

NIR dOSAVI

NIR d

(3.6)

3.6.6 Red Edge Modified Simple Ratio Index (REMSR)

Simple ratio vegetation index is more sensitive to chlorophyll change as compared to

NDVI, and more influenced by environmental factor. Wu et al. (2008) used the modified simple

ratio as opposed to simple ration to avoid disturbances.

( / Re ) 1

( Re ) 1

NIR dedgeREMSR

NIR dedge

(3.7)

3.6.7 Wide Dynamic Range Vegetation Index (WDRVI)

This index uses weighting coefficient in order to reduce the disparity in NDVI caused by

NIR and Red signals. This is effective when vegetation density is moderate to high when NDVI

30

is greater than 0.6 (Gitelson A. 2004). The weighting coefficient of 0.2 was used as suggested by

Henebry et al. 2004.

( * Re )

( * Re )

NIR dWDRVI

NIR d

(3.8)

3.6.8 Atmospherically Resistant Vegetation Index (ARVI)

This index was developed as enhancement of NDVI that is relatively resistant to

atmospheric factors and originally developed to be used for with MODIS. It uses blue reflectance

to correct red reflectance for atmospheric scattering (Kaufman et al. 1992)

( (2*Re ) )

( (2*Re ) )

NIR d BlueARVI

NIR d Blue

(3.9)

3.7 Vegetation cover mapping

After UAV image processing, the final products are the orthomosaic of images and

vegetation indices. The next important is the spatial mapping of vegetation cover using UAV

images and field data. For this, resample of UAV images for 5m and 10m are done. Degrade

function from raster generalization tool in Erdas Imagine was used to resample data. The big

pixels are made using the mean of values from all the original small pixels which is also known

as the focal mean operator (Hexagon, 2021). The scaling factor was selected based on the

original pixel size to resample the images to match field data. The vegetation cover mapping

workflow is presented in Fig. 3.9.

31

Figure 3.9: Workflow to produce vegetation cover and disturbance map due to different military

training activities

3.8 Model Selection and Accuracy assessment

A total of thirteen spectral variables are used to find the final model: eight different

vegetation indices calculated and five original bands in the images. Vegetation cover data from

the Daubenmire frame was used as a dependent variable. After the selection of variables, the first

step is to look at the correlation among selected variables. Linear correlation analysis is used to

measure the strength and relationship among variables. For this, a correlation matrix is calculated

for each box before and after disturbances, to find the relationship among variables as highly

32

correlated variables could lead to multicollinearity, and the regression coefficient can be

unreliable.

The linear stepwise regression was used to select the final model for each box, where

variables are added or removed at each step accordingly to the explanatory value of the model

(Voorde et al. 2008). It is a step-by-step iterative process that selects the independent variables

that are used in the final model. Root mean square error (RMSE) and relative RMSE (RRMSE)

for each model were calculated. Variance inflation factor (VIF) was utilized to detect any

multicollinearity in regression analysis. Any variables that has high VIF (>10) were removed and

the final models were selected based on regression analysis and VIF.

RMSE is simply the standard deviation of residuals. Residuals measures how far the data

points are from regression line, where RMSE measures the spread of these residuals. It is

calculated using following formula (Barnston 1992).

[ ]

(3.10)

Where,

zp= predicted values, zo= observed values, and N= sample size.

RRMSE is calculated by dividing RMSE by the sample mean.

*100RMSERRMSE

mean (3.11)

3.9 Cross Validation

Cross-validation is done to split the N observations of our dataset into K subsets where

each subset is used for validation and the remaining set is used for estimating parameters. To

assess the performance of the models, we looked at the RMSE and RRMSE of vegetation cover

33

predictions through the leave-one-out cross-validation (LOOCV). The major advantage of using

LOOCV is the robustness, where the training set is as similar to as possible data from the real

world, the same observation is not used for training and evaluating the prediction while the major

disadvantage lies in the computational burden (Vehtari et al. 2012).

3.10 Vegetation Change Mapping

There are two main ways to model vegetation change: Post estimation and pre-

estimation. The post-estimation emphasizes separately modeling spectral variables such as NDVI

from images acquired at two times corresponding to before and after the vegetation disturbance,

respectively. This method then generates vegetation cover maps separately and derives the

change of vegetation using the difference from two maps. The pre-estimation focuses first on

deriving the changes of spectral variables such as NDVI from images acquired at two times

corresponding to before and after the vegetation disturbance, respectively, and then on modeling

changes of vegetation cover with the differences of NDVI at two times. In this study, the post-

estimation technique for vegetation disturbance mapping is used to estimate vegetation cover

change. The image difference function will be used to difference maps from two time frames for

same area to find vegetation cover change caused by military training activities.

34

4. CHAPTER 4

RESULTS

4.1 UAV image processing

All images with an image quality of less than 0.65 were removed. The radiometric panel

was loaded and radiometric calibration was done using panel data and sun sensor. After

radiometric calibration, all images were aligned with camera locations. The GCPs were loaded

and geometric rectification was done. All images with GCPs were selected. All GCPs were

positioned at their respective locations and camera optimization was done. Three different

approaches namely reconstruction error, projection accuracy, and re-projection error from the

gradual selection tool were used to delete unnecessary tie points to maximize camera and image

calibration. Once the images were radiometrically and geometrically calibrated and unnecessary

tie points were removed, final camera optimization was done before the creation of dense cloud

points. The RMSE values after final image optimization are presented in Table 4.1. The RMSE

values for camera calibration range from 0.12 m to 13 m. The orthomosaics of all UAV images

are presented in Fig. 4.1. The final orthomosaic images had pixel size from 0.0703 to 0.078. This

small variation was due to flight height.

Table 4.1: The RMSE values of geometric images for the training boxes before and after the

military training.

Camera optimization RMSE

Box1 Before 0.13

After 0.12

Box 2 Before 0.13

After 0.13

Box 3 Before 0.13

After 0.12

35

Figure 4.1: False color composite orthomosaic images of multispectral UAV images after

processing for three training boxes.

4.2 Vegetation indices

Eight different vegetation indices were calculated and exported. The mean values of the

vegetation indices from original pixel are presented in Table 4.2. The table indicates that the

different military training activities led to the decrement in mean values of vegetation indices for

each box. NDVI value ranges from highest of 0.81 in box 3 before and lowest of 0.77 in box 1

after. Similarly, GNDVI is the only vegetation indices with values increasing in box 1 and box2

after disturbances except in box 2. Likewise, all other vegetation indices' mean values decreased

36

after disturbances. This can well be observed from vegetation indices maps where we can

distinctly see the disturbances led to vegetation loss.

Table 4.2: Mean values of calculated vegetation indices from original images

Box1 Box2 Box 3

Before After Before After Before After

NDVI 0.8 0.77 0.83 0.78 0.81 0.79

GNDVI 0.65 0.66 0.68 0.67 0.7 0.72

MSAVI2 0.88 0.86 0.9 0.87 0.9 0.88

NDRE 0.48 0.37 0.5 0.4 0.51 0.42

OSAVI 0.8 0.77 0.83 0.78 0.81 0.79

RE MSR 0.69 0.49 0.75 0.54 0.77 0.57

WDRVI 0.31 0.23 0.39 0.27 0.35 0.28

ARVI 0.68 0.61 0.71 0.64 0.68 0.64

The maps for each vegetation index from original pixels are presented in Fig. 4.2 to Fig.

4.9. All the vegetation index maps show spatial variability. That is, the VI values differ from

place to place, corresponding to the spatial variability of military training intensity. Moreover,

except for GNDVI, all the vegetation indices show a trend of decrease after the training.

Figure 4.2: The maps Normalized Difference Vegetation Index (NDVI) for each box before and

after military training, showing the training induced vegetation cover loss.

37

Figure 4.3: The maps of Green Normalized Difference Vegetation Index (GNDVI) for each box

before and after military training, showing the training induced vegetation cover loss.

Figure 4.4: The maps of Red Edge Normalized Difference Vegetation Index (RENDVI) for each

box before and after military training, showing the training induced vegetation cover loss.

38

Figure 4.5: The maps of Optimized Soil Adjusted Vegetation Index (OSAVI) for each box


Figure 4.6: The maps of Modified Soil Adjusted Vegetation Index (MSAVI2) for each box


39

Figure 4.7: The maps of Red Edge Modified Simple Ratio (REMSR) for each box before and

after military training, showing the training induced vegetation cover loss.

Figure 4.8: The maps of Wide Dynamic Range Vegetation Index (WDRVI) for each box before

and after military training, showing the training induced vegetation cover loss.

40

4.3 Field data

Field data contains Daubenmire frame vegetation cover data for 5 m and 10 m plot sizes.

The descriptive statistics of vegetation cover for 5 m and 10 m plot sizes are presented in Table

4.3 and Table 4.4, respectively, where mean, standard deviation and confidence interval for each

box were calculated.

In Table 4.3, the mean values of vegetation cover data from 5 m plots for box 1 before

and after the military training, box 2 before and after the military training, box 3 before and after

the military training are 49.95, 49.32, 54.56, 56, 53.81 and 51.2, respectively. Similarly, in Table

4.4 vegetation cover data from 10m field plot for box 1 before and after the military training, box

Figure 4.9: The maps of Atmospherically Resistant Vegetation Index (ARVI) for each box before

and after military training, showing the training induced vegetation cover loss.

41

2 before and after the military training, box 3 before and after the military training are 50.03,

49.31, 54.57, 56.02, 53.61 and 51.2, respectively. The tables also show the standard deviation

and confidence interval (95%) for each box. The military training activities led to the decreased

mean values and increased standard deviations of vegetation cover for box 1, box 3 and the

pooled dataset, but not for box 2.

