Invited Paper

An Integrative Hydrological, Ecological and Economical (HEE) Modeling System for Assessing Water Resources and Ecosystem Production: Calibration and Validation in the Upper and Middle Parts of the Yellow River Basin, China
Xianglian Lia, Xiusheng Yang*a, and Wei Gaob a Department of Natural Resources Management and Engineering, University of Connecticut, 1376 Storrs Road, Storrs, CT 06269-4087, USA b USDA UVB Monitoring and Research Program, Natural Resource Ecology Labortory, Colorado State University, Fort Collins, CO 80523, USA
ABSTRACT
Effective management of water resources in arid and semi-arid areas demands studies that cross over the disciplinaries of natural and social sciences. An integrated Hydrological, Ecological and Economical (HEE) modeling system at regional scale has been developed to assess water resources use and ecosystem production in arid and semi-arid areas. As a physically-based distributed modeling system, the HEE modeling system requires various input parameters including those for soil, vegetation, topography, groundwater, and water and agricultural management at different spatial levels. A successful implementation of the modeling system highly depends on how well it is calibrated. This paper presented an automatic calibration procedure for the HEE modeling system and its test in the upper and middle parts of the Yellow River basin. Previous to calibration, comprehensive literature investigation and sensitivity analysis were performed to identify important parameters for calibration. The automatic calibration procedure was base on conventional Monte Carlo sampling method together with a multi-objective criterion for calibration over multi-site and multi-output. The multi-objective function consisted of optimizing statistics of mean absolute relative error (MARE), Nash-Sutcliffe model efficiency coefficient (ENS), and coefficient of determination (R2). The modeling system was calibrated against streamflow and harvest yield data from multiple sites/provinces within the basin over 2001 by using the proposed automatic procedure, and validated over 1993-1995. Over the calibration period, the mean absolute relative error of simulated daily streamflow was within 7% while the statistics R2 and ENS of daily streamflow were 0.61 and 0.49 respectively. Average simulated harvest yield over the calibration period was about 9.2% less than that of observations. Overall calibration results have indicated that the calibration procedures developed in this study can efficiently calibrate the modeling system in the study area. Annual validation results for average streamflow and harvest yield showed relative large errors which were associated with irrigation water use and reservoir impact. The validation results of streamflow for sites in upper reaches have shown close relationship with observations which indicated the liability of calibrated parameter values in predicting watershed responses. The information and results provided by the study will be helpful to watershed modelers and model users in calibrating complex watershed models and contribute knowledge to interdisciplinary modeling for water resources management in the study area. Keywords: Automatic calibration; Integrated HEE modeling system; Monte Carlo sampling; Sensitivity analysis; Upper and middle parts of the Yellow River basin.

1. INTRODUCTION
Watershed models have been widely used in the past decade due to the advances in computer speed and storage, development of GIS, and availability of comprehensive digital databases. These models are generally characterized by a multitude of parameters. Parameter values of these models are known to have limited scientific foundation unless conditioned on observed data through a process of calibration (Beven, 1993). Therefore, calibration has been necessary to evaluate the ability of a watershed model in sufficiently predicting constituent yields and streamflow for a specific application. Due to the large number of parameters, therefore inevitable uncertainty, of a distributed watershed model, a calibration process usually involves sensitivity analysis of model parameters, and followed by model validation and uncertainty
*

Corresponding author: Xiusheng Yang, Professor; [email protected]; phone 1 860 486 0135; fax 1 860-486-5408
Remote Sensing and Modeling of Ecosystems for Sustainability III, edited by Wei Gao, Susan L. Ustin, Proc. of SPIE Vol. 6298, 62982I, (2006) 0277-786X/06/$15 doi: 10.1117/12.680713

