Sensitivity of a large-scale hydrologic model to quality of input data obtained at different scales; distributed versus stochastic non-distributed modelling

Sensitivity of a large-scale hydrologic model to quality of input

data obtained at different scales; distributed versus stochastic

non-distributed modelling

Georges Vachaud*, Tao Chen

Laboratoire d’etude des Transferts en Hydrologie et Environnement (LTHE, UMR 5564 CNRS, Univ. J. Fourier Grenoble1, INPG, IRD),

BP 53, F-38041 Grenoble Cedex 09, France

Received 22 May 2001; revised 12 March 2002; accepted 15 March 2002

Abstract

The amount of information available to run a spatially distributed model is often very much less than the ideal. The aim of

this study is to estimate the impact of degradation of information about the spatial distribution of input parameters, and second

the scale at which this information is obtained.

A well defined agricultural catchment, with an important database concerning spatial and temporal observations has been

used for this purpose: the agricultural catchment of la Cote St Andre, 60 km North East of Grenoble in the South East of France.

The methodological framework is based on the Areal Non-point Source Watershed Environmental Response Simulation model.

A 3 year simulation using georeferenced variables (crops and soil types) and annual changes in crop rotation is first

developed as a reference. This is compared to simulations results obtained during the same period, with the same climatic data,

but with the following degradation of quality of other inputs: firstly, the spatial distribution of soils and crop is ignored; both

variables being defined by their areal coverage obtained from local information; secondly, the same inputs are deduced from a

database obtained at the European scale. In both cases, a Latin Hypercube Sampler is used to stochastically generate sets of

samples corresponding to the probability distribution of variables. The study is based on comparisons between modelled

outputs: drainage of water and leaching of nitrate below the root zone of crops at the catchment scale.

When information is local, and in absence of lateral flow (runoff), distributed modelling and purely stochastic modelling

provide identical catchment average values; on the contrary, the use of the European database may introduce important biases

concerning the proportion of land uses and of soils. In both cases, however, the lack of information concerning the location of

sensitive areas in terms of risks of pollution may be considered as an important weakness of stochastic models.

This work was done in the frame of ‘CAMSCALE—Upscaling predictive models and catchment water quality’ a European

Union DGXII—Environment funded programme co-ordinated by SSLRC, Silsoe, UK. q 2002 Elsevier Science B.V. All rights

reserved.

Keywords: Distributed modelling; Stochastic modelling; Latin Hypercube Sampler; Nitrate leaching; Groundwater recharge; Catchment

1. Introduction

Most detailed research models are developed from

an examination of environmental processes at the

0022-1694/02/$ - see front matter q 2002 Elsevier Science B.V. All rights reserved.

PII: S0 02 2 -1 69 4 (0 2) 00 0 69 -0

Journal of Hydrology 264 (2002) 101–112

www.elsevier.com/locate/jhydrol

* Corresponding author.

E-mail address: [email protected] (G. Vachaud).

point scale or over very small areas, often over very

short time intervals. Moving from this purely research

position to an applied or policy-driven approach

involving large areas of land and time-scales of

months or years means changing the scale of

operation of the models. This, in turn, almost

invariably means that the amount of information

available to run the model is very much less than the

ideal. Improvement of the understanding of the effect

of changes of spatial and temporal scale within input

data to process-based models, as reflected in com-

parisons between modelled output and concentrations

of reference substances, such as nitrate nitrogen in

subsurface water at the catchment scale, was the

principal objective of the CAMSCALE project

(upscaling, predictive models and catchment water

quality), a 3-year multidisciplinary research project

funded by the Commission of the European Commu-

nities under the Framework IV Programme (Climate

and Environment—DGXII). It is in the framework of

this project that the results reported here were

obtained. More complete information concerning

objectives of the project, partnership, results and

overall conclusions can be found at the project web

site.1

When insufficient measured data are available,

other less precise sources of data need to be used, with

a consequent impact on the certainty of model

predictions (Binley et al., 1991). Variability with

respect to model input data is recognised as a

potentially significant source of uncertainty in model

predictions (Refsgaard et al., 1999; Finke et al., 1996;

Freissinet et al., 1999). Besides the effect of the

measurement technique itself (Davis et al., 1999;

Hack-ten-Broek and Hegmans, 1996) the source of

the uncertainty can arise from two aspects. The first

is the variability inherent within the parameters, while

the second is a function of the scale or resolution of

the data source.

This is the second of two papers addressing these

issues. The first article dealing with the sensitivity of

model outputs to variability of soil parameters within

a soil class (Vachaud and Chen, 2002), and introdu-

cing the methodological framework was based on the

concept of Areal Non-point Source Watershed

Environmental Response Simulation (ANSWERS,

Beasley et al., 1980; Bouraoui et al., 1997a), coupled

with a Latin Hypercube Sampler (LHS), to stochas-

tically generate a set of samples within a soil class.