Table 4.3: Descriptive statistics for each of military training boxes for 5 m by 5 m plot size

Time

Disturbance

level Mean

Standard

deviation

Confidence Interval

(95%)

Box1

Before High 36.59 12.21 28.82, 44.34

Medium 49.6 8.15 44.42, 54.78

Low 51.44 5.67 47.67, 55.20

Undisturbed 62.15 5.91 58.38, 65.90

Average All 49.95 1.77 46.37, 53.51

After High 35.46 8.18 30.25, 40.66

Medium 45.75 5.83 42.04,4 9.45

Low 50.13 6.99 45.68, 54.57

Undisturbed 65.92 8.00 60.83, 71.00

Average All 49.32 1.9 45.49, 53.13

Box2

Before High 42.98 8.56 37.54, 48.41

Medium 47.69 6.04 43.84, 51.53

Low 56 6.85 51.64, 60.36

Undisturbed 71.58 9.02 65.85, 77.32

Average All 54.56 1.92 50.70, 58.42

After High 44.27 6.16 40.36, 48.18

Medium 50.15 7.11 45.63, 54.66

Low 56.75 6.79 52.43, 61.06

Undisturbed 72.81 5.68 69.20, 76.42

Average All 56 1.80 52.38, 59.61

Box3

Before High 41 9.97 34.66, 47.34

Medium 47.98 13.15 39.62, 56.33

Low 52.63 5.47 49.14, 56.10

Undisturbed 73.63 3.91 71.14, 76.10

Average All 53.81 2.17 49.44, 59.17

After High 31.81 9.98 25.47, 38.15

Medium 44.75 12.36 36.89, 52.60

Low 58.25 7.44 53.52, 62.97

Undisturbed 69.96 6.54 65.80, 74.11

42

Average All 51.2 2.46 46.24, 56.15

All

Boxes

Before All 52.77 1.14 50.52, 55.46

After All 52.17 1.21 49.77, 54.56

Table 4.4: Descriptive statistics for each of military training boxes for 10 m by 10 m plot size

Time

Disturbance

level Mean

Standard

deviation

Confidence Interval

(95%)

Box 1

Before All 50.03 2.97 43.50, 56.56

After All 49.31 3.54 41.52, 57.10

Box 2

Before All 54.57 3.7 46.43, 62.7

After All 56.02 3.56 48.63, 63.4

Box 3

Before All 53.61 3.96 44.89, 62.32

After All 51.2 4.6 41.08, 61.32

All

Boxes Before All 52.74 2.03

48.62, 56.85

After All 52.17 2.22 47.66, 56.68

4.4 Results from 5 m by 5 m field plots

4.4.1 Vegetation indices

The mean values of vegetation indices from 5 m resampled UAV images are presented in

Table 4.5. There is no change in the mean values of vegetation indices on the resampled data as

compared to original pixels except in box 2 before the military training (Table 4.2). The

resampling changed the mean values of the VIs in box 2 but the changes are not significant.

Table 4.5: Mean values of calculated vegetation indices from 5 m resampled images

Box1 Box2 Box 3


NDVI 0.8 0.77 0.8 0.78 0.81 0.79

GNDVI 0.65 0.66 0.66 0.67 0.7 0.72

MSAVI2 0.88 0.86 0.87 0.87 0.9 0.88

NDRE 0.48 0.38 0.49 0.4 0.51 0.42

OSAVI 0.8 0.77 0.8 0.78 0.81 0.79

REMSR 0.69 0.49 0.73 0.54 0.77 0.57

WDRVI 0.317 0.23 0.38 0.27 0.35 0.28

ARVI 0.67 0.61 0.684 0.64 0.68 0.65

43

4.4.2 Correlation analysis

The results of correlation coefficient between vegetation cover and all the spectral

variables from 5m resample images are presented in Table 4.6. The significance of correlation

between each variable with vegetation cover was examined based on Student-T distribution:

21

2

r

nrt

at the significant levels of 0.05 and 0.1. The correlations for all spectral

variables except REMSR, RENDVI, B2 and B3 are statistically significant for box 1 before the

military training. Except B4 and B5, all the spectral variables have significant correlations with

vegetation cover for box 1 after the military training and box 2 before the military training. All

the spectral variables are significantly correlated with vegetation cover for box 2 after the

military training. Similarly, all the spectral variables except B4 are significantly correlated with

vegetation cover for box 3 before the military training. However, the correlations of all the

spectral variables except WDRVI, B1, B2, B4 and B5 are not significant at given significance

level for box3 after the military training. For the pooled dataset from all three boxes, all the

spectral variables except for band 4 have significant correlations with vegetation cover. Overall,

the vegetation indices greatly improved the correlation with vegetation cover compared with the

original bands for all the boxes before and after the military training except box 3 after the

military training.

44

Table 4.6: Correlation coefficient between vegetation cover and spectral variables obtained from

linear correlation analysis

Box1 Box2 Box 3 All Boxes

Variables Before After Before After Before After Before After

NDVI 0.68* 0.56* 0.71* 0.56* 0.54* 0.15 0.59* 0.36*

GNDVI 0.36* 0.60* 0.75* 0.66* 0.59* 0.17 0.58* 0.27*

MSAVI2 0.69* 0.57* 0.68* 0.56* 0.56* 0.09 0.59* 0.39*

RENDVI 0.18 0.46* 0.76* 0.66* 0.35* 0.15 0.47* 0.31*

OSAVI 0.68* 0.56* 0.71* 0.56* 0.54* 0.15 0.59* 0.36*

REMSR 0.16 0.42* 0.77* 0.63* 0.33* 0.17 0.47* 0.28*

WDRVI 0.64* 0.53* 0.74* 0.56* 0.49* 0.24** 0.56* 0.30*

ARVI 0.66* 0.53* 0.73* 0.55* 0.51* 0.21 0.57* 0.32*

B1 0.36* 0.68* 0.51* 0.64* 0.51* 0.39* 0.45* 0.33*

B2 0.22 0.35* 0.36* 0.64* 0.19 0.47* 0.15** 0.13**

B3 0.36 0.47* 0.56* 0.53* 0.39* 0.04 0.43* 0.27*

B4 0.29** 0.085 0.23 0.37* 0.32 0.34* 0.015 0.0085

B5 0.29** 0.23 0.1 0.52* 0.54* 0.34* 0.21* 0.24*

(* indicating that the correlation coefficient is significantly different from zero at the significant

level of 0.05 and ** indicating that the correlation coefficient is significantly different from zero

at the significant level of 0.1)

The matrix of correlation coefficients among the spectral variables for each of the boxes

was analyzed to account for the multicollinearity among the spectral variables in the model

(Tables 4.7 to 4.12). The correlation matrix is used to see the association and direction of

association among all variables to be used in regression analysis. NDVI was highly correlated

with WDRVI, OSAVI, MSAVI2, and ARVI for all boxes. Similarly, RENDVI was highly

correlated with REMSR. GNDVI was highly correlated with NDVI for box 1 after the military

training, box 2 before the military training, and box 2 after the military training. Overall, the

vegetation indices are highly correlated with each other but relatively have lower correlations

with the original bands.

45

Table 4.7: The matrix of correlation coefficients among the variables for box 1 before the military training.

Table 4.8: The matrix of correlation coefficients among the variables for box 1 after the military training.

cover NDVI WDRVI REMSR OSAVI RENDEVI MSAVI2 GNDVI ARVI B1 B2 B3 B4 B5

cover 1

NDVI 0.67506 1

WDRVI 0.635901 0.994265 1

REMSR 0.158693 0.615445 0.652193 1

OSAVI 0.675062 1 0.994265 0.615443 1

NDRE 0.178546 0.632466 0.667037 0.99878 0.632463 1

MSAVI2 0.687193 0.998934 0.98838 0.600586 0.998934 0.618389 1

GNDVI 0.363519 0.815501 0.843874 0.920164 0.815499 0.929849 0.802626 1

ARVI 0.657641 0.998737 0.998079 0.62984 0.998737 0.646123 0.995495 0.826233 1

B1 -0.30166 -0.44955 -0.4372 -0.34196 -0.44953 -0.37103 -0.45169 -0.44818 -0.45194 1

B2 0.221214 0.190974 0.18853 -0.104 0.190997 -0.12293 0.191495 -0.03988 0.181435 0.744743 1

B3 -0.35567 -0.58323 -0.57666 -0.40074 -0.58321 -0.42968 -0.58295 -0.52546 -0.59011 0.964099 0.649671 1

B4 0.294961 0.369562 0.376538 0.065627 0.369585 0.050249 0.365791 0.178425 0.363955 0.619748 0.969863 0.509366 1

B5 0.286323 0.516454 0.540173 0.390821 0.516475 0.374546 0.505888 0.461018 0.518367 0.477812 0.857299 0.357372 0.941393 1