Proc. of SPIE Vol. 6298 62982I-1

analysis. Sensitivity analysis is to estimate the rate of change in the model output with respect to changes in model inputs. The most sensitive parameters identified in sensitivity analysis are often used to calibrate a model. Model calibration typically involves modification of parameter values and comparison of the simulated and observed outputs of interest until a desired objective is achieved. When calibration objective is achieved, model should be validated over a different period. Model validation uses the estimated parameter values as inputs and compares the simulated and observed outputs to examine whether the model calibrated with a particular dataset or system is representative to other dataset or system. If the objective is not achieved for validation period, calibration and/or model assumptions might be re-inspected. There are a number of techniques available for sensitivity analysis (Hamby, 1994; Saltelli, 2000; Saltelli et al., 2004). These techniques can be broadly grouped as local and global approaches (Saltelli et al., 1999). In local techniques, output responses are determined by changing one input at a time while fixing all other factors to constant nominal values. The further the perturbation moves away from the nominal value, the less reliable the analysis results become (Helton, 1993). The advantage of local approaches is that they are simple and easier to perform, and the disadvantage of local approaches lies in that they do not account for interactions between inputs, if any exists. Also, the more nonlinear the relationship between inputs and output variables, the more difficult and unreliable it is to employ local techniques. Studies use this method can be found in Lenhart et al.(2002). Global sensitivity analysis methods, such as the HornbergerSpearYoung method, examine the entire space of inputs, including their combinations, allowing output responses as a result of all inputs and their possible interaction (Hornberger and Spear, 1981; Young, 1983; Muleta and Nicklow, 2005). Global sensitivity method is actually based on Monte-Carlo and/or Fourier amplitude simulations, in which many different model runs with randomly chosen parameters sets. Sensitivity is usually obtained by comparing the cumulative distributions of one parameter in each set. Sensitive parameters show strong differences between the distributions. However, it has disadvantage of the computational effort, as the number of calculations to describe the parameter space is a power function (Beven, 2001). Equivalent results can be obtained by both approaches as showed by studies of Lenhart et al.(2002) and Muleta and Nicklow (2005). For quantitative evaluation of parameter sensitivity, there are several methods available such as simple regression (Brenkert et al., 1988), multiple and piecewise multiple regression (Downing et al., 1985), regression coefficients and partial regression coefficients (Gardner et al., 1981, Bartell et al., 1986), stepwise regression and correlation ratios and differential sensitivity analysis (Worley, 1987, Griewank and Corliss, 1991). Studies that discuss the use of statistical regressions in evaluating parameters sensitivity are available in Iman et al. (1981a; 1981b), Morgan and Henrion (1990), and Iman and Helton (1991). Methods used for watershed model calibration can be broadly divided into manual and automatic calibration. Manual calibration is the process of adjusting the simulation outputs by changing parameters to match the observed values. For manual calibration, one has to select the parameters to be calibrated, define the variation ranges, consider mutual dependencies between parameters, and formulate criteria for the termination of the optimization. The success of a manual calibration essentially depends on the experience of the modelers and their knowledge of the basic approaches and interactions in the model. Moreover, it can be extremely time consuming. Also, a manual calibration is always subjective to some extent (Botterweg, 1995). Automatic calibration has been employed to improve these shortcomings, in which Monte Carlo based methods have recently become popular for calibration of hydrological models. Monte Carlo analysis has also been used as a basis for recursively estimating parameters to indicate model structure errors (Wagener et al., 2003 and Smith and Wheater, 2004) and assessing significance of input and calibration data errors (McIntyre and Wheater, 2004). Kalos and Whitlock (1986) gave an overall review of the methods of simplest form of Monte Carlo analysis (plain random sampling of model input parameters from uniform prior distributions). The method involves random sampling from the distribution of inputs and successive model runs until a statistically significant distribution of outputs is obtained (Sobol, 1994; Fishman, 1996). Since the method requires a large number of samplings, the computational time could be prohibitively expensive. The Latin Hypercube Sampling is one widely used modified Monte Carlo method which greatly reduces the number of simulation (Stein, 1987; Loh, 1996). In this method, the whole parameter space, consisting of all the uncertain parameters, is partitioned into cells having equal probability, and they are sampled in an efficient manner such that each parameter is sampled once from each of its possible segments. The advantage of this approach is that the random samples are generated from all the ranges of possible values, thus giving insight into the extremes of the probability distributions of the outputs. The method has been applied in probabilistic assessment of sensitivity of model parameters in hydrological simulation(Gwo et al., 1996; Liou and Yeh, 1997; Rong et al., 1998). EPA (Environmental Protection Agency, 1996) also has a detailed guidelines for the application Latin Hypercube Sampling methods in risk assessment. Most of the applications of Monte Carlo analysis have been to relatively simple, lumped watershed models, generally with fewer than ten uncertain inputs (e.g. Smith and Wheater,

Proc. of SPIE Vol. 6298 62982I-2

2004), or to distributed models but only varying a small subset of the parameters (e.g. Anderton et al., 2002). Study with a relatively large number of parameters was available in Meixner et al. (1999). They demonstrated the benefit of measuring sensitivity using multiple model outputs, rather than just one aggregated measure of output, and that surprisingly few (500) parameter samples were needed to identify the main univariate sensitivities. The application of Monte Carlo analysis to SWAT model calibration can be found in Eckhardt et al. (2003) and White and Chaubey (2005). In this paper, automatic calibration of a HEE modeling system is considered. The integrated hydrological, ecological, and economical (HEE) modeling system has been developed to address water management concerns at regional scale, with a specific focus on basin dominated by irrigation practices (Li and Yang, 2005). As a newly developed modeling system, evaluation of the inherent errors and model uncertainties, calibration and validation is essential. Therefore, the objectives of this study are to: (1) perform sensitivity analysis to identify important parameters for calibration of the HEE modeling system; (2) present a framework of automatic calibration with multi-site and multi-output to increase the efficiency and accuracy of the modeling results; and (3) conduct calibration and validation with outputs of streamflow and harvest yield at different sites in the upper and middle parts of the Yellow River basin to illustrate the use of the proposed framework. The results from this study are expected to be helpful in understanding calibration of complex watershed model and to contribute scientific knowledge to interdisciplinary modeling for water resources management in the study area.

2. THE STUDY AREA

The study area was the upper and middle parts of the Yellow River basin, a typical arid and semi-arid area in North China (95 - 114 E 32 - 42 N), with a drainage area of 772,000 km2 (Figure 1). Annual precipitation in the basin ranges from around 290 mm in the high plateau to over 670 mm in the southeastern basin. Based on a 1 1 km digital elevation model (DEM), the study area was divided into 119 subbasins and 119 streams by BASINS/SWAT interface. Together with land use and soil information, 1021 hydrologic response units (HRU) were generated which represent the smallest spatial homogenous units in hydrologic behaviors. Twelve different land use types were distinguished. Major land use types are pasture, about 55% of the total area, agricultural land use, about 27.8%, forest, about 6.5% and other land use. Major soil types include cultivated loessial soils (18%), cinnamon soils (9.5%), castanozems (6.7%), and Felty soils (5.67).