Transport parameters were then determined for each

sample with the use of pedotransfer functions (Rawls

and Brakensiek, 1989), and the results obtained from

200 realisations compared to those of a single

realisation corresponding to aggregation of the soil

class to its barycentre. For Silt, Loam, Silt loam and

coarser grained classes, it was shown that ANSWERS

is insensitive to this intraclass variability under

appropriate rainfall regimes.

In this paper, the same deterministic model

framework will be used to investigate two further

points:

† The effect of the loss of information about the

spatial distribution of soils and crops within a

catchment. ANSWERS will be used either with a

georeferenced database and coupled with a Geo-

graphical Information System to obtain distributed

modelling, or within a purely stochastic mode in

which inputs variables concerning soils and crops

over the catchment are obtained with a LHS

generator to satisfy areal coverage values given by

local observations.

† The effect of the loss of information about areal

coverage values, when upscaling the source of

information from local to European level, using the

same methodological framework.

A well defined agricultural catchment, with an

important database concerning spatial and temporal

observations, has been used for this exercise: the

agricultural catchment of la Cote St Andre, 60 km

North East of Grenoble in the South East of France.

2. Study catchment and data source

2.1. The catchment

This catchment has already been described in

detail by Bouraoui et al. (1997a,b). With a total

area of 180 km2, its major part (about 80%) is a

flat plain, surrounded by hills. The soil is shallow

(average thickness of 1 m) and rich in organic

1 http://www.silsoe.cranfield.ac.uk/sslrc/ourorg/projects/

camscale/.

G. Vachaud, T. Chen / Journal of Hydrology 264 (2002) 101–112102

matter; it is mostly of sandy loam texture with

high stone contents becoming progressively

coarser with increasing depth. Below, there is a

very coarse glacial deposit extending to about 20–

30 m, with a water table aquifer at a depth varying

around 10 m below the soil surface. It is

essentially an agricultural area partitioned between

irrigated crops (mostly maize and tobacco), dry

farming crops (maize, wheat and sunflower), and

prairies. An intensive interdisciplinary study was

initiated in 1991, aiming to optimise agricultural

practices, and to develop sustainable management

schemes with respect to important threats on

aquifer non-point source pollution. Two different

scales were investigated. First, at the local scale,

with field experiments carried out during three

successive years (1991–1993) on three types of

soil cover (maize, grass and bare soil) with

different levels of fertilisation. Their aim was

mainly to characterise the terms of the water

balance and the leaching of nitrate below the root

zone, (Normand et al., 1997). Second, at the

catchment scale, there were two aims: to describe

the water table aquifer behaviour (including long-

term monitoring in several wells of water table

depth and nitrate concentration); and to study

agricultural practices and land use management for

a large number of farms in order to obtain spatial

estimates of yield, water consumption, fertilisation

and agricultural costs (Bel et al., 1999). The first

level of investigation (local) has been used

(Vachaud and Chen, 2002) to validate the model-

ling approach. This study deals with the second

level of investigation, the catchment scale.

2.2. The sources of data

Soil data. Within the catchment and the surround-

ing region, soil parameters (soil type, number of

layers and depths, sand, clay and silt content, organic

matter content, physico-chemical parameters) were

obtained from measurements carried out on 138

sampling pits. These were augmented by a digitised

soil map of the area at the scale of 1:250,000

reclassified into nine types of soil subclasses, follow-

ing the French soil classification, and by tabulated

values of silt, clay, sand, Cation Exchange Capacity

(CEC), and organic matter content values. The

dominant subclass (66% of the catchment) is

‘Alluvions recents de basse terrasse’ (Table 1). The

classification can also be compared with the USDA

texture class system yielding three large classes of

soils: ‘Loam’, ‘Silt loam’ and ‘Sandy loam’ (Table 1).

Altitude. The whole catchment was digitised in a

Digital Terrain Model with a spatial resolution of

50 m and a vertical precision of 1 m. The elevation is

around 500 m, with an average slope of 1%.

Land use data. Two different types of observation

were obtained:

In 1991, an intensive survey, at the scale of the

whole catchment, provided a land use map

representative of the agricultural practices. Data

were classified into six types of cover: natural

prairies (18%); irrigated crop—essentially maize

(10%); dry cultivation crops: maize (13%), winter

wheat (34%), sunflower (15%) and fallow (10%)

(Fig. 1(a)).

Interannual observations in representative farms

Table 1

Soil units on the la Cote St Andre catchment

French soil classification subclass % Coverage USDA soil texture class % Coverage

Colluvions limoneuses anciennes 2.06 Silt loam 25

Placages limoneux moyennement differencies 1.74

Placages limoneux differencies hydromorphes 13.45

Basses terrases 66.15 Loam 73

Placages limoneux differencies 13.45

Moraines rissiennes 5.24

Conglomerat en pente forte 0.97

Colluviums et cones de dejections anciens 7.77

Moraines des pentes fortes 0.70 Sandy loam 2

G. Vachaud, T. Chen / Journal of Hydrology 264 (2002) 101–112 103

identified temporal changes of land use distribution

resulting from crop rotations. The most important

points concerning agricultural practices are: the

period of rotation cycles is usually 2 or 3 years;

irrigated crops are normally followed by irrigated

crops (maize followed by maize, with bare soil in

winter, is predominant), a consequence of the

existence of underground networks to supply

irrigation; prairies are permanent; interannual

rotation is thus largely confined to dry cultivation

crops and fallow.