cover WDRVI REMSR OSAVI NDVI RENDVI MSAVI GNDVI ARVI B1 B2 B3 B4 B5

cover 1

WDRVI 0.558547 1

REMSR 0.464579 0.7617 1

OSAVI 0.597422 0.988634 0.691797 1

NDVI 0.597418 0.988634 0.691794 1 1

NDRE 0.498318 0.79082 0.99635 0.729689 0.729685 1

MSAVI 0.609649 0.975693 0.653818 0.997533 0.997533 0.695349 1

GNDVI 0.647951 0.926926 0.914599 0.903044 0.903041 0.936759 0.884754 1

ARVI 0.565009 0.996657 0.724483 0.996742 0.996742 0.757981 0.989101 0.912165 1

B1 -0.73397 -0.78912 -0.51444 -0.82938 -0.82939 -0.55743 -0.84059 -0.76019 -0.80627 1

B2 -0.42965 -0.43028 -0.42439 -0.45529 -0.45531 -0.44914 -0.46267 -0.50246 -0.45163 0.731237 1

B3 -0.50042 -0.87101 -0.57541 -0.89438 -0.89439 -0.61213 -0.89822 -0.7784 -0.89285 0.883906 0.757063 1

B4 -0.11046 0.058674 -0.04348 0.029177 0.029159 -0.05489 0.016133 -0.02627 0.032695 0.355196 0.864395 0.378821 1

B5 0.239381 0.533833 0.526483 0.477992 0.477974 0.523406 0.44893 0.53206 0.494687 -0.06512 0.459144 -0.06226 0.816446 1

46

Table 4.9: The matrix of correlation coefficients among the variables for box 2 before the military training. cover GNDVI MSAVI RENDVI NDVI OSAVI REMSR WDRVI ARVI B1 B2 B3 B4 B5

cover 1

GNDVI 0.603387 1

MSAVI 0.57671 0.868701 1

NDRE 0.668407 0.952784 0.807517 1

NDVI 0.588892 0.90413 0.991866 0.830905 1

OSAVI 0.588888 0.904125 0.991865 0.830894 1 1

REMSR 0.667495 0.949067 0.776244 0.997425 0.805731 0.80572 1

WDRVI 0.587539 0.931223 0.949061 0.837702 0.981074 0.981076 0.823279 1

ARVI 0.591019 0.921504 0.976283 0.839583 0.995831 0.995832 0.819113 0.994498 1

B1 -0.51837 -0.67771 -0.70151 -0.76819 -0.6715 -0.67148 -0.74023 -0.60322 -0.64537 1

B2 -0.4205 -0.50153 -0.38988 -0.64217 -0.37233 -0.3723 -0.62862 -0.33105 -0.3592 0.91277 1

B3 -0.54888 -0.75054 -0.81456 -0.80595 -0.78594 -0.78592 -0.77308 -0.71727 -0.75999 0.981128 0.835112 1

B4 -0.37899 -0.38224 -0.27483 -0.55503 -0.25169 -0.25166 -0.54298 -0.20397 -0.23531 0.863488 0.98878 0.766704 1

B5 -0.2 -0.13068 -0.04331 -0.31698 -0.01045 -0.01042 -0.3052 0.045574 0.010524 0.733588 0.920409 0.607079 0.962212 1



cover 1

WDRVI 0.670151 1

REMSR 0.862595 0.825416 1

OSAVI 0.647156 0.990441 0.783845 1

NDVI 0.647154 0.99044 0.783841 1 1

NDRE 0.86713 0.822959 0.997791 0.786026 0.786022 1

MSAVI 0.633535 0.977854 0.758625 0.997326 0.997326 0.762976 1

GNDVI 0.81849 0.94404 0.954063 0.920507 0.920505 0.956779 0.902153 1

ARVI 0.6417 0.995942 0.786659 0.997843 0.997843 0.786334 0.991017 0.921914 1

B1 -0.7891 -0.82819 -0.91064 -0.82013 -0.82013 -0.91743 -0.8114 -0.92051 -0.80726 1

B2 -0.79537 -0.79633 -0.86485 -0.77206 -0.77206 -0.87286 -0.75384 -0.8822 -0.7741 0.896628 1

B3 -0.59293 -0.96137 -0.71421 -0.9772 -0.9772 -0.71677 -0.97737 -0.86406 -0.97579 0.794923 0.799927 1

B4 -0.48279 -0.05779 -0.4223 -0.03274 -0.03274 -0.43702 -0.02091 -0.26834 -0.02751 0.442432 0.633844 0.111903 1

B5 0.65925 0.867255 0.848033 0.833397 0.833392 0.836077 0.811211 0.884284 0.840801 -0.73698 -0.5742 -0.7116 0.12074 1

47



cover 1

WDRVI 0.492523 1

REMSR 0.33476 0.891362 1

OSAVI 0.543219 0.990532 0.890518 1

NDVI 0.543218 0.990532 0.890517 1 1

NDRE 0.345734 0.896088 0.998655 0.899984 0.899983 1

MSAVI 0.561458 0.979295 0.883728 0.997753 0.997753 0.895462 1

GNDVI 0.587193 0.90127 0.853368 0.898982 0.898981 0.854426 0.889338 1

ARVI 0.512665 0.996685 0.893888 0.997844 0.997844 0.901584 0.991697 0.892451 1

B1 -0.50568 -0.85843 -0.75003 -0.85164 -0.85165 -0.75896 -0.84167 -0.90145 -0.8493 1

B2 -0.19126 -0.54184 -0.5264 -0.52135 -0.52135 -0.53503 -0.5075 -0.63859 -0.5262 0.850929 1

B3 -0.39331 -0.93791 -0.84344 -0.9378 -0.9378 -0.85691 -0.93244 -0.80222 -0.94352 0.892707 0.701786 1

B4 0.324465 -0.14904 -0.31461 -0.14252 -0.14253 -0.3261 -0.14166 -0.0581 -0.15586 0.342096 0.672257 0.441108 1

B5 0.536526 0.673498 0.639741 0.674209 0.674202 0.628321 0.66692 0.711041 0.668474 -0.38297 0.076484 -0.39225 0.525799 1



cover 1

WDRVI -0.23712 1

REMSR -0.17034 0.891627 1

OSAVI -0.14785 0.986728 0.877024 1

NDVI -0.14785 0.986728 0.877025 1 1

NDRE -0.14623 0.885287 0.996623 0.88124 0.88124 1

MSAVI -0.0867 0.966862 0.862367 0.995279 0.995279 0.872238 1

GNDVI 0.165596 0.629337 0.687075 0.665704 0.6657 0.709068 0.678098 1

ARVI -0.21341 0.997707 0.883009 0.994761 0.994761 0.880802 0.980376 0.634588 1

B1 -0.39397 -0.3993 -0.47526 -0.4303 -0.43032 -0.47863 -0.45439 -0.43463 -0.39885 1

B2 -0.47056 -0.12533 -0.2846 -0.15811 -0.15812 -0.3033 -0.18285 -0.67248 -0.12021 0.748612 1

B3 -0.04284 -0.72727 -0.68271 -0.7166 -0.71662 -0.66784 -0.70971 -0.38192 -0.71954 0.852472 0.422057 1

B4 -0.34166 0.379596 0.134252 0.383916 0.383896 0.140139 0.368724 0.213232 0.394011 0.605822 0.514216 0.299443 1

B5 -0.34244 0.764967 0.623223 0.76133 0.761315 0.62866 0.743033 0.515378 0.771598 0.22286 0.251957 -0.12538 0.851811 1

48

4.4.3 Vegetation cover models

Table 4.13 shows the final vegetation cover model for each box and also the overall

pooled data. In addition, the results from cross-validation are also presented to assess model

accuracy of predictions. VIF for each model was calculated and the predictors with VIF greater

than 10 were removed from the final model to reduce multicollinearity among the predictors.

The RMSE and RRMSE values for modelling and cross-validation were calculated to evaluate

the regression models. The RMSE values for separate modelling of the training boxes ranges

from 8.38% to 11.34%, while the overall vegetation cover model has RMSE values of 10.99%

and 13.13%, respectively for before and after military training induced vegetation disturbances.

Similarly, the RRMSE value ranges from 15.36% to 22.14% for separate modelling of the

training boxes, while the RRMSE values of 20.82% and 25.16% were respectively obtained for

before and after military training. The results implied that separately modelling the military

training induced vegetation disturbances for the training boxes led to greater accuracy than

modelling the vegetation disturbances by pooling the data together. Moreover, the accuracy of

modelling was higher before the training than that after the training. The similar results were

found by cross-validation of the model predictions. In Fig. 4.10 to Fig. 4.13, the residuals of

vegetation cover predictions by using cross-validation were graphed against the observations,

which show that there are no obvious overestimations and underestimations taking place for all

the models.

The models were based on 13 spectral variables, including 8 vegetation indices and 5

original image bands. Vegetation indices that were widely used to map vegetation at different

settings were selected and were limited due to limited number of bands on original image. For

box 2 before the military training the REMSR and band 3 were selected in the final model. The

49

REMSR was selected because of the highest correlation with vegetation cover, while band 3 was

picked up due to its relatively lower correlation with REMSR. Similarly after military for box 2,

GNDVI and band 5 were selected because of highest correlation with vegetation cover. While in

box 1 before military training the REMSR, band 1 and band 2 were selected. Band 1 was

selected because of the correlation with vegetation cover, while REMSR and band 2 were picked

because of their relatively lower correlation with band 1. For box 1 after training REMSR, NDVI

were selected because of highest correlation, while band 4 was picked due to its relatively lower

correlation with other variables. For box 3 before military training, REMSR, MSAVI and

GNDVI were selected because of highest correlation with vegetation cover. OSAVI, band 1 and

band 3 were selected for final model. Band 1 was selected because of highest correlation with

vegetation cover, while OSAVI and band 3 were picked up due to lower correlation with band 1.

For overall pooled data, before the disturbance, REMSR and NDVI were selected as they are

highly correlated with vegetation cover. Similarly, OSAVI, B1 and B3 were selected for overall

pooled data after the disturbances because of high correlation with vegetation cover.