Figure 1. Upper and middle parts of the Yellow River basin showing subbains, streams, monitoring gages and reservoirs.

3. METHODOLOGIES
3.1. The HEE modeling system The integrated hydrological, ecological, and economical (HEE) modeling system aims to evaluate resources use, agriculture management, and rural development at regional scale. The modeling system adapted SWAT, the Soil and Water Assessment Tool (Arnold et al., 1998), as hydrological components to simulate the water balance in terms of soil moisture, evapotranspiration, streamflow, and groundwater table change. Its ecological module was integrated into the hydrological module to compute the ecosystem production of biomass and yield for different land use types. And economical module estimated the monetary values of crop yield and other ecosystem services. The model was implemented in a holistic approach, and able to produce simulation results at daily time steps with a spatial resolution of hydrological response unit (HRU). The modeling system is driven by daily climatic data and simulates dynamic and accumulative hydrological, ecological, and economical variables for each hydrological response unit. The outputs of the modeling system include basic water balance components, harvest yield and crop production at different spatial and temporal resolution. More information about the HEE modeling system can be found in Li and Yang (2005). 3.2 Parameter screen 1) Literature investigation and parameterization

Proc. of SPIE Vol. 6298 62982I-3

There were over three hundred parameters, mainly from the adapted hydrological model, in the integrated modeling system. By comprehensive investigation of literature related to the hydrological model, parameter number was reduced to 27 that were relatively important to the estimation of streamflow, biomass and crop production (Table 1). The first 11 parameters were assumed uniform value over the entire basin, while the rest parameters could differ among each subwatershed, depending on soil type, land use, topographic features, and/or management. The estimation of the parameters in Table 1 was still discouraging. Parameterization method was used to further reduce the number of parameters to be analyzed. Parameterization was a technique for transferring model parameters of a given spatial unit to other spatial units in the watershed (Muleta and Nicklow, 2005). For this study, a representative spatial unit was selected in which the model assumed homogeneity of parameters and variables. Available information about the parameters was used to develop relationships between required parameters of representative unit and corresponding parameters of other homogeneous units (e.g. HRUs). Some assumptions about the spatially varied parameters were made to further reduce the parameter numbers. For example, Mannings number for channels were broadly grouped. Table 1. Parameters identified by literature investigation
Name
SMFMX SMFMN MSK_CO1 MSK_CO2 BLAI CHTMX RDMX T_OPT T_BASE GSI HVSTI CH_N1 CH_N2 CH_COV ALPHA_BNK GW_DELAY ALPHA_BF GWQMN GW_REVAP REVAPMN RCHRG_DP SOL_AWC CN* OV_N LAT_TTIME ESCO EPCO

Full name
Maximum melt rate for snow during where C pertains to the air temperature. Minimum melt rate for snow during the year (occurs on winter solstice) Calibration coefficient used to control impact of the storage time constant for normal flow Calibration coefficient used to control impact of the storage time constant for low flow Max leaf area index. Max canopy height. Max root depth. Optimal temp for plant growth. Min temp plant growth. Max stomatal conductance (in drought condition). Potential harvest index Manning's "n" value for the tributary channels. Manning's "n" value for the main channel. Channel cover factor. Baseflow alpha factor for bank storage. Groundwater delay. Baseflow alpha factor Threshold depth of water in the shallow aquifer required for return flow to occur. Groundwater "revap" coefficient. Threshold depth of water in the shallow aquifer for "revap" to occur. Deep aquifer percolation fraction. available water capacity SCS runoff curve number for moisture condition II. Manning's "n" value for overland flow. Pasture Lateral flow travel time. Soil evaporation compensation factor. Plant uptake compensation factor.

Range Min.
0 0 0 0 0.5 0.1 0 11 0 0 0 0.01 0.01 -0.001 0 0 0.001 0 0.02 0 0 0 39 0.05 0 0 0

Max.
10 10 10 10 10 20 3 38 18 5 1 30 0.15 1 1 500 1 5000 0.2 500 7 1 68 0.3 180 1 1

*: CN2 or CNOP. It is CN2 in initial simulation and will be CNOP when there is operation. 2) Sensitive analysis Though the number of parameters has been reduced greatly through parameter screen, the number of the parameters was still large particularly for a large basin like the Yellow River basin, where long period data record was usually inadequate. Its essential to reduce the number of calibratable parameters as much as possible. Therefore, sensitivity

Proc. of SPIE Vol. 6298 62982I-4

analysis was conducted to determine the influence of a set of parameters had on simulated total streamflow and harvest yield. To perform sensitivity analysis for the modeling system, plain random sampling method of was used. Parameters were firstly generated according given distribution. Then, the modeling system was called with the different parameter sets, and outputs were accumulated and averaged for each run. The loops stopped when the differences of average outputs between two consecutive loops were less than 0.1%. Spearman's rank correlation method was used to evaluate the importance of each parameter. The rank correlation method is nonparametric (distribution-free) rank statistic as a measure of the strength of the associations between two variables (Lehmann and D'Abrera, 1998). The Spearman rank correlation coefficient can be used to give an R-estimate to measure the association strength between two variables. The Spearman rank correlation coefficient is defined by:
n

r' 1 6
i 1

d2 N ( N 2 1)

(1)

where r' is the rank correlation coefficient [1-,1], d the difference in statistical rank of corresponding variables (an approximation to the exact correlation coefficient), N the total number of ranks. Important parameters are defined as those with coefficient large than 0.10 indicating significant effect on the streamflow and/or the harvest yield over the simulation period. These identified important parameters will be estimated in calibration procedure. 3.3 Calibration of the modeling system 1) Calibration procedure
Start

