You are on page 1of 13

Journal of Hydrology 273 (2003) 164176 www.elsevier.

com/locate/jhydrol

Parameter uncertainty and the signicance of simulated land use change effects
Klaus Eckhardta,*, Lutz Breuerb, Hans-Georg Fredeb
b

METCON Umweltmeteorologische Beratung, Jappopweg 9h, D-25421, Pinneberg, Germany Institute of Landscape Ecology and Resources Management, Justus-Liebig-University, Gieen, Germany Received 18 December 2001; revised 12 November 2002; accepted 15 November 2002

Abstract Uncertainty in parameters characterising different land covers leads to uncertainty in model predictions of land use change effects. In this study, a new approach is presented which allows a model to be assessed to see whether it is suitable for investigating land use change scenarios in the sense that different land covers can be signicantly distinguished in their effects on model output. It consists of the following steps: (a) The uncertainty in land cover-dependent parameters is quantied. (b) The model of an articial catchment with representative characteristics and uniform land cover is established. (c) Using this articial catchment, Monte Carlo simulations are carried out to determine the uncertainty in the model response to different land covers. (d) By comparing the results for two covers, respectively, a dimensionless test statistic, the distinction level, is calculated. The distinction level is a normalised probability that two independent realisations of land covers which are parameterised within their range of natural uncertainty will yield distinct model responses. If the distinction level is greater than or equal to 90%, the land covers are assumed to have a signicantly different effect on the model output. An example of the application of the new method is provided using the eco-hydrologic model SWAT-G. The land covers forest, pasture and arable land can be signicantly distinguished by their long-term means of surface runoff, groundwater recharge and streamow. The minimum proportion of the catchment area on which land cover must change in order to obtain signicantly distinct model responses depends on the land covers involved and the considered hydrologic variable. In the case of a change between pasture and forest and with regard to average streamow, this minimum proportion amounts to about 25%, a value that compares well with the results of paired catchment studies. q 2003 Elsevier Science B.V. All rights reserved.
Keywords: Land use change; Land cover; Parameter uncertainty; Uncertainty analysis; Model validation; Model error

1. Introduction A major issue in river basin management studies is the assessment of land use change effects. Distributed models relying on a physically based description of the runoff generation and the effects of different land
* Corresponding author. Fax: 49-4101-693857. E-mail address: eckhardt@metcon-umb.de (K. Eckhardt).

covers are a widespread tool for answering the question of how land use changes affect hydrologic processes, erosion and transport of water constituents in a catchment. In these models, land cover properties, which are relevant for the considered processes, have to be characterised by plant-specic parameters. Technically, simulation of a land cover change signies a modication of the values of these parameters in

0022-1694/03/$ - see front matter q 2003 Elsevier Science B.V. All rights reserved. PII: S 0 0 2 2 - 1 6 9 4 ( 0 2 ) 0 0 3 9 5 - 5

K. Eckhardt et al. / Journal of Hydrology 273 (2003) 164176

165

certain parts of a catchment. Thus, reliable results can only be obtained if the parameter values for the land covers involved are known with some accuracy. However, a review of the literature (Breuer et al., 2003) reveals a high degree of uncertainty in the parameterisation of land covers. This uncertainty has several causes. Observation of some parameters is problematic. Therefore, measurements may be scarce or different measurement techniques may give systematically different results. Moreover, the regionalisation of point measurements is difcult because of the natural variability of plant characteristics for reasons of climate, soil, stand age, etc. Yet, even if these cause-variables and their effects were perfectly known, some of this information would be lost again when establishing the model because of the need to aggregate plant characteristics over the spatial subunits of the model. Against the background of these uncertainties, it is questionable whether different land covers can be signicantly distinguished in their effect on the simulation results at all. If two different land covers do not produce a signicantly distinct model response, however, the distinction between these covers is redundant. We can cope with the problem of parameter uncertainty by carrying out Monte Carlo simulations (e.g. Binley et al., 1991; Nandakumar and Mein, 1997). Yet, Monte Carlo simulations are computationally demanding. In most cases, effects of land use changes are therefore assessed by comparing only two model runs, each one representing a different parameterisation of the model as, for example, in the papers of Dunn and Mackay (1995), Storck et al. (1998), Matheussen et al. (2000), Wechsung et al. (2000), Bronstert et al. (2001), and Fohrer et al. (2001). In the present paper, a method is developed which makes it possible to determine if a change in parameterisation between two model runs will result in a signicant model reaction. First, Monte Carlo simulations are carried out independently for each land cover taking into account the whole range of parameter uncertainty and providing frequency distributions showing the uncertainty in model output. These frequency distributions then are compared in pairs. A dimensionless test statistic is derived which expresses the probability that two independent