50


assessment of predictions for each of three boxes using cross-validation

Cross-validation

Time Variables Estimate VIF

RMSE

(%)

RRMSE

(%) RMSE(%) RRMSE(%)

Box1 Before Intercept 87.11754

REMSR -26.9387 1.2

B1 -0.09743 2.66

B2 0.026717 2.38 8.42 16.86 9.27 18.55

After Intercept -58.2102

REMSR 15.00923 1.96

NDVI 154.255 1.93

B4 -0.00318 1.06 10.64 21.57 11.46 23.23

Box2 Before Intercept -23.5

REMSR 112.4 2.19

B3 0.000181 2.19 8.38 15.36 8.82 16.16


GNDVI 270.6581 4.89

B5 -0.00413 4.89 9.05 16.16 10.01 17.87


REMSR -250.88 5.06

MSAVI 295.63 6.58

GNDVI 361.93 5.31 10.13 18.82 11.07 20.57

After Intercept 146.5

REMSR -146.7 4.82

B1 -0.2175 7.12

B3 0.07895 6.14

B5 0.00701 4.05 11.34 22.14 12.64 24.69

Overall Before Intercept -44.46

REMSR 24.78 1.77

NDVI 105.9 1.77 10.99 20.82 11.21 21.24


OSAVI 104.484 3.1

B1 -0.0417 3.56

B3 0.0217 6.98 13.13 25.16 13.55 25.97

51

Figure 4.10: The Residuals of predicted vegetation cover graphed against the fitted values for box 1 before and after

military training activities at the plot size of 5 m by 5 m.



52



Figure 4.13: The Residuals of predicted vegetation cover graphed against the fitted values for the pooled data from all

boxes before and after military training activities at the plot size of 5 m by 5 m

53

4.4.4 Mapping vegetation covers change

Equations 4.1 to 4.8 are eight models obtained from the datasets of 5 m by 5 m plot size.

Based on the models, a vegetation cover prediction map for each box was derived. The maps are

shown in Fig. 4.14 to Fig. 4.16. In addition, the maps include the vegetation cover change

derived from differencing the vegetation cover predictions from the models before and after the

military training. The vegetation cover and vegetation cover change varied spatially with most of

the areas showing the decrease of vegetation cover after the military training.

1_ 87.12 26.94* 0.0974* 1 0.27* 2Box before REMSR B B 4.1

1_ 58.21 15.009* 154.255* 0.0032* 4Box after REMSR NDVI B 4.2

2_ 23.5 112.4* 0.000181* 3Box before REMSR B 4.3

2_ 77.915 270.66* 0.00413* 5Box after GNDVI B 4.4

3_ 253.94 250.88* 295.63* 361.93*Box before REMSR MSAVI GNDVI 4.5

3_ 146.5 146.7* 0.2175* 1 0.07895* 3 0.00701* 5Box after REMSR B B B 4.6

_ 44.46 24.78* 105.9*All before REMSR NDVI 4.7

_ 16.91 104.45* 0.0417* 1 0.0217* 3All after OSAVI B B 4.8

Table 4.14 shows the vegetation changes for each box after military training activities

using image differencing tool. The values are presented at a threshold value of 1% change in

vegetation cover. Box 3 has the highest decrease in vegetation cover with total of 31.505

hectares (ha) of decrease and very few increased in cover percentage with value of 1.225 ha. Box

2 has highest increase in vegetation cover with 9.76 ha total increase and 13.824 ha of decrease.

Box 1 has 23.46 ha of decrease in vegetation cover of 1% whereas only 3.75 ha increased in

vegetation cover.

54

Table 4.14: Vegetation cover change statistics for each box derived from image differencing of

the maps before and after militrainig.

Change Box1 Box2 Box3

Decreased 23.45 13.82 31.50

Some Decrease 0 1.10 0.99

Unchanged 1.73 1.02 0.40

Some Decrease 1.60 0 0.58

Increased 3.73 9.76 1.22

(*at least 1% decrease in vegetation cover)

Figure 4.14: Vegetation cover and change maps for box 1 from 5 m resolution resampled data.

55



56

4.5 Results from 10 m by10 m field plots

4.5.1 Vegetation indices

The mean values of vegetation indices from 10 m resampled UAV images are presented

in Table 4.15. There is no change in the mean values of vegetation indices due to the 10 m

resampling as compared to those from the original pixels and the 5 m resampling except in box 2

before the military training, and the changes in box 2 were significant.

Table 4.15: Mean values of calculated vegetation indices from 10 m resampled images

Box1 Box2 Box 3


NDVI 0.8 0.77 0.77 0.77 0.81 0.79

GNDVI 0.65 0.66 0.64 0.66 0.695 0.72

MSAVI2 0.88 0.86 0.84 0.86 0.9 0.88

NDRE 0.48 0.38 0.47 0.39 0.51 0.42

OSAVI 0.8 0.77 0.773 0.77 0.81 0.79

RE MSR 0.7 0.49 0.702 0.54 0.77 0.57

WDRVI 0.32 0.23 0.37 0.27 0.35 0.28

ARVI 0.67 0.61 0.66 0.64 0.68 0.65

4.5.2 Correlation analysis

The results of correlation coefficient between vegetation cover and all the spectral

variables from 10 m by 10 m plots are presented in Table 4.16. All the vegetation indices except

REMSR and RENDVI are significantly correlated with vegetation cover for box 1 before the

military training, but none of original bands. All the spectral variables except B4 and B5 have

significant correlations with vegetation cover for box 1 after the military training. Except for B2,

B4 and B5, all the spectral variables are significantly correlated with vegetation cover for box 2

before and after the military training. For box 3, out of 13 spectral variables, only NDVI,

GNDVI, MSAVI2, OSAVI, ARVI and B1 are significant before the training, and all the VIs

except for GNDVI and the original bands have significant correlations with vegetation cover. For

57

the pooled dataset from all three boxes, all the spectral variables except for band 2 and band 4

have significant correlations with vegetation cover before the military training, while after the

military training, only MSAVI2, band 1 and band 3 are significantly correlated with vegetation

cover.

Table 4.16: Correlation coefficient between vegetation cover and spectral variables obtained

from linear correlation analysis for 10 m resampled data

Box1 Box2 Box 3 All Boxes

Variables Before After Before After Before After Before After

NDVI 0.81* 0.59* 0.73* 0.82* 0.58* 0.57** 0.65* 0.26

GNDVI 0.49** 0.68* 0.75* 0.90* 0.63* 0.01 0.62* 0.22

MSAVI2 0.82* 0.60* 0.70* 0.81* 0.63* 0.52** 0.66* 0.28**

RENDVI 0.36 0.54** 0.77* 0.88* 0.35 0.53** 0.52* 0.25

OSAVI 0.81* 0.59* 0.73* 0.82* 0.58* 0.57** 0.65* 0.26

REMSR 0.33 0.51** 0.79* 0.86* 0.34 0.56** 0.53* 0.22

WDRVI 0.77* 0.54** 0.76* 0.83* 0.43 0.63* 0.60* 0.2

ARVI 0.80* 0.55** 0.75* 0.81* 0.50** 0.61* 0.63* 0.21

B1 0.38 0.83* 0.51** 0.91* 0.54** 0.2 0.46* 0.35*

B2 0.17 0.50** 0.4 0.91* 0.31 0.44 0.19 0.17

B3 0.49 0.54** 0.54** 0.77* 0.37 0.33 0.44* 0.30**

B4 0.24 0.22 0.25 0.60* 0.12 0.38 0.04 0.09

B5 0.27 0.14 0.27 0.80* 0.34 0.54 0.26* 0.11

The matrix of correlation coefficients among the spectral variables for each of the boxes

was analyzed to account for the multicollinearity among the spectral variables in the model

(Tables 4.17 to 4.22). The correlation matrix was used to see the association and direction of

association among all variables to be used in regression analysis. NDVI was highly correlated

with WDRVI, OSAVI, MSAVI2, and ARVI for all boxes. Similarly, RENDVI was highly

correlated with REMSR. GNDVI was highly correlated with NDVI for box 1 after military

training, box 2 before military training and box 2 after military training. Overall vegetation

indices are highly correlated with each other but have relatively lower correlations with original

bands.

58


cover NDVI WDRVI REMSR OSAVI RENDVI MSAVI GNDVI ARVI B1 B2 B3 B4 B5

cover 1

NDVI 0.815755 1

WDRVI 0.774797 0.995774 1

REMSR 0.333285 0.769592 0.801288 1

OSAVI 0.815753 1 0.995774 0.769593 1

NDRE 0.360982 0.787264 0.816158 0.998747 0.787263 1

MSAVI 0.827758 0.999323 0.991796 0.757212 0.999323 0.775956 1

GNDVI 0.496551 0.861915 0.892626 0.93666 0.861916 0.942298 0.849016 1

ARVI 0.80062 0.999189 0.998318 0.78023 0.999189 0.79703 0.997149 0.871753 1

B1 -0.37894 -0.37322 -0.37187 -0.29174 -0.37319 -0.31967 -0.37096 -0.31198 -0.38105 1

B2 0.171537 0.226009 0.219281 0.127378 0.226041 0.108411 0.229354 0.171001 0.213636 0.796183 1

B3 -0.48416 -0.51288 -0.5082 -0.35816 -0.51285 -0.38906 -0.51177 -0.3917 -0.52143 0.965745 0.705471 1

B4 0.241457 0.360455 0.364164 0.275291 0.360486 0.258055 0.359387 0.354695 0.352088 0.702879 0.976948 0.606846 1

B5 0.274266 0.515279 0.531461 0.529387 0.515306 0.511675 0.509188 0.577271 0.512568 0.54897 0.895295 0.448585 0.959471 1



cover 1

WDRVI 0.53839 1

REMSR 0.517446 0.774306 1

OSAVI 0.588124 0.992224 0.735041 1

NDVI 0.588123 0.992224 0.735043 1 1

NDRE 0.544217 0.79557 0.998822 0.760542 0.760544 1

MSAVI 0.604755 0.983469 0.711461 0.998346 0.998346 0.738741 1

GNDVI 0.680821 0.906362 0.936757 0.896657 0.896659 0.94885 0.886148 1

ARVI 0.546592 0.997386 0.749846 0.997607 0.997607 0.77342 0.992619 0.896058 1

B1 -0.8356 -0.73803 -0.65587 -0.77713 -0.77714 -0.67869 -0.78951 -0.83127 -0.75023 1

B2 -0.49635 -0.35461 -0.54829 -0.3861 -0.38611 -0.55484 -0.39774 -0.56837 -0.37592 0.739888 1

B3 -0.54028 -0.86792 -0.68644 -0.89212 -0.89213 -0.70995 -0.89909 -0.83146 -0.88929 0.848832 0.717488 1

B4 -0.21527 0.06337 -0.23578 0.03146 0.031445 -0.23345 0.017306 -0.17664 0.03845 0.411104 0.899982 0.390769 1

B5 0.141404 0.539106 0.334445 0.492201 0.492187 0.339857 0.467814 0.382039 0.504434 -0.02418 0.540694 -0.05704 0.832242 1