The calibration procedure is shown by flowchart in Figure 2. Firstly, the calibration control parameters, such as subbasin numbers, simulation period and parameter ranges, and observations of streamflow and harvest yield were read in for each corresponding spatial unit. Secondly, parameter space were generated according to given distribution and the random combinations of these parameters were sampled based on conventional MonteCarlo random sampling method, with which parameter set was selected randomly from generated parameter space. Input files were then modified with each parameter combination and the HEE modeling system was executed. After that, simulated outputs were read and compared with corresponding observations. Then, an objective function, defined in next section, was estimated and the estimated statistics were written out for each run. Parameters, with which largest statistic values of fitness and smallest error of output comparison were obtained for a calibration site, were selected as optimized values and fixed for subbasins above the calibration site. Finally, above steps were repeated for calibration of other sites in lower reaches. Due to the complexity of natural conditions in the study area, the modeling system was calibrated from upper to lower reaches. The method was integrated as part of the implementation of the modeling system.

Read control parameters and observations Read data for sub-basin (sub-k) from upper reach to lower reach: k = 1 to L Generate valid parameter space:
j (j=1

to M)

Read j and change parameters in each input file with j Call HEE j+1 Read model outputs; calculate objective function; and write out No j>M? Yes The optimal values of parameter set are the one with the smallest AMRE, largest values ENS, R2 Fixed the parameters in current sub-basin and repeat the steps for the sub-basin in lower reach. k>L? Yes End No

k+1

Figure 2. Calibration flowchart for the modeling system in the upper and middle parts of the Yellow River basin

2) Evaluation criteria and objective function Available statistic methods used in evaluating watershed model performance include: mean and standard deviation (SD) of the simulated outputs, relative error (RE), coefficient of determination (R2), and Nash-Suttcliffe simulation efficiency (ENS) (Nash and Sutcliffe, 1970). In general, R2 value is an indicator of strength of relationship between the observed and

Proc. of SPIE Vol. 6298 62982I-5

simulated values. ENS indicates how well the plot of observed versus simulated value fits the 1:1 line. If the values of R2 and ENS are less than or very close to zero, prediction of the modeling system is considered unacceptable or poor. If the values of R2 and ENS are close to one, then prediction of the modeling system is near perfect. However, there are no explicit standards specified for assessing the model prediction using these statistics. In this study, we use the mean absolute relative error, coefficient of determination and Nash-Sutcliffe efficiency coefficient to measure the model performance. The definitions of these statistics are: (1) Mean absolute relative error (MARE):
n

MARE
i 1

X i' / X i 1 / n

(2)

where X is observation, X' the simulation, i the record and n the total number of record. The optimal statistical value occurs when the value does reach zero. This criterion is applicable to both streamflow and harvest yield. (2) Coefficient of determination (R2):
2 n

(Qi R2
i 1 n

Q )(Q
2 n i 1

' i

Q)

'

(3)
(Q
' i

(Qi
i 1

Q)

Q)

' 2

where Q is observed streamflow, Q' the simulated streamflow Q bar and Q' bar are the mean observed and simulated streamflow respectively. All other variables have the same meaning as in Equation (1). The optimal statistical value occurs when the value reaches 1. This statistics measure the goodness of fit of simulations and observations. (3) Nash-Sutcliffe efficiency coefficient (ENS):
n

(Qi E NS 1
i 1 n

Qi' ) 2
(4)

(Qi
i 1

Q)

All variables have the same meaning as in Equations (1) and (2). The optimal statistical value occurs when the value reaches 1. With the statistic criteria above, the objective function is to:
T n

Min
t 1 j 1

( MARE qjt
n

MARE yjt ) and

(5)

Max
j 1

(R

2 qj

E NSj )

where, t stands for the calibration year, j the site, MAREq and MAREyd the mean absolute relative error for streamflow and harvest yield. 3.4 Setup of calibration and validation of the modeling system The HEE modeling system was calibrated against daily streamflow measurements of year 2001 at eight monitoring sites (Figure 1) and annual harvest yield for six provinces for the year. Sorooshian and Gupta (1995) pointed out that the length of data should be at least 20 times that of the number of calibrable parameters. For this study, the length of the data at all sites in the basin is 1 year (365 samples) for streamflow and 1 year for yield. Since there are 12 calibrable parameters, the available data were sufficient for streamflow calibration. Land use and soil data were from the Institute of Geography and Natural Resources Research (IGANRR), Chinese Academy of Sciences (CAS). Land uses were assumed unchanged for the whole simulation period. Water use data were got from statistics of Yellow River basin report for water use at provincial level. In particular, irrigation was set according to local agricultural management practices, and water for other use was assumed to be based on monthly statistics. Three reservoirs with daily measured discharges were set to be operated while other reservoirs without measurements were set according to annually average

Proc. of SPIE Vol. 6298 62982I-6

discharges or were not triggered. Please be advised that, because the initial conditions of the modeling have to be stabilized, therefore the model have been ran for 3 years (1999-2001) but only the last year is taken into consideration. With parameters estimated during the calibration process, the modeling system was validated over the period of 19931995. Land use and soil data were not changed while water use data were set according to annual water use reports. Reservoirs were not triggered due to absent of discharge information. Outputs of monthly streamflow and annually harvest yield were compared with observations for different validation sites. Scattering plot and statistic measurements of modeling efficiency coefficient and coefficient of determination were used to evaluate the performance.