realisations of the corresponding land covers will yield different values of the considered model output variable. If the test statistic is greater than a threshold, the two land covers are assessed to produce a signicantly distinct model response. Changes should only be simulated between land covers that meet this criterion. After demonstrating the application of the new method using the example of a conceptual distributed model, the problem of model validation by paired catchment studies is addressed. There is widespread agreement that a good correspondence of measured and calculated streamow at the catchment outlet is not a sufcient criterion for the validity of a physically based model but that additional knowledge on internal variables is required (e.g. Beven, 1989). In the special case of models intended to simulate land cover change, paired catchment studies and the investigation of historical catchment conditions, combined with corresponding streamow measurements, could contribute to a more reliable assessment of the model efciency. With regard to this approach, a step-by-step land cover change from arable land to pasture and then to forest is simulated. The objective is to quantify how much of a catchments land cover has to be changed to get a signicant hydrologic response and so to support the decision on whether it is worthwhile carrying out a study of the abovementioned type or not. The starting points of this paper were investigations on land cover changes in a German low mountain range (Fohrer et al., 2001). The focus of the study is therefore on mid-latitude temperate climate conditions and catchments with elevations ranging from about 100 to 1000 m a.s.l., predominantly with steep slopes and shallow soils. Nevertheless, the methodology can also be applied to different conditions.

2. Description of the model SWAT-G The model being used in this study is the version SWAT-G (Eckhardt et al., 2002) of the conceptual distributed model SWAT (Soil and Water Assessment Tool) (Arnold et al., 1998). SWAT has been developed to predict the impact of management on water balance, erosion and transport of nutrients and

166

K. Eckhardt et al. / Journal of Hydrology 273 (2003) 164176

pesticides in meso- to macroscale basins. Major model components include: weather, hydrology, soil, temperature, plant growth, nutrients, pesticides, and land management. The smallest spatial sub-units resolved by the model are hydrotopes (hydrologic response units) which are assumed to be homogeneous with respect to their hydrologic properties. In each of the hydrotopes, water balance is represented by several storage volumes: canopy storage, snow, soil prole, shallow aquifer, and deep aquifer. The soil prole can be subdivided into multiple layers. Soil water processes include inltration, evaporation, plant uptake, lateral ow, and percolation to lower layers. Percolation from the bottom of the soil prole recharges the shallow aquifer. A recession constant is used to lag ow from the aquifer to the stream. Other shallow aquifer components include evaporation, pumping withdrawals, and seepage to the deep aquifer. Lumped modelling of hydrological and chemical processes, plant growth and management in the hydrotopes yields the runoff which, together with its sediment, nutrient and pesticide loads, is then routed through the channel network to the catchment outlet. Time step length is one day. The crop model in SWAT is a simplication of the EPIC crop model (Williams et al., 1984). Growth can only occur if the daily mean temperature exceeds a plant-specic base temperature. The temperature excess, counted in heat units, is accumulated over the time. The phenological development of the plants is controlled by comparing the actually accumulated heat units to the predened heat unit sum required for maturity of the plant. SWAT uses Monteiths approach to estimate potential biomass (Monteith, 1977), coupled with water, temperature and nutrient stress adjustments. The leaf area index is simulated as a function of heat units and varies between plantspecic potential minimum and maximum values. Potential evapotranspiration is estimated with the Penman Monteith method (Monteith, 1965). Canopy evaporation is a function of potential evapotranspiration, maximum interception capacity and the ratio of actual and potential maximum leaf area index. Plant water uptake from the soil is simulated as a function of potential evapotranspiration, leaf area index and rooting depth and is limited by soil water content.