59



cover 1

WDRVI 0.764988 1

REMSR 0.7911 0.960011 1

OSAVI 0.73456 0.976523 0.936866 1

NDVI 0.73455 0.97652 0.936866 1 1

NDRE 0.776185 0.960766 0.995533 0.956721 0.956723 1

MSAVI 0.700851 0.942454 0.903155 0.992143 0.992145 0.932961 1

GNDVI 0.75297 0.986427 0.984066 0.953527 0.953525 0.979874 0.913667 1

ARVI 0.75065 0.992549 0.952057 0.995411 0.995409 0.963855 0.975711 0.973804 1

B1 -0.50967 -0.76312 -0.77696 -0.86942 -0.86943 -0.82416 -0.91366 -0.74951 -0.82661 1

B2 -0.40233 -0.66089 -0.71222 -0.77739 -0.77741 -0.76345 -0.82706 -0.67285 -0.73052 0.974794 1

B3 -0.54768 -0.81266 -0.79996 -0.91222 -0.91223 -0.84716 -0.95142 -0.78918 -0.87301 0.991223 0.947795 1

B4 -0.25688 -0.44767 -0.52524 -0.59975 -0.59977 -0.58765 -0.67283 -0.46093 -0.53606 0.90511 0.963475 0.856693 1

B5 0.268196 0.158444 0.074674 -0.02466 -0.02468 0.0025 -0.12585 0.144428 0.05506 0.4931 0.627115 0.413185 0.80189 1



cover 1

WDRVI 0.831684 1

REMSR 0.864515 0.853029 1

OSAVI 0.822674 0.991436 0.802559 1

NDVI 0.822672 0.991435 0.802556 1 1

NDRE 0.8803 0.850161 0.99725 0.802905 0.802902 1

MSAVI 0.814265 0.978822 0.771233 0.997094 0.997095 0.772862 1

GNDVI 0.90784 0.955489 0.957759 0.927827 0.927826 0.961455 0.90611 1

ARVI 0.813896 0.995659 0.808709 0.998573 0.998573 0.807041 0.992246 0.930453 1

B1 -0.91154 -0.86927 -0.97732 -0.83179 -0.83178 -0.9824 -0.80645 -0.96678 -0.83264 1

B2 -0.91035 -0.92217 -0.91561 -0.8962 -0.8962 -0.92599 -0.87438 -0.95897 -0.90018 0.950525 1

B3 -0.77472 -0.96645 -0.72244 -0.98442 -0.98442 -0.72263 -0.98625 -0.86731 -0.98343 0.76556 0.873486 1

B4 -0.60469 -0.37257 -0.66456 -0.32219 -0.32219 -0.68549 -0.2929 -0.53887 -0.32575 0.68023 0.68759 0.309905 1

B5 0.800663 0.892696 0.941241 0.852043 0.85204 0.928531 0.82581 0.947049 0.858515 -0.90672 -0.82874 -0.75879 -0.37342 1

60



cover 1

WDRVI 0.429308 1

REMSR 0.337506 0.770691 1

OSAVI 0.575863 0.974797 0.819873 1

NDVI 0.575865 0.974798 0.819872 1 1

NDRE 0.355087 0.763752 0.99903 0.817705 0.817706 1

MSAVI 0.633531 0.945335 0.831055 0.994213 0.994213 0.8314 1

GNDVI 0.628133 0.782395 0.763701 0.83604 0.836038 0.757878 0.842234 1

ARVI 0.499078 0.992876 0.801448 0.992989 0.99299 0.797609 0.975718 0.794552 1

B1 -0.53794 -0.76246 -0.57959 -0.74426 -0.74427 -0.58028 -0.71994 -0.80822 -0.74423 1

B2 -0.30502 -0.35002 -0.37474 -0.32617 -0.32618 -0.38421 -0.30935 -0.44667 -0.33223 0.808909 1

B3 -0.36669 -0.86217 -0.67727 -0.82381 -0.82382 -0.68267 -0.79468 -0.54278 -0.85844 0.786059 0.621984 1

B4 0.127301 0.019426 -0.26269 0.03771 0.037699 -0.27974 0.038286 0.148144 0.01149 0.214615 0.667705 0.430985 1

B5 0.340076 0.615126 0.53109 0.660786 0.660776 0.514496 0.666271 0.695678 0.628683 -0.24032 0.315759 -0.14835 0.676198 1



cover 1

WDRVI -0.62847 1

REMSR -0.55674 0.972498 1

OSAVI -0.56976 0.992516 0.960524 1

NDVI -0.56977 0.992519 0.960528 1 1

NDRE -0.53435 0.967324 0.997998 0.961898 0.9619 1

MSAVI -0.51899 0.980515 0.947427 0.997016 0.997014 0.952605 1

GNDVI -0.01414 0.540746 0.628364 0.58539 0.585378 0.654732 0.60986 1

ARVI -0.6157 0.998541 0.965439 0.997022 0.997024 0.962691 0.988271 0.546712 1

B1 -0.20337 -0.11066 -0.25849 -0.05599 -0.05601 -0.23861 -0.03774 -0.17603 -0.07168 1

B2 -0.43889 0.033769 -0.15177 0.027585 0.027585 -0.16576 0.012996 -0.65332 0.052099 0.680565 1

B3 0.331551 -0.62987 -0.68068 -0.56113 -0.56116 -0.64706 -0.52555 -0.14924 -0.59648 0.752874 0.224643 1

B4 -0.37656 0.44954 0.307473 0.514297 0.514277 0.328197 0.537928 0.330455 0.487035 0.775721 0.458711 0.367504 1

B5 -0.53641 0.826173 0.739615 0.870098 0.870086 0.755803 0.883333 0.560351 0.850438 0.407884 0.243064 -0.09329 0.864221 1

61

4.5.3. Vegetation cover models

Table 4.23 shows the results of the final vegetation cover model and cross validation for

each box and for the overall pooled data for 10 m resampled data. When the vegetation cover

was modelled separately for the training boxes, the RMSE values of the modelling ranges from

3.44% to 10.21%, whereas the overall pooled data resulted in the RMSE values of 9.07% and

11.2%, respectively, for before and after the military training. Similarly, the RRMSE values

ranges from 6.88% to 19.04% for the separate modellings of the training boxes, while the

RRMSE values of 16.57% and 21.43% were respectively obtained for before and after the

military training for the pooled dataset. The same implications that were found from the datasets

of 5 m by 5 m plot size could be applied to the datasets of 10 m by 10 m plot size. That is,

separately modelling the military training induced vegetation disturbances for the training boxes

led to greater accuracy than modelling the vegetation disturbances by pooling the data together,

and the accuracy of modelling was higher before the training than that after the training.

However, the box 3 has the worst performance of both modelling and predictions. The results

were validated by cross-validation of the model predictions. In addition, there are no

systematical overestimations and underestimations found in terms of the residuals of predictions

against the observations in Fig. 4.17 to Fig. 4.20.

For box 1 before the military training, REMSR and NDVI were selected and after

military training, REMSR and GNDVI were selected because these variables have highest

correlation with vegetation cover. Similarly for box 2 before military training, NDVI and band 5

were selected as they have highest correlation with vegetation cover, while GNDVI and band 5

were selected for box 2 after military training because of the same reason. For box 3 before

military training, GNDVI was selected because it has highest correlation with vegetation cover.

62

Similarly for box 3 after military training, REMSR, band 1 and band 5 were selected because

REMSR is highly correlated with vegetation cover, and band1 and band 5 have low correlations

with REMSR. Band 1 and band 5 were selected for the pooled dataset before the disturbance,

while REMSR and band 1 were selected for the pooled dataset after the disturbance as they are

highly correlated with vegetation cover and weakly correlated with each other.


assessment of predictions for each of three boxes using cross-validation

Cross-validation

Time Variables Estimate VIF RMSE(%) RRMSE(%) RMSE(%) RRMSE(%)


REMSR -164.06 2.45

NDVI 276.2 2.45 3.44 6.88 4.38 8.75


REMSR -362.3 8.16

GNDVI 822.7 8.16 7.69 15.59 11.82 23.97


NDVI 152.1525 1

B5 0.001559 1 7.54 13.82 9.114 16.7

After Intercept -130

GNDVI 392.8 9.69

B5 -0.00677 9.69 4.194 7.49 5.033 8.98


GNDVI 391.8

10.21 19.044 12.58 23.46

After Intercept 103.8391

REMSR -46.3409 2.57

B1 -0.16791 3.19

B5 0.08482 5.55 9.704 18.95 15.74 30.74

Overall Before Intercept 53.6034

B1 -0.0427 1.16

B5 0.00311 1.16 9.07 16.57 9.69 18.37

After Intercept 120

REMSR -101.9 3.29

B1 -0.0734 3.68

B5 0.00534 2.522 11.2 21.43 12.71 24.32

63

Figure 4.17: The Residuals of predicted vegetation cover graphed against the fitted values for box 1 before and

after military training activities at the plot size of 10 m by 10 m.

