Document Type : Research Paper
Authors
^{1} University of Hormozgan
^{2} University of Tehran
^{3} Soil Conservation and Watershed Management Research Institute Tehran
Abstract
Keywords
Runoff simulation using SWAT model and SUFI2 algorithm
(Case study: Shafaroud watershed, Guilan Province, Iran)
E. Taghvaye Salimi^{1, 5*}, A. Nohegar^{2}, A. Malekian^{3}, M. Hosseini^{4}, A. Holisaz^{1}
1. Department of Range and Watershed Management, University of Hormozgan, BandarAbbas, Iran
2. Department of Environmental PlanningManagement and Education, University of Tehran, Tehran, Iran
3. Department of Range and Watershed Management, University of Tehran, Karaj, Iran
4. Soil Conservation and Watershed Management Research Institute (SCWMRI), Tehran, Iran
5. Department of Range and Watershed Management, University of Guilan, SowmehSara, Iran
* Corresponding author’s Email: edristaghvaei@guilan.ac.ir
(Received: May. 26.2015 Accepted: Nov. 11.2015)
ABSTRACT
Reliable estimates of runoff are required as a part of the information sets that help watershed managers make informed decisions on water resources planning and management. This study was carried out in Shafaroud watershed located in the north of Iran. In order to achieve the best runoff simulation in the study area, first rainfall data of four stations during 1998 to 2011 were collected and combined with other maps of the study area, such as Digital Elevation Model (DEM), land use and soil as input data in the form ofSoil and Water Assessment Tools (SWAT) model. After running the model, the Sequential Uncertainty Fitting (SUFI2) algorithm in SWAT calibration and uncertainty program (SWATCUP) were used to evaluate the data uncertainty and the most accurate simulation. The first three years (19982000) of rainfall data for warmup and the next 7 years (20012007) for the calibration and final 4 years (20082011) were used for the validation period. Finally, with multiple simulations, the uncertainty of the parameters was assessed with Pfactor, Rfactor, and NS coefficients. The results of validation period ( =0.85, NS=0.74) confirmed the potential of SUFI2 algorithm of SWATCUP program for simulating runoff data in the study area.
Key words: Shafaroud watershed, Simulation, SUFI2, SWATCUP.
INTRODUCTION
More detailed information on the status of rainfall runoff also facilitate decisions on future programs for watershed managers, a step towards the preservation of natural resources for sustainable development. Recently, rainfallrunoff models are widely used with hydrologists to simulate watersheds runoff and play a key role in water resources management (Bilondi et al. 2013). In other hand, several programs and techniques have been developed to reduce parameters uncertainty and achieve to best fit of parameters in the hydrological modeling (Singh et al. 2013).
The SWAT model (Arnoled et al. 1998) is a continuoustime semidistributed hydrological model for application at the watershed scale (Krysanova & Srinivasan 2015). This model has been widely used to land use change effect assessment (Shen et al. 2010; De Girolamo & Lo Porto 2012; Yang et al. 2012; Du et al. 2013; Huang et al. 2013; Niu & Sivakumar 2014; Lin et al. 2015), sediment prediction(Shen et al. 2012; Rostamian et al. 2013), climate change (Andersson et al. 2006; Zhang et al. 2012; Huang et al. 2015), water quality (Debele et al. 2008; Zhang et al. 2011) and simulation of evapotranspiration (Wang et al. 2006). Many computer programs have been developed by hydrologists for parameters uncertainty analysis in river basin model, such as, generalized likelihood uncertainty estimation (GLUE; Beven & Binley 1992), sequential uncertainty fitting (SUFI2; Abbaspouret et al. 2004), parameter solution (ParaSol; Van Griensven & Meixner 2006) and Markov chain Monte Carlo (MCMC; Kuczera & Parent 1998; Vrugt et al. 2008).
The SWATCUP (Abbaspour et al. 2007b) is a computer program that links the Sequential Uncertainty Fitting (SUFI2) algorithm to SWAT model.
Up to now, researchers used SUFI2 algorithm for model calibration and uncertainty analysis of parameters of SWAT model. Narsimlu et al. (2015) in Kunwari River basin applied SUFI2 algorithm in 19year period (19872005) for model calibration, sensitivity and uncertainty analysis. Fukunaga et al. (2015) investigated application of the SWAT hydrologic model to a tropical watershed at Brazil. Nyeko (2015) assessed the capabilities and limitations of SWAT model in modeling watershed that has limited field and hydrologic data for possible use in water resources management. Romanowicz et al. (2005) investigated Sensitivity of the SWAT model to the soil and land use data in the Thyle catchment of Belgium country. Schuol & Abbaspour (2006) used SWAT to simulate water quantity of the four million area in West Africa and applied Sufi2 algorithm on parameters uncertainty. Defersha & Melesse (2012) applied SWAT to evaluate the impacts of land use changes on runoff and sediment yield in the Mara River basin, Kenya. Krysanova & Srinivasan (2015) assessed five projects of different applications of SWAT covering the following themes: impacts of climate change, impacts of land cover change and combined impacts of climate change and human intervention in water management. Bossa et al. (2012) applied the SWAT model in the Republic of Benin, West Africa to evaluate the effects of different soil databases on modeling of hydrological processes and sediment yield. Vilaysane et al. (2015) applied SWAT model to test the capability of the model for predicting stream flow and also used SUFI2 algorithm for calibration and uncertainty analysis in Xedone river basin. Singh et al. (2013) used GLUE and SUFI2 algorithms to simulate daily and monthly streamflow for the period 1993–2002 in the Krishna River basin. Their study revealed excellent correlation during monthly calibration, and good model match between the observed and simulated streamflows. Lin et al. (2015) in their study investigated the effects of land use and land cover changes on runoff response using SWAT model.