3. Methodology 3.1. Catchment For the present study, the use of complex natural catchments has several disadvantages: Reactions of different parts of the catchment will mix which makes it difcult to identify signals originating from certain areas of different land cover types. Furthermore, these differences can be caused by factors other than land cover, e.g. soil type and hillslope. Finally, a complex catchment structure requires a complex model representation. This complicates the modication of model parameters and increases the computational demands. Therefore, instead of a natural catchment a simply structured articial catchment is considered. It is formed as a V-shaped valley (Fig. 1) with a base area of 2 km2. The articial catchment is designed as part of a low mountain range. Main characteristics of the articial catchment are representative for the Dill catchment, a mesoscale low mountain range catchment (area: 693 km2) situated in the south-east of the Rhenish Massif in Germany. The mean elevation of the articial catchment amounts to 400 m a.s.l., the hillslope is 15% and the soil mainly consists of a shallow cambisol (soil depth: 1.1 m) over a hard rock aquifer. Deeper gley is only found alongside the stream, on 10% of the area. The weather input data is taken from a station that is situated in the center of the Dill catchment, the Dillenburg station (508440 N, 88160 E) of the DWD (Deutscher Wetterdienst German Weather Service). The mean annual air

Fig. 1. Schematic of the articial catchment.

K. Eckhardt et al. / Journal of Hydrology 273 (2003) 164176

167

temperature is 6.6 8C, the mean annual precipitation amounts to 880 mm. We calibrated the model of the articial catchment by comparing the calculated streamow at its outlet with measured streamow at the outlet of the Dill catchment. During this calibration, the cambisol was assumed to be covered by forest and the gley by pasture. Fig. 2 shows the monthly means of measured and calculated streamow over the six considered hydrologic years. The long-term mean values (measured: 342 mm/a, calculated: 341 mm/a) as well as the seasonal variations of the streamow correspond well. Summer runoff is underestimated because we assume that 90% of the articial catchment is covered by forest while this is the case in only 56% of the Dill catchment. Differences in the times of concentration in the articial and the natural catchment are irrelevant since, in this study, we are only interested in the long-term mean response of the catchment and, therefore, consider monthly means. The model efciency (Nash and Sutcliffe, 1970) over the 72 monthly mean values attains a value of 0.88. We conclude that we can consider the articial catchment as being representative for a natural catchment, not necessarily in its short-term reactions, but with respect to its mean response to

variations conditions.

in

the

meteorological

boundary

3.2. Land covers and their parameterisation Four different land cover types are simulated: mixed coniferous forest (Picea abies and Pinus sylvestris), mixed deciduous forest (Fagus sylvatica and Quercus petrea), pasture andrepresenting arable landa cropping sequence of summer barley (Hordeum vulgare), winter rape (Brassica napus) and winter wheat (Triticum aestivum). The model description in Section 2 has already named some of the parameters whose uncertainty will be taken into account: base temperature, minimum and maximum leaf area index, maximum interception capacity and maximum rooting depth. Albedo is needed to calculate the net energy available for evapotranspiration. Furthermore, maximum values for stomatal conductance and plant height have to be specied which are entered into the Penman Monteith formula for the calculation of potential evapotranspiration (Monteith, 1965). Finally, because the surface runoff volume is calculated by using a modication of the curve number technique (USDA-SCS, 1972), the curve number will be considered.

Fig. 2. Monthly streamow measured at the outlet of the Dill catchment/Germany (stream gauge Alar; 508350 N, 88280 E) and calculated with the articial catchment.

168

K. Eckhardt et al. / Journal of Hydrology 273 (2003) 164176

Table 1 Considered land cover-dependent parameters and their assumed uncertainty Parameter Coniferous forest Deciduous forest Pasture Arable land Barley Curve number Base temperature (8C) Maximum leaf area index Minimum leaf area index Maximum interception capacity (mm) Maximum stomatal conductance (mm/s) Maximum plant height (m) Maximum rooting depth (m) Albedo
a

Rape 6882a 46 2.04.5 0 1.92.3 2.55.5 1.01.2 1.11.7 0.210.26

Wheat 6882a 46 2.04.5 0 1.92.3 5.0 12.0 1.01.2 1.11.7 0.210.26

5060 48 4.514.5 2.59.5 2.56.5 0.54.0 18.032.0 1.52.5 0.100.13

5060 610 5.012.0 1.02.0 2.54.5 1.03.0 16.030.0 1.51.8 0.190.26

5567 34 1.59.0 1.02.0 1.03.0 2.0 10.0 0.41.5 0.70.9 0.200.28

6882a 58 1.82.4 0 1.72.1 3.57.5 1.01.2 0.81.2 0.210.26

Between seed and harvest; curve number for fallow: 85.