4. RESULTS AND DISCUSSION

4.1 Parameter sensitivity of the modeling system Parameter sensitivity analysis has identified 12 important parameters for calibration as shown in Table 2 below. The most sensitive parameters for streamflow and harvest yield were soil available water content (SOL_AWC) and harvest index (HVSTI) respectively. The modeling system showed less sensitive to curve number (CN) and Manning's "n" value for the tributary channels (CH_N1). Both streamflow and harvest yield are sensitive to parameters of maximum leaf area index (BLAI), maximum root depth (RDMAX), and soil compensate factor (ESCO). Table 2. Sensitivity ranking of parameters based on ranked correlation coefficient statistic # Parameters r'-streamflow r' -yield 1 SMFMN 0.13 2 BLAIa -0.11 0.55 3 RDMXa -0.16 0.11 4 T_OPTa -0.31 5 T_BASEa -0.41 6 HVSTIa 0.57 7 CH_N1 -0.10 8 GWQMN -0.20 9 SOL_AWCa -0.93 10 CN2 0.08 11 ESCO 0.14 0.25 12 EPCO 0.29 a: for representative land use of pasture and soil type of cultivated loessial soil.

Some of above parameters were compared with these identified in SWAT documentation or other studies relevant to SWAT. Parameters, such as curve number, soil available water content, maximum leave area index, optimal and base temperatures, were also identified as sensitive parameters in modeling watershed water yield in studies of Srinivasan et al., (1998) and Lenhart et al. (2002). Some studies also found parameters, such as maximum stomatal conductivity and channel width, were important in water yield estimation while they were not in this study (Arnold and Allen, 1999; Lenhart et al. 2002). Curve number was broadly recognized as an important parameter for calibration in many of the studies. Also, it is a key parameter in estimation surface runoff. Therefore, curve number was listed as an important parameter for calibration though its ranked correlation coefficient was less than one. All parameters in Table 2 were modified during calibration. 4.2 Calibration results 1) Calibrated parameter values of the modeling system The calibrated parameter values for basin level parameters and the calibrated parameter values ranges of subbasin/HRU level parameters are listed in Table 3. The calibrated parameters of minimum melt rate for snow on winter solstice (SMFMN) was 0.82. The optimized values for vegetation parameters of maximum leave area index (BLAI), maximum root depth (RDMAX), optimal temperature (T_OPT), base temperature (T_BASE) and harvest index (HVSTI) for representative vegetation of pasture grass were 4.66, 1.82, 28.18, 13.86, and 0.68 respectively. Manning's "n" value for the tributary channels ranged from 0.03 to 1.14 for different reaches. Groundwater parameter (GWQMIN) was 8.74 to 83.93 for different subbasins. Curve number (CN) was from 59 to 77 for representative land use of pasture. Soil evaporation compensation factor (ESCO) varied from 0.08 to 0.66 for representative soil type of cultivated loessial soil and plant

Table 3. Calibrated parameter values, 2001 # Parameters Parameter values/ranges 1 SMFMN 0.82 2 BLAIa 4.66 3 RDMXa 1.82 4 T_OPTa 28.18 5 T_BASEa 13.86 6 HVSTIa 0.68 7 CH_N1 0.03-0.14 8 GWQMN 8.74-83.93 9 SOL_AWCa 0.18-0.50 10 CNa 59-77 11 ESCO 0.08-0.66 12 EPCO 0.25-0.37 a: for representative land use of pasture and soil type of cultivated loessial soil.

Proc. of SPIE Vol. 6298 62982I-7

uptake compensation factor (EPCO) was from 0.25 to 0.37 for representative land use of pasture. Soil available water capacity (SOL_AWC) was from 0.08 to 0.50 for cultivated loessial soil. Parameter values listed here were for the representative land uses or soils. The actual parameter values varied greatly for different land uses, soil types and/or subbasins except for basin level and vegetation parameters. The range of the calibrated parameter values implied the diversity of climate, topography and soil within the basin. In most other studies of automatic model calibration, there was only one optimized value for a particular parameter of a representative unit. All other parameter values would be derived from this representative parameter value according to pre-assumed relationship. In this study, geographic properties of the basin varied greatly in spatial extent which required a wide variety of parameters existing in modeling the basin. Therefore, a set of optimal parameters was estimated for subbasins and HRUs above each calibration site. In this point of view, the procedure proposed in this study can be seen as semi-automatic calibration. 2) Hydrological outputs The overall statistics of the daily comparisons between the simulated and observed streamflow for the calibration sites are shown in Table 4. The mean absolute relative errors between the simulated and observed average daily streamflows for the calibration sites were 1.15% ~ 9.08% with an average of 6.71%. R2 values for the regression of simulated values against observations for the eight calibration sites were from 0.43~0.74 with a mean value of 0.61. The efficiencies ENS for the calibration sites ranged from 0.34 to 0.71 with a mean value of 0.49. In this study, R2 and Nash-Sutcliffe values over 0.5 and 0.4 respectively were considered acceptable given the large impact of human activities in the study area. In general, the daily calibration results for the study area satisfied the requirements for these statistics except Nash-Sutcliffe values for sites of Shizuishan and Toudaoguai where daily reservoir information was not sufficient. The general trend of the simulated daily streamflow for the calibration sites compared well with corresponding observations in year 2001, as shown in Figure 3 for two selected sites. The best fit of the simulated and observed streamflow was observed at sites of Lanzhou and Tangnaihe in upper reaches. Simulated streamflow was relatively less than that of observations in most of calibration sites, an example plot of Tangnaihei was shown in Figure 3, except that of Lanzhou. Generally, the modeling system captured the major trends of the streamflow though it slightly underestimated the extreme streamflow peaks in the calibration period. Table 4. Statistics of the simulated and observed streamflow for calibration sites, 2001
Sites Tangnaihei Lanzhou Xiaheyan Shizuisan Sanhukou Toudaoguai Wubao Longmen Mean Q'a (cms) 397.62 592.90 601.06 520.14 409.72 343.10 353.46 420.16 Mean Q (cms) 435.60 586.14 654.29 571.32 375.62 363.81 370.80 453.18 MARE (%) 8.72 1.15 8.14 8.96 9.08 5.69 4.68 7.29 R2 (-) 0.74 0.77 0.66 0.43 0.64 0.47 0.68 0.50 ENS (-) 0.54 0.71 0.41 0.37 0.58 0.34 0.55 0.41