Crop successions are summarised in Table 2; each

line indicates, for a given crop, the percentage of

surface, which will be cultivated the following year

with a crop identified in columns headings. As an

example, if we consider fields cultivated with wheat

on year n, 23% will be again with wheat on year

n þ 1; whereas 16% will be cropped with rainfed

maize and 35% with sunflower, while 26% will be

fallow.

Weather data. Meteorological data are available

from the Airport of Grenoble, in the middle of the

catchment. Daily values (rainfall, wind, air and soil

temperature, Potential Evapotranspiration—PET)

over 16 successive years are available. The average

annual rainfall is around 1000 mm, with a very high

interannual variability (^300 mm). Mean evapotran-

spiration computed by the Penman Monteith approach

is around 850 mm per year.

3. Spatially distributed multiannual modelling

The first stage was to generate a reference study.

This was done by simulating the hydrologic behaviour

of the catchment over a number of years, and the

amount of nitrate leaching below the root zone of

crops, making use of the largest possible source of

information available on this catchment.

3.1. Modelling tool

This study uses the ANSWERS model in the

version developed by Bouraoui (1994), Bouraoui et al.

(1997a,b) as a watershed scale model with distributed

parameters to be used to simulate on a continuous

daily time-scale water balance, aquifer recharge and

non-point source pollution (leaching of N-nitrate). A

characteristics of this model is the use of model

parameter estimation routines related to easily

obtainable properties, and/or the parameterisation of

several processes with easily obtainable parameters or

large databases. In particular, the soil transport

parameters are obtained from a set of pedotransfer

functions (Tietje, 1999) on the basis of soil texture,

Fig. 1. The reference case study, year 1991; annual rainfall:

767 mm. (a) Land use map of the catchment. (b) Distribution of

annual drainage (cumulative water loss, in mm, from January 1 to

December 31, 1991); catchment average: 294 ^ 52 mm. (c)

Distribution of annual leaching (cumulative nitrate leaching, in

KgN·NO23/ha, from January 1 to December 31, 1991); catchment

average: 82 ^ 45 KgN·NO23/ha.

G. Vachaud, T. Chen / Journal of Hydrology 264 (2002) 101–112104

organic matter and CEC (Rawls and Brakensiek,

1989); the parameters describing plant behaviour are

directly obtained from a database (Knissel, 1993).

The model uses a 1D vertical description of

water transport and nitrogen transformation in the

root zone, whereas at the soil surface it behaves as a

2D model for runoff simulation. When information

concerning the spatial distribution of input par-

ameters is available, ANSWERS can be used as a

distributed hydrologic model at a catchment scale.

The catchment is discretised into square elements,

each one with a uniform set of topographic, soil

hydrodynamic and crop parameters. For each model

element (block), the water balance equation yields

an estimate of actual evapotranspiration, runoff and

drainage (defined as the water percolating down-

wards from the root zone). Draining water transports

excess of nitrate from the root zone and produces

the nitrate leaching. In order to allow automation of

the input file creation (soil and plant parameters),

easy modification and manipulation of the cell size

and of the georeferenced data, and visualisation of

input and output data, this model is coupled with a

raster based Geographic Information System: Geo-

graphical Resources Analysis and Support System

(GRASS; US Army Corps of Engineer, Clamons and

Byars, 1997). The spatially distributed information

concerning digital terrain elevation, soil, and land use

constitutes three layers of input data, with its third layer

(land use) changing every year; two additional layers are

defined for visualisation of output data, one concerning

the drainage of water and the other the leaching of

nitrate. In both cases, this is done for a reference level

selected at a depth of 1 m below the soil surface, the

deepest observation of roots in this field.

3.2. Application and results

A grid of about 4400 blocks (each one

representing about 4 ha) was used to discretise

this catchment. In association with the GIS, every

block was identified by:

† Spatial co-ordinates,

† Soil class (from which were inferred automati-

cally the value of bulk density, hydrodynamic

parameters of the soils (saturated hydraulic

conductivity and suction at the wetting front),

and parameters concerning the nitrogen cycle via

pedotransfer function with the use of pedotrans-

fer functions. Since Vachaud and Chen (2002)

showed the absence of sensitivity of this model

outputs to the intraclass variability of transport

parameters for the soil classes describing this

catchment, each class of soil was defined by a

single set of textural parameters (sand, silt and

clay contents obtained at the barycentre of the

class),

† Land use, from which plant uptake and plant

growth were inferred automatically. Leaf Index

Area and root depth at different phenological

periods, were obtained via the plant database.