The parameter ranges are shown in Table 1. As far as the curve number is concerned, soil group B and hydrologic condition good are assumed and the tabled values are assumed to vary by ^ 10%. The denition of the other parameter ranges is based on the literature review of Breuer et al. (2003). In most cases, the number of observations was sufcient to apply a statistical test which showed normal distribution. Therefore, normal distributions are assumed for all parameters. Furthermore, it is assumed that the parameter ranges shown in Table 1 correspond to the interval [mean value 2 3 standard deviations, mean value 3 standard deviations]. Parameter values beyond the interval bounds are not taken into account. Land cover is independently parameterised in each model run. It may be argued that the parameterisation of different land covers should be correlated to some extent and that consequently an independent parameterisation will lead to an overestimation of uncertainty. Yet, formulation of any relationship between parameter values for different land covers is inevitably uncertain in itself. As land cover changes, boundary conditions at a site also change. Land cover inuences microclimate and soil conditions. Furthermore, plant characteristics depend on management practice which may neither be exactly known for the present state nor for the simulated scenario. It is therefore concluded that an independent parameterisation is admissible as a rst approximation. At least, it allows a worst-case assessment. If we nd that two land covers are signicantly

distinct with respect to their effect on the model output, then this nding will not change if we introduce additional constraints in the land cover parameterisation. 3.3. Calculation of the distinction level Uncertainty propagation is calculated by means of Monte Carlo simulations. To determine the required number of model runs, comparative calculations with 10,000 and 50,000 model runs were carried out. The resulting frequency distributions of output variables were nearly identical. Their 5th and 95th percentiles differed by less than 2%. So it was concluded that 10,000 model runs in each Monte Carlo simulation are sufcient to obtain representative distributions of the model output. Each model run covers six hydrologic years and is based on a set of plant parameter values randomly drawn from the normal distributions dened in Section 3.2. The rst year is used as a warm-up period. The remaining ve hydrologic years are evaluated with respect to the mean values of three hydrologic variables: surface runoff, groundwater recharge and streamow. These variables are selected because of their importance for erosion, drinking water supply, ooding risk assessment, model calibration, etc. Each Monte Carlo simulation yields a frequency distribution for each of the variables. We want to detect whether different land covers can be distinguished with respect to their hydrologic

K. Eckhardt et al. / Journal of Hydrology 273 (2003) 164176

169

Fig. 3. Schematic of frequency distributions of a hydrologic variable y for two land covers A and B.

effects. To this purpose, the results for two land covers respectively are compared. In the following, it is assumed that the potential range of the considered output variable y is partitioned into N intervals yi. Fig. 3 schematically shows two frequency distributions obtained by means of Monte Carlo simulations, one distribution for a land cover A and one for a land cover B. By normalising the frequencies in the intervals yi i 1; ; N to the number of model runs per Monte Carlo simulation, approximate values for the probabilities pyi i 1; ; N for nding the value of the variable y in one of the intervals yi are calculated. In most studies, only two model runs are compared to assess the effects of a land cover change, one realisation of land cover A and one realisation of land cover B. Distinct model responses may be observed, but there is a certain probability that the parameterisation of the land covers in these two model runs was incorrect and hence that the observed effect is only feigned by an erroneous choice of parameter values. A correct parameterisation of the land covers might lead to the same model response, for example a value in the interval y0 . Let py0 A denote the probability that for land cover A the calculated value of y lies in the interval y0 and py0 B the corresponding probability for land cover B. The probability py0 A; B that two model runs, a realisation of land cover A and a realisation of land cover B, will yield a model response y in the

same interval y0 is py0 A; B py0 Apy0 B 1

The overall probability Py(A,B) that two independent realisations of the land covers A and B will yield the same result with respect to y is obtained by summation over all intervals: Py A; B
N X i1 N X i1

pyi A; B

{pyi Apyi B}

The distinction between two land covers A and B is signicant, if the probability Py(A,B) that they will lead to the same hydrologic response is small. Yet, what signies small? Even the probability Py(A,A) or Py(B,B) that two independent realisations of the same land cover will yield identical results can be quite small if the corresponding frequency distribution is wide. Therefore, throughout this study Py(A,B) is normalised with respect to max {Py(A,A), Py(B,B)} and land cover change effects are assumed to be signicant if Py A; B # 0:1 max{Py A; APy B; B} or if Sy A; B $ 0:9 3