a: Q': simulated streamflow: Q: observed streamflow There were several other factors which can affect the calibration results. Firstly, operation of reservoirs and water withdrawal from stream changed the natural river environment. The simulated streamflow for Toudaoguai was an example of irrigation water use impact which has resulted in higher estimation in August, September, and October. For better simulation of water behaviors, detailed information of reservoir operation and water use was required, which could not be simply reached by adjusting model parameters. This impact could also be seen from the low statistic values of streamflow comparison at sites of Longmen, Shizuishan, sanhukou, and Toudaoguai as shown in Table 4. Secondly, the streamflows of some small tributary rivers, which contributed streamflow to the calibration sites, were not calibrated. They might affect the results of a particular calibration site locating in lower reaches. Due to the unavailability of observation gages in these tributaries, an individual parameter for a representative unit had to be assumed uniform for all HRUs above a particular calibration site. Thirdly, parameter of minimum melt rate for snow on winter solstice along might not be enough to simulate snowmelt in spring at the river head. It was observed that there was a poor match during spring months with snowmelt as shown in Tangnaihe in Figure 3. To further improve the calibration results, the above factors need to be examined carefully.

Proc. of SPIE Vol. 6298 62982I-8

1400

1400 1200 1000 (cms) 800 600 400 200 0 0

Tangnaihei R2=0.74 obervation simulation

1200 simulation (cms) 1000 800 600 400 200 0

Tangnaihei ENS=0.54

50

100

150

200 Day

250

300

350

500 1000 observation (cms)

1500

1600 Lanzhou 1400 1200 1000 (cms) 800 600 400 200 0 0 50 100 150 200 Day 250 300 350 R2=0.77 obervation simulation

1400 1200 simulation (cms) 1000 800 600 400 200 0 0 500 1000 observation (cms) 1500 Lanzhou ENS=0.71

Figure 3. The observed and simulated daily streamflow for sites of Tangnaihei and Lanzhou in the upper parts of the Yellow River basin, 2001. 3) Ecological production The comparisons of simulated and observed unit harvest yield for each province in 2001 are shown in Table 5. The unit harvest yields were underestimated basin-widely with an average relative error of 9.2%. The relative error of the simulated and observed unit harvest yield for Shanxi, Shaanxi and Gansu were less than 10% while those for Qinghai, Ningxia, and Inner Mongolia were more than 10%. Ecological production for other land uses such as grassland and forest was not calibrated due to difficulties in obtaining accurate observations.

Table 5. Calibration statistics for unit harvest yield for year 2001 Observation Simulated MARE Province (kg/ha) (kg/ha) (%) Shanxi 2650.6 2473.0 6.70 Shaanxi 2632.2 2382.1 9.50 Gansu 2420.3 2342.8 3.20 Inner Mongolia 3707.8 3299.9 11.00 Ningxia 2627.9 2262.3 13.91 Qinghai 1635.3 1457.1 10.90

The accuracy of simulated harvest yield has been improved by adjusting model parameters in automatic calibration. However, yield estimation was also affected by other inputs, such as new crop variety, which were hard to be quantified in the current modeling system. Also, plant areas changed every year which was also hard to specify in the current modeling system. Therefore, in addition to parameter adjustment, the simulation capability of the modeling system also needs to be improved to capture the impact from the above mentioned factors. 4) Automatic calibration efficiency The calibration efficiency obtained by means of the automatic calibration in the study area was acceptable. By using the relative errors, R2 and ENS as objective function, more than 2400 simulation runs were carried out. The computational time depends on the division of hydrological response units, simulation period, number of output variables, and the