They used two different landuse scenarios (1985 and 2006, with reduced forest and increased cropland and urbanized area) in Jinjiang catchment. Shen et al. (2012) used SWAT model to simulate sediment and streamflow in Three Gorges reservoir basin. Their research showed that sediment simulation presented greater uncertainty than streamflow. Yang et al. (2008) tried to find the best uncertainty analysis techniques in Chaohe basin. They compared five algorithms (e.g. GLUE, ParaSol, SUFI2, MCMC and PSO) to a distributed watershed model (SWAT) in north China. In this study, we focused on application of SUFI2 algorithm for prediction of stream flow and uncertainty analysis in the Shafaroud watershed. The main objective of this study is to test feasibility and capability of the SUFI2 algorithm for runoff simulation of the study area, which will contribute to the preservation of natural resources in the Shafaroud watershed and thereby is useful for sustainable development.
MATHERIALS AND METHODS
Study area
Shafaroud watershed is located in Guilan Province at north of Iran, between longitudes 48˚ 39´ 34˝ and 49˚ 8´ 11˝ East and latitudes 37˚ 24´ 58˝and 37˚ 34´ 18˝ north with a drainage area of 336.89 (Fig. 1). The altitude of the catchment ranges from 168 m to 2895 m. The main river with a total length about 40.95 km and located in the north of the catchment.
The numbers of meteorology stations were four stations and discharge data was measured at one gauge, located at the outlet.
The majority of land is used for forest, agriculture and pasture.
SWAT and SWATCUP
Soil and water assessment tools (SWAT) is a semiphysically based model for assessing the impact of management and climate on water supplies, sediment, and agricultural chemical yields in catchments (Narsimlu et al. 2015). In SWAT, a catchment is divided into multiple subcatchments whit hydrologic response units (HRUs) that consist of homogeneous land use, management, topographical, and soil characteristics (Abbaspour et al. 2007a). Each subcatchment is split into multiple hydrological response units (HRUs) based on topography, management, land use and soil types (Wang & Kalin 2011).
SWATCUP is a computer program for calibration of SWAT models. It enables sensitivity analysis, calibration, validation, and uncertainty analysis of SWAT models (Abbaspour et al. 2007b).
Fig. 1. Location of Shafaroud Watershed.
SUFI2 Algorithm
Uncertainty in Sequential Uncertainty Fitting (SUFI2) algorithm is defined as the difference between simulated and observed variables (Rostamian et al. 2013). The uncertainty is determined by the 95% prediction uncertainty band calculated at the 2.5% and 97.5% levels of the output variables (Abbaspour et al. 2004, 2007b).
Pfactor
The Pfactor (percentage of measured data bracketed by the 95% prediction boundary) often named 95PPU (Percentage Prediction Uncertainty). The 95PPU is calculated at the 2.5% and 97.5% levels of the cumulative
Distribution of an output variable obtained through Latin hypercube sampling (Abbaspour 2011). The range of the Pfactor varies from 0 to 1, with values is close to 1 indicating good fitness between simulated and observed values (Yang et al. 2008).
Rfactor
Another measure quantifying the strength of a calibration/uncertainty analysis is the Rfactor, which is the average thickness of the 95PPU band divided by the standard deviation of the measured data. The calibrated parameter ranges can be generated with an acceptable value of the Rfactor and Pfactor.
The Rfactor is given by Eq. (1) (Yang et al. 2008; Narsimlu et al. 2015):
Where and are the upper and lower boundaries of the 95UB and is the standard deviation of the observed data.
NS objective function
NashSutcliffe function has been used for assessment of model performance. This
Function is calculated by using the following equation Eq. (2) (Nash & Sutcliffe 1970):
Where is the groundbased measurements; is the model predicted data and is the mean of the groundbased measurements.
Coefficience
The range of determination coefficient ( ) is 0 to 1 that explain the relationship between
Observed variance and simulated values. The is given by Eq. (3) (Pluntke et al., 2014):
Where and are the observed and simulated values, respectively.
RESULTS
Setup SWAT Model
According to the Soil and Water Assessment Tools (SWAT) model, the following main data was used: landuse, soil characteristics, topography and climate data. First, the raster maps (e.g. topography, landuse, soil) were imported in ArcSWAT 2012 interface.
In the next step, soil and landuse characteristics were overlaid for each subcatchment. In addition, the weather data were defined. Finally, it was ran and simulated a 14year period with 3 years warmup from 1998 through 2011.
Calibration and Sensitivity Analysis
For calibration model we used SWATCUP program with SUFI2 algorithm which can read output data from ArcSWAT interface. In this section, fourteen parameters were selected for calibration that influence streamflow. Sensitively analysis was performed and its results indicated the most sensitive parameters that illustrated in Table 1. According to Table 1, the most sensitive parameters are soil bulk density (SOL_BD) and SCS curve number for moisture condition II (CN2) because of Pvalue close to 0 and tstat bigger than other parameters.
In the next step, model simulated and compared monthly simulated and observed streamflows using SUFI2 algorithm. We calibrated a 7year period from 2001 to 2007 and validated a 4year period from 2008 to 2011. Analysis of hydrographs indicates that the calibrated model slightly underestimate the peak runoff (Fig. 2). The size of uncertainty band (95PPU) is shown in Fig. 2 which confirms the uncertainty is very high. After defining the initial values of the fourteen parameters, it was specified for selecting appropriate parameters ranges. It could be reduce the band of uncertainty. Furthermore, after three iterations with 500 model runs, the best calibration illustrated in Fig. 3, where value was 0.86, Pfactor of 0.51, Rfactor of 0.54 and NS was 0.77. With this calibration, the best ranges of parameters were obtained (Table 2). According to the last calibration, the best parameters values were imported (Table 2) in SWAT model and validated using data set for the period of 2008 to 2011and compared the plot of observed and simulated data.
Table 1. Sensitively analysis of parameters.
Index 
Parameter 
Definition 
t_stat 
pvalue 
Process 
Sensitivity 
1 
ALPHA_BF 
Baseflow alpha factors (1.days^{1}) 
0.29 
0.77 
Groundwater 
very low 
2 
GWQMN 
Threshold depth in shallow aquifer (mm) 
0.53 
0.60 
Groundwater 