Figure 4.18: The Residuals of predicted vegetation cover graphed against the fitted values for box 2 before and

after military training activities at the plot size of 10 m by 10 m.

64



Figure 4.20: The Residuals of predicted vegetation cover graphed against the fitted values for overall pooled data

before and after military training activities at the plot size of 10 m by 10 m.

65

4.5.3 Mapping vegetation covers change

Equations 4.9 to 4.16 are eight models obtained from the datasets of 10 m by 10 m plot

size. Based on the models, a vegetation cover prediction map for each box was derived. The

maps are shown in Fig. 4.21 to Fig. 4.23. In addition, the maps include the vegetation cover

change derived from differencing the before and after vegetation cover maps. Overall, the

vegetation cover and its change differed from place to place for all the boxes. The decrease of

vegetation cover dominated the training boxes especially for box 1 and box 3, implying that the

military training decreased the vegetation cover.

1_ 55.32 164.06* 276.2*Box before REMSR NDVI 4.9

1_ 308.7 362.3* 822.7*Box after REMSR GNDVI 4.10

2_ 87.08 152.15* 0.001559* 5Box before NDVI B 4.11

2_ 130 392.8* 0.00677* 5Box after GNDVI B 4.12

3_ 203.1 391.8*Box before GNDVI 4.13

3_ 103.84 46.3409* 0.16791* 1 0.08482* 5Box after REMSR B B 4.14

_ 53.60 0.0427* 1 0.00311* 5All before B B 4.15

_ 120 101.9* 0.0734* 1 0.00534* 5All after REMSR B B 4.16

Table 4.24 shows the vegetation cover changes for each box after the military training

activities using image differencing tool. The values are presented at a threshold value of 1%

change in vegetation cover. Box 3 has the highest decrease in vegetation cover with total of

28.48 hectares (ha) of decrease and very few pixels increased in cover percentage with 3.91 ha.

Box 2 has 10.085 ha increase in vegetation cover and 13.75 ha decrease. Box 1 has 17.23 ha of

decrease in vegetation cover, whereas only 8.25 ha increase.

66

Table 4.24: Vegetation cover change statistics for each box derived from image differencing

from the before and after military training vegetation cover maps.

(*at least 1% decrease in vegetation cover)


Change Box1(ha) Box2(ha) Box3(ha)

Decreased 17.23 13.75 28.48

Some Decrease 2.295 0 0.792

Unchanged 1.287 1.17 0.6112

Some Decrease 2.11 1.9 1.182

Increased 8.25 10.085 3.91

67



68

5. CHAPTER 5

6. DISCUSSION AND CONCLUSIONS

5.1 Discussion

The modelling and prediction of vegetation cover for army installations are challenging

because the military training activities are diverse, cumulative, and interrelated (Rijal 2017). In

addition, the training activities that have occurred in previous years and resultant vegetation

cover due to the intensity of training have a significant impact on predicting vegetation cover

(Howard et al. 2013). Formation of roads and gullies due to previous military training can be

easily visible before disturbance occurred which can lead to underestimation of vegetation cover

loss that happened in the current year. In this study, high-resolution multispectral UAV images

were used as the potential for mapping vegetation cover and disturbances caused by military

training activities.

The pre-processing of UAV images is a complicated process and requires time and

configuration of processing units should be able to process image without difficulty. Due to

variable flight heights, unbalance UAVs and small coverage, the UAV images require advanced

geometric and radiometric corrections and calibration, and mosicing processing before their use.

Moreover, there is a need to set up a measure that can be used to assess the quality of the images.

In this study, a sharpness level of greater than 0.65 was utilized. However, the parameter might

vary depending on different studies. In addition, there is also a need to determine the optimal

number of GCPs per area unit. However, this was not done in this study due to limited time.

The basic configuration recommended by Agisoft requires up to 32GB of RAM with

2.0+ GHz and 700+ CUDA cores GPU with advanced configuration up to 128 GB RAM and

Extreme configuration of more than 128 GB RAM. 16 GB RAM, Intel core 7 CPU with Intel

69

UHD Graphics configuration was used for pre-processing UAV images. Table 5.1 provides

details about time taken to process UAV images. The total time of image processing can be

reduced if computational intensity is increased.

Table 5.1: Total time spent to process UAV images for six boxes

Image Item Total time

Image alignment and calibration 30 hours

Dense cloud point and orthomosaic 18 hours

Vegetation indices calculation and export 9 hours

This study found that the use of high-resolution multispectral UAV imagery has the

potential for vegetation mapping. Similar potential was observed and reported on various

researches to map vegetations at different settings (Arnold et al. 2013, Feng et al. 2015, Korehisa

et al. 2014). Vegetation monitoring and assessment are among the most common and widely

used applications of remote sensing using vegetation indices calculated from reflectance

measurements using different bands (Purevdorj et al. 1998, Jafari et al. 2007, Ali et al. 2018).

Vegetation indices such as NDVI are widely used to map the occurrence of vegetation and its

health. In this study, eight different vegetation indices along with original bands from UAV

images were used to model vegetation cover. We applied a method similar to Ali et al (2018) and

Jafari et al (2007) where the researchers used different vegetation indices to map and assess

vegetation cover. Jafari et al (2007) concluded that perennial plant cover prediction was more

accurate within a land system, so different vegetation indices are able to predict vegetation cover

in our study, as more of areas are covered with herbaceous plants and shrub land (Figure 3.1).

Eight different vegetation indices namely NDVI, GNDVI, RENDVI, REMSR, OSAVI,

MSAVI2, WDRVI, ARVI and five different bands from original images were used for modeling.

The selections of these variables were based on their ability to map vegetation cover at different

settings and are among widely used indices for mapping. The significant and application of

70

different vegetation indices can be found on review article named significant remote sensing

vegetation indices from Xue et al. (2017). The limitation in selection of variables was the

original image bands.

We found that the disturbance intensity was high where the intensity of training was high

in central region of FR, where three training areas were located and their UAV images were

taken. Our finding is consistent with that from Howard et al (2013), Wang et al (2014) and Rijal

et al (2017), that the western and central regions of FR were areas with high training intensity

that led to corresponding high vegetation cover disturbance. This was observed from the

decreased mean values of VIs after the military training for each box. The mean value of each of

the calculated vegetation indices, except GNDVI, decreased after the military training activities,

concluding that the vegetation cover was impacted by the military training activities. The mean

values of VIs at the 5m spatial resolution were not very much different from those at the original

resolution. We observed a very small variation of change in VIs for box 2 before disturbance at

the 5m resolution, but the change was more visible for 10m resolution. This might have occurred

as a result of improper handling of flight plan and flight box. The problem was observed on the

bottom left part of the study box, where the flight box and original image are identical, thus there

are not enough pixels to resample using scaling factor. The problem might have arisen also due

to the edge effect from the resampling technique.

It was found that the VIs were highly correlated with each other, which led to high

multicollinearity, but had relatively lower correlation with the original bands. Thus, in the final

models, very often a VI and an original band were selected. The widely selected VIs in the final

models included REMSR, NDVI and GNDVI due to their significant and high correlation with

vegetation cover.

71

The intensity and disturbance severity of military training activities are different for

different areas across FR. Using a global model with the same regression coefficients may lack

the ability of modelling spatiotemporal variable military training intensity and corresponding

disturbances, thus could lead to larger errors (Rijal 2017). This finding was supported by the

results of this study in which it was noticed that regardless of spatial resolutions, separately

modelling the relationship of vegetation cover with the spectral variables led to higher accuracy

of both modelling and predictions of vegetation cover. The overall pooled dataset only

performed better than the dataset of box 3 at the spatial resolution of 10m by 10 m.

Moreover, the accuracy of modelling and prediction was higher before the training than

that after the training. This was mainly because the military training activities led to the decrease

of vegetation cover and thus increased the impact of spectral reflectance from soils on the VIs.

Another reason might be because the military training activities led to high variation of

vegetation cover within the training boxes. In addition, in this study the results of modelling and

predicting vegetation cover from 5 m by 5 m and 10 m by 10 m plot sizes were compared. It was

found that overall, the 5 m spatial resolution resulted in more accurate predictions, but the

differences of accuracy were not significant. The prediction accuracy, to some extent, varied

depending on the training boxes, plot sizes, and before and after the training. However, the

spatial resolution of 10 m by 10 m led to a limited number of sample plots, which may become a

problem to model the relationship of vegetation cover with spectral variables (Miao et al. 2020).

There is also the potential of using the datasets from the sub-plots of 1 m by 1 m. However, due

to limited time we did not complete the comparison, which will be the future work.

The use of satellite remote sensing to assess military land conditions at FR had been

previously studied by Fang et al (2002), Howard et al (2013), Wang et al (2014), and Rjial

72

(2017). This study to assess military training-induced disturbance cover mapping using UAV

imagery is a pioneer study at FR Army Installation. This study resampled the UAV images to the

spatial resolutions of field data to make consistency between field data and UAV images. Dash et

al (2018) resampled the original UAV image pixels from MicaSense RedEdge-3 to 5m spatial

resolution to compare the results with satellite data to monitor forest health. The results suggest

that the resampled UAV data is more sensitive than the finer spatial resolution to detect the stress

in vegetation down to individual trees. In this study, we did not make the comparison between

UAV and satellite data. We will do it in the near future.