Proc. of SPIE Vol. 6298 62982I-9

computer capability as well. In this study, it took about 3 days to optimize calibration parameters for 1021 HRUs, for a simulation period of 3 years on a Pentium IV computer with Windows XP. Major advantages of the automatic calibration developed in this study over manual calibration are that it is time-saving and easy to repeat. The time for both automatic and manual calibrations in one complete run is the same. However, the process of manually comparing outputs for each run really takes time while that of automatic calibration can be done in just minutes. With a little modification or without modification, the automatic calibration procedure can be easily used for model calibration or similar studies. For example, once a model or modeling system is automatically calibrated, the model users will be free of other work, such as further improvement of the model or modeling system. Therefore, the automatic calibration procedure in this study is an efficient tool to both modelers and users. 4.3 Validation results Statistics of annual and monthly validation results over 1993-1995 are shown in Table 6. Mean absolute relative errors for streamflow and harvest yield were somewhat high which were believed to be associated with irrigation water use, reservoir impact and peak flow simulation error. However, the validation results for the sites in very upper reaches, where there was less impact from human activities, Table 6. Validation statistics for streamflow and harvest showed a close relationship between measured output and yield, 1993-1995 simulated output as indicated by R2 and ENS of 0.68 and 0.37, respectively. The validation results for sites in upper For the study area For Tangnaihei Outputs & Lanzhou reaches indicated that the calibrated parameter values ENS R2 ENS MARE* (%) R2 were acceptable in predicting watershed response for Streamflow 14 0.44 0.13 0.68 0.37 outputs of streamflow and harvest yield given available Harvest 22 information. Moreover, validation results in Table 6 were yield expected to be improved when more detailed information * MARE is annual output statistics; R2 and ENS are monthly regarding irrigation water use and reservoirs discharge output statistics. were available.

5. CONCLUSION
This study presented an automatic calibration procedure for multi-site and multi-output for an integrated Hydrological, Ecological and Economical (HEE) modeling system, and examined the procedure in the upper and middle parts of the Yellow River basin, China. Before calibration, twelve important parameters were identified as sensitive parameters to the HEE modeling system by broad literature review and sensitivity analysis. The automatic calibration results of daily streamflow for eight sites and annual harvest yield for six provinces showed fair relationship with that of observations in the study period. The general trend of the simulated streamflow for the calibration sites also compared well with corresponding observations. The automatic calibration procedure has proved to be time-saving in processing outputs comparison and finding optimal parameter values for the modeling system as shown by the time used for modeling system calibration and its results. Validation results for sites in upper reaches have shown close relationship with observations though that for sites in middle reaches did not provide similarity with observations. The large error of validation results for sites in middle reaches indicated that factors, such as water use and reservoir operation, needed to be set up carefully to further improve the validation results. Future work needs to be focused on uncertainty analysis to further examine the performance of the modeling system in the study area.

REFERENCES [1]. Anderton, S., J. Latron, and F. Gallart, 2002. Sensitivity analysis and multi-response, multi-criteria evaluation of a physically based distributed model. Hydrological Processes, 16: 333-353 [2]. Arnold, J. G., R. Srinivasan, R. S. Muttiah and J. R. Williams, 1998. Large area hydrologic modeling and assessment part I: model development. Journal of the American Water Resources Association, 34(1): 73-89. [3]. Arnold, J.G. and P.M. Allen, 1999. Automated methods for estimating baseflow and ground water recharge from streamflow records. Journal of the American Water Resources Association, 35(2): 411-424. [4]. Bartell S.M., J.E. Breck, R.H. Gardner, 1986. Individual parameter perturbation and error analysis of fish bioenergetics models. Canadian Journal of Fisheries and Aquatic Sciences 43: 160-168. [5]. Beven, K.J., 1993. Prophecy, reality and uncertainty in distributed hydrological modeling, Advances in Water Resources, 16: 41-51