The land use map for 1992 and 1993 were

generated from that obtained in 1991 with a

stochastic generator. The plots of land with

irrigated crop or prairies remain identical to

those observed in 1991 (no crop rotation for

those two covers). In contrast, the plots with

fallow or rainfed crops were stochastically

redistributed within the catchment from year to

year to satisfy the crop rotation index given in

Table 2

Crop rotation index

Land use Irrigated maize Rainfed maize Wheat Sunflower Prairie Fallow

Irrigated maize 100 0 0 0 0 0

Rainfed maize 0 48 22 25 0 5

Wheat 0 16 23 35 0 26

Sunflower 0 17 76 0 0 7

Prairie 0 0 0 0 100 0

Fallow 0 27 8 22 0 43

For every crop (row), the number in columns represents the percentage of the surface cultivated in the following year with the crop identified

in column heading.

G. Vachaud, T. Chen / Journal of Hydrology 264 (2002) 101–112 105

Table 2 and, at the same time, to represent the

same global percentage as observed in 1991.

Planting and harvesting dates, as well as dates

of application of fertiliser for the crops are given

in Table 3. They represent usual farming

practises in the catchment. Fallow corresponds

to bare soil.

Actual evapotranspiration, drainage, runoff, nitro-

gen uptake and nitrate leaching were simulated for

every individual block during the period 1991–1993.

Daily values of PET, rainfall and irrigation (for

maize) were recorded during the 3 year intensive

experiment during the same period were used as input

data; annual values are given in Table 4. The complete

simulation (3 years, daily time step, 4400 blocks)

takes approximately 2 h with a usual 700 MHz PC

computer.

Results obtained for every block were stored and

maps created by the Geographic Information System

to express the spatial distribution of selected outputs.

An important result is that no runoff was ever obtained

during the simulation, nor was any runoff observed

over the field, a result of the combination of highly

permeable soil and absence of slope. An example of

yearly cumulative values of drainage of water and

leaching of nitrate is given in Fig. 1(b) and (c) for the

first year (1991). These maps clearly identify the

localised ‘hot spots’, and illustrate the large varia-

bility, at the yearly scale, of water drainage and of

nitrate leaching depending upon the combination type

of soil/type of crop. Quite clearly the amount of

drainage is very high under irrigated maize, a

consequence of excessive irrigation added to the fact

that this cultivation is usually followed by bare soil in

winter. It is also clear, on Fig. 1(c), that for fallow, the

amount of annual nitrate leaching can be in the range

150–200 kgN/ha, a number that may raise doubts

concerning the sustainability of such a practise. These

results are confirmed by observation obtained on

experimental plots on this catchment (Normand et al.,

1997).

It is also possible to obtain statistics on these

outputs for either blocks cultivated with a given crop,

or the whole catchment (Table 5). These show that the

most important factor affecting water drainage and

nitrate leaching below the root zone is the annual

rainfall (amount and distribution). This is quite clear

for crops under continuous cultivation, such as

irrigated maize or meadow. With an annual rainfall

varying from 767 mm in 1991 to 1010 mm in 1992,

water drainage below meadow varies by nearly 100%;

nitrate leaching below irrigated maize shows that

fertilisation at predetermined dates does not corre-

spond to optimal practices. Nitrate leaching is also

more variable than water flow.

Table 3

Agronomic characteristics for the catchment simulation experiment

Sowing

(day)

First application fertiliser

(kgN/ha) (day)

Second application fertiliser

(kgN/ha) (day)

Harvest

(day)

Irrigated maize 112 50 (111) 160 (167) 297

Maize (dry cultivation) 111 50 (111) 120 (167) 297

Wheat 283 60 (33) 100 (79) 195

Sunflower 116 70 (116) – 244

Same values (day and amount) used for consecutive years.

Table 4

Annual climatological data

PET (mm, total of year) Annual rainfall (mm) Irrigation (mm, total of year)

1991 886 767 268

1992 787 1010 120

1993 823 1115 108

G. Vachaud, T. Chen / Journal of Hydrology 264 (2002) 101–112106

4. Stochastic non-distributed modelling

Clearly, a large amount of information on soil

characteristics and land use are necessary to run such

a spatially distributed model (Rhodenburg et al.,

1986). It is quite rare to have the corresponding set of

input data available, particularly in a large catchment,

and some authors apply effective parameters derived

from local measurements (Djurhuus et al., 1999). In

contrast, different sources of information concerning

the percentage coverage of those variables are often

accessible, either at the catchment scale or from large-

scale data sets, such as National or European

databases. The aim of this section is to compare the

results obtained in the previous case study with those

resulting from only using statistical information

obtained either at the catchment scale or at the

European level. In both cases, Latin Hypercube

Sampling of key input parameters is used to generate

stochastic simulations in the framework of

ANSWERS.

4.1. Modelling tool

A Latin Hypercube Sampling algorithm (McKay

et al., 1997), which is a multivariate stratified random

sampling technique (Pebesma and Heulevink, 1999),

was selected, since it permits stochastic generation of

sets of variables satisfying two constraints: the

characteristics of each variable population (for each

variable the probability distribution function of the

generated set is prescribed) and the areal distribution

of variables (in our case the percentage cover of each

crop, or each soil class on the catchment). An

advantage of equiprobable stratified sampling, such

as used in LHS, is that in contrast with classical Monte

Carlo simulation with simple random sampling

(Jensen and Mantoglou, 1992) smaller number of

samples is usually required to represent the parameter

space. This is particularly true, when the number of

strata is equivalent to the number of generated

samples (in our case 200), a situation defined as

corresponding to a maximal stratification. For catch-

ment averages, the sampling may be up to 10 times

more efficient than a classical Monte Carlo sampling

method. Input variables generated by LHS were then

used in ANSWERS following the method described in

the first article of this set.