170

K. Eckhardt et al. / Journal of Hydrology 273 (2003) 164176

with Sy A; B 1 2 Py A; B max{Py A; APy B; B} 5

respectively. In the following, Sy(A,B) is called the distinction level of the land covers A and B with respect to the variable y. The distinction level Sy(A,B) is a normalised probability that two independent realisations of land covers A and B will yield different values of the model output variable y. If two identical frequency distributions are compared, then Sy(A,B) 0. If the distributions do not overlap, then Sy(A,B) 1. A common problem of statistical tests is how to dene the signicance criterion. In practice a threshold value of 0.95 or 0.99 is often chosen. Considering the highly schematic character of the articial catchment, we decided for the less restrictive criterion (4). If criterion (4) is not met, there is an elevated probability that two land covers will induce the same model response. Yet, if the simulation of a land cover change does not result in a signicant model reaction, a distinction between the considered land covers is redundant. Therefore, changes should only be simulated between land covers that meet the criterion (4).

and pasture (PAST) or forest (FRSD: deciduous forest, FRSE: coniferous forest) are highly signicant with respect to surface runoff (distinction level: 1.0). The surface runoff for deciduous and coniferous forest is similar because the curve numbers for both covers are the same. Differences are mainly caused by the more elevated interception in coniferous forest due to its smaller albedo (leading to a higher potential evapotranspiration) and greater interception capacity. Yet, these differences are not so important that coniferous and deciduous forest could be signicantly distinguished with respect to mean surface runoff (distinction level: 0.62). The frequency distributions of pasture and deciduous forest overlap, mainly because their curve number ranges do so. Nevertheless, pasture and deciduous forest as well as pasture and coniferous forest are signicantly distinct with respect to mean surface runoff (Ssurface runoff(PAST,FRSD) 0.99, Ssurface runoff(PAST,FRSE) 1.00). 4.1.2. Groundwater recharge The smallest groundwater recharge is found beneath forest. Again, there is a difference between deciduous and coniferous forest because of their different evapotranspiration. The higher evapotranspiration of coniferous forest is caused by its smaller albedo and greater mean leaf area index, interception, plant height and rooting depth. Yet, the difference is not sufcient to signicantly distinguish these two types of forest with respect to mean groundwater recharge (distinction level: 0.85). Pasture shows a signicantly higher groundwater recharge than the other land covers. Despite the low evapotranspiration of CROP, groundwater recharge beneath arable land is smaller than beneath pasture because on arable land much more surface runoff is generated (Fig. 4a). The frequency distributions for CROP and FRSD (deciduous forest) overlap. Nevertheless, CROP and FRSD as well as CROP and FRSE (coniferous forest) are signicantly distinct with respect to mean groundwater recharge (Sgroundwater recharge(CROP,FRSD) 0.99, Sgroundwater recharge(CROP,FRSE) 1.00). 4.1.3. Streamow The differences of the land covers, with respect to the streamow, are reciprocally proportional to the

4. Results 4.1. Changes between uniform land covers in the whole catchment First, we compare different versions of the articial catchment with one and only one land cover in the whole catchment, respectively. The resulting frequency distributions for mean surface runoff, groundwater recharge and streamow are shown in Fig. 4. The corresponding distinction levels are listed in Table 2. 4.1.1. Surface runoff On arable land (CROP) most surface runoff is generated because the soil is poorly covered, or even bare, over longer periods, interception is low and the curve number high. In our simulations, this effect is very pronounced because the catchment hillslope of 15%. Therefore, the differences between arable land

K. Eckhardt et al. / Journal of Hydrology 273 (2003) 164176

171

Fig. 4. Normalised frequency distributions of (a) surface runoff, (b) groundwater recharge, and (c) streamow assuming a uniform land cover on the whole catchment area.