Proc. of SPIE Vol. 6298 62982I-10

[6]. Beven, K.J., 2001. In: Rainfall Runoff Modelling. The Primer Wiley. Chichister, England. pp: 217 225. [7]. Botterweg, P., 1995. The users influence on model calibration results: an example of the model SOIL, independently calibrated by two users. Ecological Modeling, 81: 7181. [8]. Brenkert, A.L., R.H. Gardner, S.M. Bartell, and F.O. Hoffman, 1988. Uncertainties associated with estimates of radium accumulation in lake sediments and biota. In: Reliability of Radioactive Transfer Models, G. Desmet, ed. Commission of the European Communities, Elsevier Applied Science, London. pp: 185-192. [9]. Downing, D.J., R.H. Gardner, and F.O. Hoffman, 1985. An examination of response-surface methodologies for uncertainty analysis in assessment models. Technometrics, 27: 151-163. [10]. Eckhardt, K., S. Haverkamp, N. Fohrer, and H.-G. Frede, 2002. SWAT-G, a version of SWAT99.2 modified for application to low mountain range catchments. Physics and Chemistry of the Earth, 27: 641-644. [11]. Environmental Protection Agency, U. S., 1996. Summary report for the workshop on Monte Carlo Analysis. EPA/630/R-96/001. [12]. Fishman, G. S., 1996. Monte Carlo Concepts, Algorithms and Applications. Springer-Verlag, New York Inc, 698 pp. [13]. Gardner R, R. ONeill, J. Manakin, J. Carney, 1981. A comparison of sensitivity analysis and error analysis based on a stream ecosystem model. Ecol Model 12: 173-190. [14]. Griewank, A. and H. Corliss, (Eds.), 1991. Automatic differentiation of algorithms: theory, implementation, and application. Philadelphia, PA. Society for Industrial and Applied Mathematics. [15]. Gwo, J.P., L.E. Toran, M.D. Morris, and G.V. Wilson, 1996. Subsurface stormflow modeling with sensitivity analysis using a Latin-Hypercube Sampling technique. Ground Water, 34(5): 811-818. [16]. Hamby D.M., 1994. A review of techniques for parameter sensitivity analysis of environmental models. Environmental Monitoring and Assessment, 32(2): 135-154. [17]. Helton, J.C., 1993. Uncertainty and sensitivity analysis techniques for use in performance assessment for radioactive waste disposal. Reliability Engineering and System Safety, 42(2-3): 327-367. [18]. Hornberger, G.M., and R.C. Spear, 1981. An approach to the preliminary analysis of environmental systems. Journal of Environmental Management, 12: 7-18. [19]. Iman, R. L., J.C. Helton, and J.E. Campbell, 1981a. An approach to sensitivity analysis of computer models, part 1. introduction, input variable selection and preliminary variable assessment. Journal of Quality Technology, 13(3):174-183. [20]. Iman, R. L., J.C. Helton, and J.E. Campbell, 1981b. An approach to sensitivity analysis of computer models, part 2. ranking of input variables, response surface validation, distribution effect and techniques synopsis. Journal of Quality Technology, 13(4):232-240. [21]. Iman, R.L., and J.C. Helton, 1991. The repeatability of uncertainty and sensitivity analyses for complex probabilistic risk assessments. Risk Analysis 11(4):591-606. [22]. Kalos, M.H., and P.A. Whitlock, 1986. Monte Carlo Methods, Vol. I: Basics. John Wiley and Sons, New York. [23]. Lehmann, E. L. and D'Abrera, H. J. M. Nonparametrics, 1998. Statistical Methods Based on Ranks, revised. Englewood Cliffs, NJ: Prentice-Hall, 323pp [24]. Lenhart, T., K. Eckhardt, N. Fohrer, and H.G. Frede, 2002. Comparison of two different approaches of sensitivity analysis. Physics and Chemistry of the Earth, Parts A/B/C, 27(9-10): 645-654. [25]. Li, X., and X. Yang, 2005. Integrative modeling of hydrological, ecological, and economical systems for water resources management at river basin scale. Proceedings of SPIE: Remote Sensing and Modeling of Ecosystems for Sustainability, Vol.5884-6: 47-57. [26]. Liou, T.S., and H.D. Yeh, 1997. Conditional expectation for evaluation of risk groundwater flow and solute transport - one-dimensional analysis. Journal of Hydrology, 199(3-4): 378-402. [27]. Loh, W.L., 1996. On Latin Hypercube Sampling. The Annals of Statistics, 24: 2058-2080 [28]. McIntyre, N.R., and H.S. Wheater, 2004. Calibration of an in-river phosphorus model: prior evaluation of data needs and model uncertainty, Journal of Hydrology, 290: 100 - 116 [29]. Meixner, T., H.V. Gupta, L.A. Bastidas, and R.C. Bales, 1999. Sensitivity analysis using mass flux and concentration, Hydrological Processes, 13: 2233-2244 [30]. Morgan, M.G., and M. Henrion, 1990. Uncertainty, a guide to dealing with uncertainty in quantitative risk and policy analysis. Cambridge University Press, New York. [31]. Muleta, M.K., and J.W. Nicklow, 2005. Sensitivity and uncertainty analysis coupled with automatic calibration for a distributed watershed model. Journal of Hydrology, 306: 127-145

Proc. of SPIE Vol. 6298 62982I-11

[32]. Nash J.E., and J.V. Sutcliffe, 1970. River flow forecasting through conceptual models - Part I: A discussion of principles. Journal of Hydrology, 10: 282-290. [33]. Rong, Y., R.F. Wang, and R. Chou, 1998. Monte Carlo simulation for a groundwater mixing model in soil remediation of tetrachloroethylene. Journal of Soil Contamination, 7(1): 87-102. [34]. Saltelli, A., 2000. What is sensitivity analysis? In: K.C. a. E.M.S. A. Saltelli (Ed.). Sensitivity Analysis, Wiley, New York. [35]. Saltelli, A., S. Tarantola, F. Campolongo, and M. Ratto, 2004. Sensitivity analysis in practice: a guide to assessing scientific models. John Wiley & Sons, Ltd, Joint Reseach Center of the European Commission, Ispra, Italy. [36]. Saltelli, A., S. Torantola, and K.P.S. Chan, 1999. A quantitative model-independent method for global sensitivity analysis of model output. Technometrics, 41(1): 39-56. [37]. Smith, R.M.S, and H.S. Wheater, 2004. Multiple objective evaluation of a simple phosphorus transfer model, Hydrological Processes, 18: 1703 1720. [38]. Sobol, I.M., 1994. A primer for the Monte Carlo method. CRC Press LLC, 107 pp. [39]. Srinivasan, R., T.S. Ramanarayanan, J.G. Arnold, and S.T. Bednarz, 1998. Large area hydrologic modeling and assessment Part II: model application. Journal of the American Water Resources Association, 34(1): 91-101. [40]. Stein, M. 1987. Large sample properties of simulations using latin hypercube sampling. Technometrics 29: 143151. [41]. Wagener, T., 2003. Evaluation of catchment models. Hydrological Processes, 17(16): 33753378 [42]. White, K. L. and I. Chaubey, 2005. Sensitivity analysis, calibration, and validations for a multisite and multivariable swat model. Journal of the American Water Resources Association, 41(5): 1077-1089. [43]. Worley, B.A., 1987. Deterministic uncertainty analysis. ORNL-6428. Oak Ridge National Laboratory, Oak Ridge, Tennessee. [44]. Young, P.C., 1983. The validity and credibility of models for badly defined systems. In: M. B. Beck and V. Straten (Ed.). Uncertainty and forecasting of water quality, Springer-Verlag, New York. 69-98.

Proc. of SPIE Vol. 6298 62982I-12