4.2. Simulation with local source variables

First the modelling tool was built with the use of

information obtained at the local scale (catchment)—

as defined in Section 2.2 of this article—on the

following basis:

† Soil classes. For each soil class, a set of 200

realisations was generated to determine the histo-

grams of transport parameters accounting for the

soil texture variability within soil classes. To

achieve this goal, first for every variable (%sand,

%silt, %clay, %organic matter, CEC) 200 values

Table 5

Comparison between annual cumulative values of water losses

(drainage) and nitrate losses (leaching) resulting from distributed

modelling for the dominant crops3 of the year for the whole

catchment (catchment average ^ standard deviation)

Drainage water

(mm)

Nitrate leaching

(KgN/ha)

1991

Irrigated maize 381.8 ^ 3.0 95.3 ^ 9.8

Rainfed maize 280.1 ^ 9.9 69.7 ^ 9.4

Winter wheat 251.6 ^ 13.9 99.8 ^ 20.2

Sunflower 272.7 ^ 9.3 66.5 ^ 11.2

Meadow 244.7 ^ 7.1 10.9 ^ 3.9

Fallow (bare soil) 406.9 ^ 3.9 168.2 ^ 23.9

Catchment 294.5 ^ 52.4 81.6 ^ 45.5

1992

Irrigated maize 540.8 ^ 2.9 101.2 ^ 8.3

Rainfed maize 500.8 ^ 5.4 79.5 ^ 12.9

Winter wheat 491.8 ^ 7.3 170 ^ 22.2

Sunflower 478.7 ^ 7.2 109.9 ^ 13.7

Meadow 459.6 ^ 5.5 2.6 ^ 3.8

Fallow (bare soil) 617.4 ^ 14 180.3 ^ 25.2

Catchment 509.4 ^ 51.2 109.2 ^ 63.8

1992

Irrigated maize 581.9 ^ 19.5 105.3 ^ 5.9

Rainfed maize 523.2 ^ 27.0 79.6 ^ 21.3

Winter wheat 460.4 ^ 30.9 132.1 ^ 18.1

Sunflower 472.2 ^ 24.1 126.4 ^ 23.7

Meadow 435.4 ^ 28.6 2.7 ^ 3.1

Fallow (bare soil) 653.9 ^ 32.4 158.7 ^ 20.3

Catchment 510.2 ^ 78.5 99.5 ^ 54

The dominant crop of year x is the crop being harvested the

same year.

G. Vachaud, T. Chen / Journal of Hydrology 264 (2002) 101–112 107

were drawn by stratified sampling, then 200 sets of

these variables were randomly built following the

method of permutation illustrated by Pebesma and

Heulevink, (1999). It is quite clear that the three

variables describing texture are not independent;

this choice was imposed by the irregular shape of

the domain defining a soil class in the textural

triangle, a point already discussed by Vachaud and

Chen (2002).

† Crop coverage. The spatial coverage of the catch-

ment consists of natural prairies (18%); irrigated

crops, essentially maize (10%); dry cultivation crops:

maize (13%), winter wheat (34%), sunflower (15%)

and fallow (10%). For every crop, we assume a

uniform distribution in which the same proportion is

found in each soil class

† Fluxes within a soil class. For every soil class, the

distribution of fluxes corresponding to water

drainage and nitrate leaching below the root zone

were obtained from a total of 200 simulations using

ANSWERS, with a weighting procedure according

to the areal coverage of crops. Assuming the areal

proportion of crop Xk to be k%, a number of

(0.01†k†200) samples were randomly drawn from

the population of results corresponding to that crop

† Fluxes at the catchment scale. For the three classes

of soil (variable Yj), each one representing an areal

proportion j% of the catchment, 200 samples were

extracted from the set of 600 realisations obtained

by the previous procedure by random drawing of a

number of (0.01†j†200) samples from the popu-

lation corresponding to every soil class.

No further information concerning the distribution

of variables in the catchment was used. Results are

expressed in Table 6 and compared to those given by

the spatially distributed approach using local sources.

They will be discussed in the last part of this article.

4.3. The use of European source information

Frequently, however, data on soils and land use are

not even available at the local scale, and in these

situations, recourse is often made to data that is available

at a larger regional scale. In Europe, in particular, large

efforts have been deployed during the past decades to

obtain statistics at European scale. The last illustration

will thus deal with the use of European sources of land

use and soil information to simulate this catchment.