One of the major application limitations of UAV imagery is that the area of interest

should be small, as we can see that the sizes of flight boxes are small compared to the whole

Military Installation. However, it is possible to map heavily disturbed areas if the military

training induced disturbances are limited to small patches. Additionally, the sample points where

vegetation cover was measured using Daubenmire plots were not consistent between flights and

resolution (5m and 10 m). Deriving vegetation cover data from 12 Daubenmire frame plots for

each box might have caused some uncertainties. For example, the results for the mean value of

vegetation cover for box 2 after the disturbance is greater than the mean value before the

disturbances. It is recommended to collect data according to the level of disturbances that

occurred in the way to match the disturbance level and plot locations should be consistent for

both before and after the disturbance. It is recommended to use the same flight height between

flights so that consistent pixel resolution is obtained. Another recommendation is the proper

planning designation or flight box and collecting data to match the spatial resolution of UAV

imagery. For example, in box 2 before the training, there were an insufficient number of pixels

on the edges of the lower left side which caused the mean value to decrease for the overall box.

73

5.2 Conclusions

Overall, this study enhanced the understanding of using UAV images to map vegetation

cover change from military training activities and led to following conclusions:

1) UAV images can be used to map vegetation cover and disturbances induced by military

training activities at FR Army Installations and similar settings. The use of different VIs and

original bands from the images were capable to capture spatial variability of vegetation cover for

each box and for overall pooled data.

2) Developing vegetation cover model separately for the training boxes led to more accurate

predictions of vegetation cover compared to using the pooled dataset for both 5 m and 10 m

spatial resolutions, which implied that the military training induced disturbances vary spatially.

This conclusion can also be observed on the descriptive statistics from vegetation cover data

collected using Daubenmire frame plots, where the standard deviation of cover data increased

after military training activities.

3) The accuracy of modelling was also higher before the training than that after the training

because the training activities increased the impact of spectral reflectance from soils on the bands

and their VIs and also led to higher spatial variation of vegetation cover.

4) The 5 m by 5 m spatial resolution images were more capable in capturing spatial variation of

the vegetation disturbances than those at 10 m by 10 m spatial resolution. The 10 m by 10 m plot

datasets have limited sample points and were unable to capture variations after military training

activities.

5) Compared with the original UAV image bands, the VIs improved the correlation with

vegetation cover, and the REMSR, NDVI and GNDVI were more frequently selected in the final

models than other VIs.

74

7. REFERENCES

Agisoft. ―Agisoft PhotoScan User Manual: Professional Edition, Version 1.4‖.

http://www.agisoft.com/pdf/photoscan-pro_1_4_en.pdf, 2018, assessed on April, 2021.

Althoff, Donald P., James W. Rivers, Jeffrey S. Pontius, Philip S. Gipson, and Philip B.

Woodford. "A comprehensive approach to identifying monitoring priorities of small landbirds on

military installations." Environmental Management 34, no. 6 (2004): 887-902.

Althoff, Peggy S., and Stephen J. Thien. "Impact of M1A1 main battle tank disturbance on soil

quality, invertebrates, and vegetation characteristics." Journal of Terramechanics 42, no. 3-4

(2005): 159-176.

Althoff, Donald P., P. S. Althoff, N. D. Lambrecht, P. S. Gipson, J. S. Pontius, and P. B.

Woodford. "Soil properties and perceived disturbance of grasslands subjected to mechanized

military training: evaluation of an index." Land Degradation and Development 18, no. 3 (2007):

269-288.

Anderson, Alan B., Patrick J. Guertin, and David L. Price. Land Condition Trend Analysis Data:

Power Analysis. No. CERL-TR-97/05. CONSTRUCTION ENGINEERING RESEARCH LAB

(ARMY) CHAMPAIGN IL, 1996.

Anderson, Alan B., Antonio J. Palazzo, Paul D. Ayers, Jeffrey S. Fehmi, Sally Shoop, and

Patricia Sullivan. "Assessing the impacts of military vehicle traffic on natural areas. Introduction

to the special issue and review of the relevant military vehicle impact literature." Journal of

Terramechanics 42, no. 3-4 (2005): 143-158.

Anderson, Alan B., Guangxing Wang, Shoufan Fang, George Z. Gertner, Burak Güneralp, and

Don Jones. "Assessing and predicting changes in vegetation cover associated with military land

use activities using field monitoring data at Fort Hood, Texas." Journal of Terramechanics 42,

no. 3-4 (2005): 207-229.

Ayers, Paul D., Alan B. Anderson, and Chunxia Wu. "Analysis of vehicle use patterns during

field training exercises to identify potential roads." Journal of Terramechanics 42, no. 3-4

(2005): 321-338.

Bailey, Robert G. "Ecoregions of the United States (map)." Ogden, UT: US Department of

Agriculture, US Forest Service, Intermountain Region (1976).

Barnston, A. "Correspondence among the Correlation (root mean square error) and Heidke

verification measures; refinement of the Heidke score. Notes and Correspondence. American

Meteorological Society Journal, 699–709." (1992).

Barron, Douglas G., Jeffrey D. Brawn, Luke K. Butler, L. Michael Romero, and Patrick J.

Weatherhead. "Effects of military activity on breeding birds." The Journal of Wildlife

Management 76, no. 5 (2012): 911-918.

http://www.agisoft.com/pdf/photoscan-pro_1_4_en.pdf

75

Bertacchi, Andrea, Vittoria Giannini, Carmelo Di Franco, and Nicola Silvestri. "Using

unmanned aerial vehicles for vegetation mapping and identification of botanical species in

wetlands." Landscape and Ecological Engineering 15, no. 2 (2019): 231-240.

Colomina, Ismael, and Pere Molina. "Unmanned aerial systems for photogrammetry and remote

sensing: A review." ISPRS Journal of Photogrammetry and Remote Sensing 92 (2014): 79-97.

Cramer, Michael, Stéphane Bovet, Manfred Gültlinger, Eija Honkavaara, Andy McGill, Martijn

Rijsdijk, Mark Tabor, and Vincent Tournadre. "On the use of RPAS in national mapping—The

EUROSDR point of view." Int. Arch. Photogram. Remote Sens. Spat. Inf. Sci (2013): 93-99.

Dale, Virginia, Daniel L. Druckenbrod, Latha Baskaran, Matthew Aldridge, Michael Berry,

Chuck Garten, Lisa Olsen, Rebecca Efroymson, and Robert Washington-Allen. "Vehicle impacts

on the environment at different spatial scales: observations in west central Georgia,

USA." Journal of Terramechanics 42, no. 3-4 (2005): 383-402.

Delaney, David K., Larry L. Pater, Lawrence D. Carlile, Eric W. Spadgenske, Timothy A. Beaty,

and Robert H. Melton. "Response of red‐cockaded woodpeckers to military training operations:

La reponse des pics à face blanche aux exercices d'Entraînement militaire." Wildlife

Monographs 177, no. 1 (2011): 1-38.

Everaerts, J., N. Lewyckyj, and D. Fransaer. "Pegasus: design of a stratospheric long endurance

UAV system for remote sensing." The International Archives of the Photogrammetry, Remote

Sensing and Spatial Information Sciences 35, no. Part B (2004).

Gertner, George, Guangxing Wang, Shoufan Fang, and Alan B. Anderson. "Mapping and

uncertainty of predictions based on multiple primary variables from joint co-simulation with

Landsat TM image and polynomial regression." Remote Sensing of Environment 83, no. 3

(2002): 498-510.

Gertner, George, Guangxing Wang, and Alan B. Anderson. "Determination of frequency for

remeasuring ground and vegetation cover factor needed for soil erosion

modeling." Environmental Management 37, no. 1 (2006): 84-97.

Gitelson, Anatoly A. "Wide dynamic range vegetation index for remote quantification of

biophysical characteristics of vegetation." Journal of Plant Physiology 161, no. 2 (2004): 165-

173.

Gitelson, Anatoly A., and Mark N. Merzlyak. "Remote sensing of chlorophyll concentration in

higher plant leaves." Advances in Space Research 22, no. 5 (1998): 689-692.

Haala, Norbert, Michael Cramer, and Mathias Rothermel. "Quality of 3D point clouds from

highly overlapping UAV imagery." Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci (2013):

183-188

76

Hansen, Dennis J., and W. Kent Ostler. "Assessment technique for evaluating military vehicular

impacts to vegetation in the Mojave Desert." Journal of terramechanics 42, no. 3-4 (2005): 193-

205.

HEXAGON, Raster Generalization, 2021,

https://geospatialcommunity.hexagon.com/s/article/Raster-Generalization-Degrade-

Downsample-etc assessed on April, 2021.

Howard, Heidi R., Guangxing Wang, Steve Singer, and Alan B. Anderson. "Modeling and

prediction of land condition for Fort Riley military installation." Transactions of the ASABE 56,

no. 2 (2013): 643-652.

Henebry, Geoffrey M., Andrés Viña, and Anatoly A. Gitelson. "The wide dynamic range

vegetation index and its potential utility for gap analysis." DigitalCommons, University of

Nebraska - Lincoln, (2004).

Hutchinson, J. M. S., A. Jacquin, S. L. Hutchinson, and J. Verbesselt. "Monitoring vegetation

change and dynamics on US Army training lands using satellite image time series

analysis." Journal of Environmental Management 150 (2015): 355-366.