172

K. Eckhardt et al. / Journal of Hydrology 273 (2003) 164176

Table 2 Distinction levels Sy(A,B) with respect to mean values of surface runoff, groundwater recharge and streamow Surface runoff FRSD FRSE PAST CROP 0.62 0.99 1.00 FRSE PAST Groundwater recharge FRSD 0.85 1.00 0.99 FRSE PAST Streamow FRSD 0.90 0.99 1.00 FRSE PAST

1.00 1.00

1.00

1.00 1.00

1.00

1.00 1.00

1.00

FRSD: deciduous forest, FRSE: coniferous forest, PAST: pasture, CROP: cropping sequence of summer barley, winter rape and winter wheat.

differences in evapotranspiration. The land cover with the smallest evapotranspiration-arable land- yields the highest streamow, followed by pasture, deciduous and coniferous forest. Now, coniferous and deciduous forest can just be signicantly distinguished (distinction level: 0.90). Aside from the distributions of FRSE and FRSD, only the distributions of deciduous forest and pasture overlap. FRSD and PAST are signicantly distinct with respect to mean streamow though (distinction level: 0.99), just like the other combinations of land covers. 4.2. Step-by-step change from one land cover to another Section 4.1 has shown that the four land covers are signicantly distinct with respect to nearly all considered output variables. Only the distinction between deciduous and coniferous forest may be problematic. Yet in practice, only in a few special cases of smaller watersheds, will a uniform land cover change take place in the whole catchment. Therefore, a land cover change in only parts of the catchment will now be simulated. The following calculations are further based on the assumption of an initially uniform land cover but now different stages of a transition to another cover are simulated (5, 10, 20, 30, 50 and 100% of cover change). The question to be answered deals with how much of the land cover has to be changed to get a signicant hydrologic response. To guarantee that only effects of the different parameterisation of the respective land covers are observed, the changes are realised on strips perpendicular to the stream at the axis of the articial catchment so that the two soils, cambisol

and gleysol, are equally affected. Table 3 summarises the results. 4.2.1. Change from arable land to pasture Fig. 5a c shows how the distinction levels Ssurface runoff(CROP,PAST), Sgroundwater recharge(CROP,PAST) and Sstreamow(CROP,PAST) change if arable land is replaced by pasture on an increasing proportion of the articial catchment. The 0.9 level is marked by a dashed line. A signicantly distinct mean surface runoff and groundwater recharge is obtained if land cover is changed on more than about 20% of the catchment area, while signicance for streamow is already attained if land cover is changed on only about 15% of the area. 4.2.2. Change from pasture to forest When replacing pasture by coniferous forest, about 40% of the catchment area has to be affected to observe a signicant change in mean surface runoff. In the case of a transition to deciduous forest, signicance is only attained if cover changes on 60 70% of the area. Signicant differences in mean groundwater recharge can be observed after a 25 30% land cover change. As regards streamow, coniferous
Table 3 Proportion of the catchment area which has to be affected by a land cover change to obtain a signicant effect on mean surface runoff, groundwater recharge or streamow CROP ! PAST Surface runoff (%) Groundwater recharge (%) Streamow (%) 20 20 15 PAST ! FRSE 40 25 25 PAST ! FRSD 6070 30 35

K. Eckhardt et al. / Journal of Hydrology 273 (2003) 164176

173

Fig. 5. Distinction levels for different stages of a change from arable land to pasture with respect to mean (a) surface runoff, (b) groundwater recharge, and (c) streamow. The respective differences in the hydrologic variable are assumed to be signicant if the distinction level is $0.9.

Fig. 6. Distinction levels for different stages of a change from pasture to coniferous forest with respect to mean (a) surface runoff, (b) groundwater recharge, and (c) streamow.

forest has to replace pasture on 25% of the catchment area while in the case of a change from pasture to deciduous forest about 35% of the catchment area has to be affected to obtain a signicantly distinct model response (Figs. 6 and 7).

4.3. Model error The main objective of this study is to assess whether, against the background of parameter uncertainty, modelled hydrologic effects of different land

174

K. Eckhardt et al. / Journal of Hydrology 273 (2003) 164176

Binley et al. (1991), the uncertainty range of model output is dened by the 5th and the 95th percentile of the respective distribution. The relative model error E is calculated as E 95th percentile 2 5th percentile=2 mode 6

This formulation is analogous to the denition of the coefcient of variation ( standard deviation/mean value). However, because some of the calculated frequency distributions are skewed, we do not use the mean value and standard deviation to characterise the shape of the calculated distributions. Instead, we use the mode, the most frequent value of the considered output variable, and half the difference between the 5th and the 95th percentile. The results are summarised in Table 4. The calculated errors depend on the considered land cover and output variable. Groundwater recharge shows the maximum error. Errors regarding streamow are of special interest because streamow usually serves to calibrate catchment models.
Table 4 Mode, 5th and 95th percentile of the frequency distributions shown in Fig. 4. The model error indicated in the last column is calculated as the ratio of half the difference between the 5th and the 95th percentile and the mode Land cover Mode 5th percentile 95th percentile Model error (%) Surface runoff (mm/d) FRSD 0.108 0.106 FRSE 0.100 0.093 PAST 0.128 0.121 CROP 0.309 0.294 Groundwater recharge (mm/d) FRSD 0.207 0.197 FRSE 0.174 0.162 PAST 0.364 0.337 CROP 0.274 0.262 Streamow (mm/d) FRSD 1.085 FRSE 1.001 PAST 1.352 CROP 1.700 1.044 0.875 1.304 1.674