Soil data. Soil data is provided by the digital

1:1,000,000 scale Soil Map of Europe. Although the

national soil information, which has been used to

construct this map is based on a variety of

methodologies, scales, classifications, etc. this infor-

mation has been harmonised at the 1:1,000,000 scale

using the FAO classification system. The soil map

shows the distribution of Soil Map Units (SMU’s).

Each SMU is a grouping of individual soil types (or

Soil Typological Units, STU’s), which are generally

found together in the landscape. The number of STU’s

within a single SMU depends upon the complexity of

the soil pattern in a given landscape, but generally

varies between two and five.

In our case, 13 STUs are identified on an area of

540 km2 (about three times larger than the catchment).

However, seven units represent (altogether) less than

5% of the area, and for the six remaining units there are

only two classes of dominant textural class:

† Medium (18% , clay , 35% and sand . 15%, or

clay , 18% and 15% , sand , 65%) for 92%

Table 6

Local database: comparison between annual cumulative values of water losses (drainage) and nitrate losses (leaching) resulting from distributed

or non-distributed modelling for the entire catchment (catchment average ^ standard deviation)

Distributed model Stochastic model (LHS)

Drainage water (mm) Nitrate leaching (KgN/ha) Drainage water (mm) Nitrate leaching (KgN/ha)

1991 294.5 ^ 52.4 81.6 ^ 45.5 289.6 ^ 57.9 79.6 ^ 45.5

1992 509.4 ^ 51.2 109.2 ^ 63.8 505.1 ^ 54.2 108.7 ^ 64.4

1993 510.2 ^ 78.5 99.5 ^ 54 496.0 ^ 94.7 98.4 ^ 53.6

G. Vachaud, T. Chen / Journal of Hydrology 264 (2002) 101–112108

† Medium fine (clay , 35% and sand , 15%) for

8%

Land use data. The European source of land use

and cropping data is the REGIO database available

from Eurostat. This gives the annual hectarage of a

variety of land uses and crops (Table 7) for the

Administrative Units (NUTS) of each European

country from around 1974. For this example, the

NUTS level 1 area of Rhone Alpes was used. Some

screening had to be done on the basis that some land

use (such as vineyards, olive plantations) or some

crops (such as rice) cannot be found in the northern

part of Rhone Alpes and that the proportion of

agricultural land on this catchment is 95% instead of

only 40% for the whole Rhone Alpes area in which

mountains are dominant. All crops representing less

than 2% were not accounted for. On this basis, the

percentage land use of the total utilised agricultural

area from this database is as follows:

† permanent grassland, 55%

† arable land: 45% with the following crops:

irrigated maize, 30%; rainfed maize: 11%, cereal

(wheat, barley) 35%, sunflower and oilseed rape:

24% (no fallow on the database)

These values have to be compared with those

following obtained from the local source:

† permanent grassland, 18%

† arable land: 82% with the following crops:

irrigated maize, 10%; rainfed maize: 13%, cereal

(wheat, barley) 34%, sunflower and oilseed rape:

15%, fallow: 10%

Land management data. There are no European

sources of land management information available to

provide such data as pesticide or fertiliser usage at the

subnational level. Fertiliser usage data has therefore

been derived using local sources of usage data

combined with the European land use data.

Meteorological data. There were no European

sources of meteorological data available at a shorter

temporal scale than monthly at the time of modelling.

Therefore, local sources of meteorological data have

been used.

On this basis, stochastic simulations were run to

generate 200 realisations with the constraints of crop

cover and soil classes given above for the 3 year

period covered previously. Results are expressed in

Table 8 and should be compared to those reported in

Table 6 for the same conditions.

5. Discussion and conclusion

In spite of the fact that several results obtained in

this study are site specific (for example, the absence of

runoff) several general points can be highlighted.

The first result to discuss concerns the comparison

between distributed and stochastic outputs obtained

with the same model and the same source of

information. Clearly, the quality of information is

quite different, since distributed modelling gives

access to the spatial distribution of fluxes within the

watershed, and so locates sensitive areas in terms of

risks of pollution, a point crucial for risk management.

On the other hand, the stochastic modelling will

provide information only on catchment-aggregated

values offluxes (average, percentiles, ranges, variance).

It is clear from the results given in Table 6 that

annual losses obtained with the stochastic non-

distributed model or with the distributed model are

quite close. In fact, they should be identical, if the

variables use in both cases satisfy the same probability

distribution functions, since in the absence of lateral

flow the spatial integration of the distributed vertical

model should give the same result as the stochastic

integration of the random field. The slight differences

Table 7

Land use and crop categories in the Eurostat Regio database

Land use categories Crop categories

Total area Cereal

Forest Wheat

Agricultural area Barley

Gardens Maize (grain)

Grassland Maize (fodder)

Permanent crops Potatoes

Vineyards Sugar beet

Olives Total oilseeds

Arable Oilseed rape

Greenfodder (on arable land) Sunflower

Tobacco

Rice

G. Vachaud, T. Chen / Journal of Hydrology 264 (2002) 101–112 109

between results in Table 6 for both models is due to

the fact that in the distributed model a soil class is

represented by a unique set of values (that of the

barycentre of the class) whereas in the non-distributed

model the same class is represented by a population of

200 samples distributed within the class domain.