Johnson, Sara, Guangxing Wang, Heidi Howard, and Alan B. Anderson. "Identification of

superfluous roads in terms of sustainable military land carrying capacity and

environment." Journal of Terramechanics 48, no. 2 (2011): 97-104.

Jones, R., D. Horner, P. Sullivan, and R. Ahlvin. "A methodology for quantitatively assessing

vehicular rutting on terrains." Journal of Terramechanics 42, no. 3-4 (2005): 245-257.

Kaufman, Yoram J., and Didier Tanre. "Atmospherically resistant vegetation index (ARVI) for

EOS-MODIS." IEEE Transactions on Geoscience and Remote Sensing 30, no. 2 (1992): 261-

270.

Koch, Daniel J., Paul D. Ayers, Heidi R. Howard, and Gary Siebert. Vehicle Dynamics

Monitoring and Tracking System (VDMTS): Monitoring Mission Impacts in Support of

Installation Land Management. ARMY RESEARCH LAB WHITE SANDS MISSILE RANGE

NM COMPUTATIONAL AND INFORMATION SCIENCES DIRECTORATE, 2012.

Leis, Sherry A., David M. Engle, David M. Leslie, and Jeffrey S. Fehmi. "Effects of short-and

long-term disturbance resulting from military maneuvers on vegetation and soils in a mixed

prairie area." Environmental Management 36, no. 6 (2005): 849-861.

Liu, Zhixiang, Youmin Zhang, Xiang Yu, and Chi Yuan. "Unmanned surface vehicles: An

overview of developments and challenges." Annual Reviews in Control 41 (2016): 71-93.

Mayer, Cedric, LM Gomes Pereira, and Thomas P. Kersten. "A comprehensive workflow to

process UAV images for the efficient production of accurate geo-information." In Proceedings of

the IX National Conference on Cartography and Geodesy, Amadora, Portugal, pp. 25-26. 2018.

https://geospatialcommunity.hexagon.com/s/article/Raster-Generalization-Degrade-Downsample-etc

https://geospatialcommunity.hexagon.com/s/article/Raster-Generalization-Degrade-Downsample-etc

https://digitalcommons.unl.edu/

https://digitalcommons.unl.edu/

77

Micasense. 2017. MicaSense RedEdge TM 3 Multispectral Camera User Manual: 1–27

Micasense, 2021. https://support.micasense.com/hc/en-us/articles/115000351194-RedEdge-

Camera-Radiometric-Calibration-Model assesse on February, 2021.

Palazzo, Antonio J., Kevin B. Jensen, Blair L. Waldron, and Timothy J. Cary. "Effects of tank

tracking on range grasses." Journal of Terramechanics 42, no. 3-4 (2005): 177-191

Quist, M.C., P.A. Fay, C.S. Guy, A.K. Knapp, and B.N. Rubenstein. 2003. Military training

effects on terrestrial and aquatic communities on a grassland military installation. Ecological

Applications 13(2): 432-442.

Qi, Jiaguo, Abdelghani Chehbouni, Alfredo R. Huete, Yann H. Kerr, and Soroosh Sorooshian.

"A modified soil adjusted vegetation index." Remote sensing of Environment 48, no. 2 (1994):

119-126.

Rijal, S., G. Wang, P. B. Woodford, H. R. Howard, J. Schoof, T. J. Oyana, L. O. Park, and R. Li.

"Comparison of military and nonmilitary land condition using an image derived soil erosion

cover factor." Journal of Soil and Water Conservation 72, no. 5 (2017): 425-437.

Raper, R. L. "Agricultural traffic impacts on soil." Journal of Terramechanics 42, no. 3-4

(2005): 259-280.

Remondino, Fabio, Silvio Del Pizzo, Thomas P. Kersten, and Salvatore Troisi. "Low-cost and

open-source solutions for automated image orientation–A critical overview." In Euro-

Mediterranean Conference, pp. 40-54. Springer, Berlin, Heidelberg, 2012.

Rondeaux, Geneviève, Michael Steven, and Frédéric Baret. "Optimization of soil-adjusted

vegetation indices." Remote Sensing of Environment 55, no. 2 (1996): 95-107.

Rosnell, Tomi, and Eija Honkavaara. "Point cloud generation from aerial image data acquired by

a quadrocopter type micro unmanned aerial vehicle and a digital still camera." Sensors 12, no. 1

(2012): 453-480.

Rouse, J. W., Rüdiger H. Haas, John A. Schell, and Donald W. Deering. "Monitoring vegetation

systems in the Great Plains with ERTS." NASA Special Publication 351, no. 1974 (1974): 309.

Shoop, S., R. Affleck, C. Collins, G. Larsen, L. Barna, and P. Sullivan. "Maneuver analysis

methodology to predict vehicle impacts on training lands." Journal of Terramechanics 42, no. 3-

4 (2005): 281-303.

Salamí, Esther, Cristina Barrado, and Enric Pastor. "UAV flight experiments applied to the

remote sensing of vegetated areas." Remote Sensing 6, no. 11 (2014): 11051-11081.

https://support.micasense.com/hc/en-us/articles/115000351194-RedEdge-Camera-Radiometric-Calibration-Model

https://support.micasense.com/hc/en-us/articles/115000351194-RedEdge-Camera-Radiometric-Calibration-Model

78

Senseman, Gary M., Scott A. Tweddale, Alan B. Anderson, and Calvin F. Bagley. Correlation of

Land Condition Trend Analysis (LCTA) Rangeland Cover Measures to Satellite-Imagery-

Derived Vegetation Indices. No. CERL-TR-97/07. CONSTRUCTION ENGINEERING

RESEARCH LAB (ARMY) CHAMPAIGN IL, 1996.

Sims, Daniel A., and John A. Gamon. "Relationships between leaf pigment content and spectral

reflectance across a wide range of species, leaf structures and developmental stages." Remote

Sensing of Environment 81, no. 2-3 (2002): 337-354.

Singer, Steve, Guangxing Wang, Heidi Howard, and Alan Anderson. "Environmental condition

assessment of US Military Installations Using GIS based spatial multi-criteria decision

analysis." Environmental Management 50, no. 2 (2012): 329-340.

Tagle Casapia, Maria Ximena. "Study of radiometric variations in Unmanned Aerial Vehicle

remote sensing imagery for vegetation mapping." Lund University GEM thesis series (2017).

Tweddale, Scott. Historical analysis of land cover/condition trends at Fort Bliss, Texas, using

remotely sensed imagery. No. ERDC/CERL-TR-01-36. ENGINEER RESEARCH AND

DEVELOPMENT CENTER CHAMPAIGN IL CONSTRUCTION ENGINEERING

RESEARCH LAB, 2001.

US Army. 1994. Integrated natural resource management plan for Fort Riley, Kansas.

Directorate of Engineering and Housing, Environmental and Natural Resources Division. Berger

and Associates, Chicago, Illinois, USA.

Van de Voorde, Tim, Jeroen Vlaeminck, and Frank Canters. "Comparing different approaches

for mapping urban vegetation cover from Landsat ETM+ data: a case study on

Brussels." Sensors 8, no. 6 (2008): 3880-3902.

Vehtari, Aki, Daniel P. Simpson, Yuling Yao, and Andrew Gelman. "Limitations of ―Limitations

of Bayesian leave-one-out cross-validation for model selection‖." Computational Brain &

Behavior 2, no. 1 (2019): 22-27.

Wang, Guangxing, G. Gertner, and Alan B. Anderson. "Sampling design and uncertainty based

on spatial variability of spectral variables for mapping vegetation cover." International Journal of

Remote Sensing 26, no. 15 (2005): 3255-3274.

Wang, Guangxing, George Gertner, and Alan B. Anderson. "Sampling and mapping a soil

erosion cover factor by integrating stratification, model updating and cokriging with

images." Environmental Management 39, no. 1 (2007): 84-97.

Wang, Guangxing, G. Gertner, Alan B. Anderson, H. Howard, D. Gebhart, D. Althoff, T. Davis,

and P. Woodford. "Spatial variability and temporal dynamics analysis of soil erosion due to

military land use activities: uncertainty and implications for land management." Land

Degradation and Development 18, no. 5 (2007): 519-542.

79

Wang, Guangxing, Tonny Oyana, Maozhen Zhang, Samuel Adu-Prah, Siqi Zeng, Hui Lin, and

Jiyun Se. "Mapping and spatial uncertainty analysis of forest vegetation carbon by combining

national forest inventory data and satellite images." Forest Ecology and Management 258, no. 7

(2009): 1275-1283.

Wang, Guangxing, Dana Murphy, Adam Oller, Heidi R. Howard, Alan B. Anderson, Santosh

Rijal, Natalie R. Myers, and Philip Woodford. "Spatial and temporal assessment of cumulative

disturbance impacts due to military training, burning, haying, and their interactions on land

condition of Fort Riley." Environmental Management 54, no. 1 (2014): 51-66.

Wilson, Scott D. "The effects of military tank traffic on prairie: a management

model." Environmental Management 12, no. 3 (1988): 397-403.

Wu, Chaoyang, Zheng Niu, Quan Tang, and Wenjiang Huang. "Estimating chlorophyll content

from hyperspectral vegetation indices: Modeling and validation." Agricultural and forest

meteorology 148, no. 8-9 (2008): 1230-1241.

80

8. VITA

Graduate School

Southern Illinois University

Bibek Ban

[email protected]

Tribhuvan University, Nepal

Bachelor of Science, Forestry, April 2014

Thesis Paper Title:



Major Professor: Dr. Guangxing Wang

MAPPING VEGETATION COVER AND DISTURBANCE BASED …

Documents