0.114 0.105 0.136 0.342 0.245 0.210 0.404 0.281 1.229 1.080 1.407 1.723

3 6 5 8 12 14 9 3 9 10 4 1

Fig. 7. Distinction levels for different stages of a change from pasture to deciduous forest with respect to mean (a) surface runoff, (b) groundwater recharge, and (c) streamow.

covers are signicantly distinct. However, the frequency distributions shown in Fig. 4 also allow potential errors in simulation results caused by parameter uncertainty to be quantied. According to

FRSD: deciduous forest, FRSE: coniferous forest, PAST: pasture, CROP: cropping sequence of summer barley, winter rape and winter wheat.

K. Eckhardt et al. / Journal of Hydrology 273 (2003) 164176

175

The respective error for forest amounts to about 10% while the errors for pasture and arable land are smaller (up to 4%).

5. Discussion and conclusions A methodology has been presented which allows us to investigate whether a model is suitable for simulating effects of land cover changes in the sense that, taking into consideration parameter uncertainty, it yields signicantly distinct results for different covers. The methodology, which we call distinction level estimation, has been applied to the model SWAT-G. SWAT-G proved to be suitable for land cover change studies. Forest, pasture and arable land yield signicantly distinct mean values of surface runoff, groundwater recharge and streamow if the whole catchment is affected by a land cover change. Yet in most cases, land cover will only change in parts of a catchment. The minimum proportion of the catchment area which has to be affected by a land cover change in order to obtain signicantly distinct model responses, has been quantied. Changes between arable land and pasture become noticeable sooner than changes between pasture and forest. Furthermore, a considerably higher proportion of land cover change is required, if pasture is replaced by deciduous forest than if it is replaced by coniferous forest. Our results can be validated to a certain extent by comparing them to observations in catchment experiments. Bosch and Hewlett (1982) reviewed results from 94 experiments throughout the world, most of them dealing with deforestation. The observational data scatters over wide ranges. Bosch and Hewlett (1982) concluded that reductions in forest cover of less than 20% apparently cannot be detected by measuring streamow. This nding has been conrmed by Stednick (1996) who compiled 95 paired catchment studies in the USA reporting on the effects of deforestation on mean streamow. This corroborates our simulation results concerning a change between pasture and Bosch and Hewlett (1982) found that, on average, a 10% change in cover caused approximately a 40 mm change in annual water yield for

coniferous forest and 25 mm for deciduous forest. The corresponding changes that we have simulated with our model of the articial catchment are smaller (13 mm change in annual water yield per 10% change in coniferous forest cover, 10 mm in the case of deciduous forest). However, absolute changes in water yield increase with the regions mean annual precipitation (Bosch and Hewlett, 1982). Most of the compiled experimental studies, though, were carried out in regions with a far higher mean annual precipitation than in the articial catchment. A comparison with Fig. 3 in the paper of Bosch and Hewlett (1982), where water yield changes are shown as a function of mean annual precipitation, reveals that our simulation results are realistic. The overall impression is that the response of the model of the articial catchment and its assessment correspond well to real catchment reactions. The present study also allowed us to quantify an upper limit for errors in simulation results caused by uncertainty in land cover-dependent parameters. In the case of average streamow, this error amounts to up to 10%. If a real catchment is considered, measurements may be used to calibrate the model and hence to diminish potential errors. The error in absolute values of model output can be estimated using the generalised likelihood uncertainty estimation (GLUE; Beven and Binley, 1992) for instance. Apart from the comparison of the model output with observational data, GLUE differs from the distinction level estimation in that all sources of uncertainty (parameterisation, boundary conditions, measurement errors, model structure) are reected, yet there is no way to distinguish their contributions to the overall model error. The distinction level estimation aims specically at parameter uncertainty. Above all, however, the main objective of the new approach is not to calculate overall uncertainty in absolute values of model output but to assess the signicance of differences in model output, due to, e.g. land cover changes, or, in other words, to assess the signicance of model reactions to changes in parameterisation. In this sense, the two methods throw light on two different aspects of model uncertainty and therefore can complement one another.