The largest advantages of the stochastic non-

distributed modelling are the gain in term of data

acquisition (no need of georeferenced input), as well

as an impressive gain in terms of simulation time

(5 min for the 3 year simulation at a daily time step

instead of 2 h with the same desk computer). The most

important limitation is that these conclusions are only

valid if there is no runoff or lateral flow in the system.

The second result is concerned with the compari-

son between non-distributed models using sources of

information obtained at different scales. A major

problem in environmental modelling is often the cost

of data acquisition, and there is a common trend to use

available data obtained quite often at a scale not in

coherence with that of the application. In this

particular example, ‘local’ information corresponds

basically to pedological and crop observations at a

scale smaller than 1:100,000 whereas those given in

the European database are provided at the scale

1:1,000,000. Two major sources of errors can be

easily identified. First, the estimation of crop coverage

obtained from the European source does not corre-

spond to that of the catchment; and second, the

textural limits of the soil classes of the European

STUs are too broad.

When applying European source cropping data to

catchments, it is imperative that model is calibrated in

the light of local estimates of the proportion of

agricultural to non-agricultural land use. It is indeed

quite obvious that proportions of land uses may differ

considerably, as it is the case for that given catchment,

between an area covering a few hundred square

kilometres and another covering some 10,000 km2.

Similarly, soil descriptions derived from European

sources cannot reach the same level of precision as the

local data. Consequently, the combined use of large-

scale data for plants and large-scale data for soils

yields annual values of cumulative drainage and

cumulative nitrate leaching which differ quite sensi-

bly for those obtained with the reference model (Table

8). It is, however, worth noting that cumulative nitrate

leaching deviates more from the reference value than

drainage, a point easily explainable by the fact that

difference of water consumption between crops is

much smaller than difference in fertilisation and in

nitrogen uptake.

If, however, the objective is to simulate various

land uses for screening purposes, reasonable approxi-

mations can be obtained using the European sources

of data. For example, in spite of the fact that estimates

of nitrate leaching deviate strongly from the reference

values (Table 8), they remain within their ranges of

uncertainty (expressed by ^one standard deviation).

As a conclusion of this study, it appears that with a

given methodological tool, such as ANSWERS,

different levels of simplification of the input data

can be selected, depending on the objectives of the

modelling and the level of acceptable losses of

information on outputs. The most detailed infor-

mation will always be given by distributed modelling;

for example, when the knowledge of spatial distribution

of fluxes within a watershed is an essential element of

environmental management. The price to be paid for

this detail is in the spatially distributed data required. If

surface lateral flow is negligible, and the objective is the

determination of catchment-aggregated information,

Table 8

European database; cumulative values of water losses (drainage) and nitrate losses (leaching) for the entire catchment (catchment

average ^ standard deviation), and deviation of annual mean value from that obtained in the catchment distributed model (reference) with local

source of inputs

Drainage water (mm) Nitrate leaching (KgN/ha)

Euro-database Deviation % from reference Euro-database Deviation % from reference

1991 330.8 ^ 61.5 12.5 116.0 ^ 52.3 42

1992 554.9 ^ 62.5 9 135.2 ^ 48.6 23

1993 556.3 ^ 102.3 9 115.8 ^ 45.6 16

G. Vachaud, T. Chen / Journal of Hydrology 264 (2002) 101–112110

then stochastic simulation with statistical values of

input variables at the scale of the watershed may

suffice. The benefits of such an approach are the

reduced cost of data acquisition and reduction in

simulation time. Finally, statistical values obtained at

a scale larger than that of the watershed may also be

used, but the degradation of information may not be

acceptable, unless this is done in the concept of

screening purposes (comparison of effects of land use)

after a careful calibration concerning in particular the

proportion of agricultural lands and the types of

agricultural system.

Acknowledgments

This work was done in the frame of ‘CAMS-

CALE—Upscaling predictive models and catchment

water quality’ a European Union DGXII—Environ-

ment funded programme (ENV4-CT97-0439). The

authors are particularly grateful to Dr J.M. Hollis, co-

ordinator, Soil Survey and Land Research Centre,

Cranfield University, Silsoe, UK, to Prof. N. Jarvis,

Department of Soil Science, Uppsala University,

Sweden, and other members of the programme for

very fruitful discussions and suggestions concerning

this study, to Dr Adrian Amstrong, ADAS

Gleadthorpe, UK, and Prof. T. Burt, Dept. of

Geography, Univ. of Durham, UK, for their careful

editorial help and to anonymous referees of the initial

version of this paper for thoughtful and helpful

suggestions on the presentation of our results.

References

Beasley, D.B., Huggins, L.F., Monke, E.J., 1980. ANSWERS: a

model for watershed planning. Transactions of ASAE 23,

938–944.

Bel, F., Lacroix, A., Mollard, A., David, C., Beaudoin, N., Mary, B.,

Vachaud, G., Vauclin, M., Garino, B., 1999. Une approche

interdisciplinaire, pluri-echelle, multipartenaire des pollutions

diffuses de l’eau: l’experience de la Cote Saint-Andre, Isere. La

Houille Blanche 6, 72–79.

Binley, A.M., Beven, K.J., Calver, A., Watts, L.G., 1991. Changing

responses in hydrology: assessing the uncertainty in physically

based model predictions. Water Resources Research 27,

1253–1261.

Bouraoui, F., 1994. Development of a continuous, physically based,

distributed parameter, non-point source model. Doctoral Thesis

(PhD), Faculty of Agricultural Engineering, Virginia Polytech-

nic Institute, Blacksburg, VA, USA, p. 110.

Bouraoui, F., Vachaud, G., Haverkamp, R., Normand, B., 1997a. A

physical approach for surface–subsurface water transport

modeling in agricultural watersheds. Journal of Hydrology

203, 75–88.

Bouraoui, F., Vachaud, G., Haverkamp, R., Chen, T., Dillaha, T.A.,

1997b. A GIS integrated distributed approach for non-point

source pollution modelling. IAHS, Public.240, Proceedings of

the Rabat Symposium, April 1997, pp. 377–384.

Clamons, S.F., Byars, B.W., 1997. GRASS 4.2 Programmer’s

Manual, US Army-Construction Engineering Research Labora-

tory, Champaign, IL, USA, p. 192.

Davis, S.H., Vertessy, R.A., Silberstein, R.P., 1999. The sensitivity

of a catchment model to soil hydraulic properties obtained by

using different measurement techniques. Hydrological Pro-

cesses 13, 677–688.

Djurhuus, J., Hansen, S., Schelde, K., Jacobsen, O.H., 1999.

Modelling mean nitrate leaching from spatially variable fields

using effective hydraulic parameters. Geoderma 87, 261–279.

Finke, P., Wosten, J., Jansen, M., 1996. Effects of uncertainty in

major input variables on simulated functional soil behavior.

Hydrological Processes 10, 661–669.

Freissinet, C., Vauclin, M., Ehlirch, M., 1999. Comparison of first-

order analysis and fuzzy set approach in the evaluation of

imprecisions in pesticide groundwater pollution using a screen-

ing model. Journal of Contaminant Hydrology 37, 21–43.

Hack-ten-Broek, M.J.D., Hegmans, J.H.B.M., 1996. Use of soil

physical characteristics from laboratory measurements or

standard series for modelling unsaturated water flow. Agricul-

tural Water Management 29, 201–213.

Jensen, K.H., Mantoglou, A., 1992. Application of stochastic

unsaturated flow theory, numerical simulations, and comparison

to field observations. Water Resources Research 28, 269–284.

Knissel, W.G., 1993. GLEAMS: groundwater loading effects of

agricultural management systems, Version 2.10, Model Docu-

mentation, University of Virginia, USDA, ARS, Tifton, GA,

USA.

McKay, M.D., Beckman, R.J., Conover, W.J., 1987. A comparison

of three methods for selecting values of input variables in the

analysis of output from a computer code. Technometrics 21,

239–245.

Normand, B., Recous, S., Vachaud, G., Kengni, L., Garino, B.,

1997. Nitrogen balance of an irrigated maize by continuous

monitoring plus sampling at harvest. Soil Science Society of

American Journal 61, 1508–1518.

Pebesma, E.J., Heulevink, G.B.M., 1999. Latin Hypercube

Sampling on Gaussian random fields. Technometrics 41,

303–312.

Rawls, W.J., Brakensiek, D.L., 1989. Estimation of soil water

retention and hydraulic properties. In: Morel-Seytoux, H.J.,

(Ed.), Unsaturated Flow in Hydrologic Modeling: Theory and

Practice, NATO ASI Series C, vol. 275. Kluwer Academic

Publishers, Dordrecht.

Refsgaard, J.C., Thorsen, M., Jensen, J.B., Kleeschulte, S.,

Sauerbeck, D., 1999. Large-scale modelling of groundwater

G. Vachaud, T. Chen / Journal of Hydrology 264 (2002) 101–112 111

contamination from nitrate leaching. Journal of Hydrology 221,

117–140.

Rhodenburg, H., Diekkruger, B., Bork, H.R., 1986. Deterministic

hydrological site and catchments models for the analysis of

agro-ecosystems. Catena 13, 119–137.

Tietje, O., 1999. Pedotransfer functions with interdependency of

hydraulic parameters. In: van Genuchten, M.Th., Leij, F.J., Wu,

L. (Eds.), Proceedings of the International Workshop on

Characterization and Measurement of the Hydraulic Properties

of Unsaturated Porous Materials, Riverside, California, October

1997, pp. 1225–1236.

Vachaud, G., Chen, T., 2002. Sensitivity of computed values of

water balance and nitrate leaching to variability of soil transport

parameters within a soil class. Development of a Latin

Hypercube Sampler combined with a deterministic model

(ANSWERS). Journal of Hydrology, 264, 89–102.

G. Vachaud, T. Chen / Journal of Hydrology 264 (2002) 101–112112