176

K. Eckhardt et al. / Journal of Hydrology 273 (2003) 164176 Eckhardt, K., Haverkamp, S., Fohrer, N., Frede, H.-G., 2002. SWAT-G, a version of SWAT99.2 modied for application to low mountain range catchments. Phys. Chem. Earth 27, 641 644. Fohrer, N., Haverkamp, S., Eckhardt, K., Frede, H.-G., 2001. Hydrologic response to land use changes on the catchment scale. Phys. Chem. Earth (B) 26, 577 582. Matheussen, B., Kirschbaum, R.L., Goodman, I.A., O Donnell, G.M., Lettenmaier, D.P., 2000. Effects of land cover change on streamow in the interior Columbia River Basin (USA and Canada). Hydrol. Proc. 14, 867 885. Monteith, J.L., 1965. Evaporation and environment. Symp. Soc. Exp. Biol. 19, 205234. Monteith, J.L., 1977. Climate and the efciency of crop production in Britain. Phil. Trans. Res. Soc. Lond., Ser. B 281, 277329. Nandakumar, N., Mein, R.G., 1997. Uncertainty in rainfall-runoff model simulations and the implications for predicting the hydrologic effects of land-use change. J. Hydrol. 192, 211232. Nash, J.E., Sutcliffe, J.E., 1970. River ow forecasting through conceptual models, Part Ia discussion of principles. J. Hydrol. 10, 282290. Stednick, J.D., 1996. Monitoring the effects of timber harvest on annual water yield. J. Hydrol. 176, 7995. Storck, P., Bowling, L., Wetherbee, P., Lettenmaier, D., 1998. Application of a GIS-based distributed hydrology model for prediction of forest harvest effects on peak stream ow in the Pacic Northwest. Hydrol. Processes 12, 889 904. USDA Soil Conservation Service, 1972. National Engineering Handbook, Hydrology, Section 4, US Government Printing Ofce, Washington, DC, Chapters 410. Wechsung, F., Krysanova, V., Flechsig, M., Schaphoff, S., 2000. May land use change reduce the water deciency problem caused by reduced brown coal mining in the state of Brandenburg? Landscape Urban. Plan. 51, 177 189. Williams, J.R., Jones, C.A., Dyke, P.T., 1984. A modelling approach to determining the relationship between erosion and soil productivity. Trans. ASAE 27, 129 144.

Acknowledgements This study has been supported by the Deutsche Forschungsgemeinschaft within the scope of the Collaborative Research Centre (SFB) 299.

References
Arnold, J.G., Srinivasan, R., Muttiah, R.S., Williams, J.R., 1998. Large area hydrologic modelling and assessment, Part I: model development. J. Am. Water Res. Assoc. 34, 7389. Beven, K.J., 1989. Changing ideas in hydrology: the case of physically-based models. J. Hydrol. 105, 157 172. Beven, K.J., Binley, A.M., 1992. The future of distributed models: model calibration and uncertainty prediction. Hydrol. Processes 6, 279 298. 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 Resour. Res. 27, 12531261. Bosch, J.M., Hewlett, J.D., 1982. A review of catchment experiments to determine the effect of vegetation changes on water yield and evapotranspiration. J. Hydrol. 55, 323. Breuer, L., Eckhardt, K., Frede, H.-G., 2003 Review of plant parameter values for hydrological and soilvegetationatmosphere transfer (SVAT) models in temperate climates. Submitted for publication. Internet database: http://www.uni-giessen.de/ ~gh1461/plapada/plapada.html, January, 2003 . Bronstert, A., Fritsch, U., Katzenmaier, D., 2001. Quantizierung des Einusses der Landnutzung und -bedeckung auf den Hochwasserabuss in Flussgebieten unter Berucksichtigung der aktuellen Bodenbearbeitung, des Zustandes der Vegeta tionsdecke und moglicher Klimaanderungen. Rep. 297 24 508, Federal Environmental Agency of Germany. Dunn, S.M., Mackay, R., 1995. Spatial variation in evapotranspiration and inuence of land use on catchment hydrology. J. Hydrol. 171, 4973.

You might also like