3 
HRU_SLP 
Average slope steepness (m.m^{1}) 
0.66 
0.51 
Geomorphology 

4 
OV_N 
Manning’s n value for overland flow* 
0.73 
0.47 
Geomorphology 

5 
SOL_Z 
Soil depth (mm) 
0.79 
0.43 
Soil 

6 
CH_K2 
Channel effective hydraulic conductivity (mm.hr^{1}) 
1.02 
0.31 
Channel 

7 
GW_DELAY 
Groundwater delay (day) 
1.19 
0.23 
Groundwater 

8 
CH_N2 
Manning’s n value for main channel* 
1.22 
0.22 
Channel 

9 
SOL_AWC 
Available water capacity of the soil layer (mm.mm^{1}) 
1.38 
0.17 
Soil 

10 
SOL_ZMX 
Maximum rooting depth of soil profile (mm) 
1.54 
0.12 
Soil 

11 
ALPHA_BNK 
Base flow alpha factor for bank storage (days) 
1.68 
0.09 
Channel 

12 
SOL_K 
Soil conductivity (mm.hr^{1}) 
1.69 
0.09 
Soil 

13 
CN2 
SCS curve number for moisture condition II* 
2.75 
0.01 
Runoff 

14 
SOL_BD 
Soil bulk density (g/ ) 
5.59 
0.00 
Soil 
very high 
*dimensionless
Fig. 2. 95% probability of uncertainty plot and comparing observed and simulated streamflow before calibration.
Table 3 illustrates the values of P and R factors, and NS in calibration (2001 to 2007) and validation (2008 to 2011) periods. Taking an analysis of the catchment at the outlet had a positive correlation with surface runoff, with of 0.85, while Pfactor, Rfactor and NS were 0.63, 0.49 and 0.74 respectively (Fig. 4).
In other words, the evaluation of the hydrograph plot showed good model match in validation period. Also coefficient of determination ( ) value of calibration and validation period showed a good correlation between observed and simulated values (Fig. 5).
Fig. 3. 95% probability uncertainty plot and comparison observed and simulated streamflow after calibration (20012007).
Fig. 4. 95% probability uncertainty plot and comparison observed and simulated streamflow in validation period (20082011).
Fig. 5. Scatter plot of river streamflow for (a) calibration period (20012007) and (b) validation period (20082011).
Table 2. Optimum ranges of parameters.
Parameter_name 
Fitted_value 
Min_value 
Max_value 
CN2 
0.233 
0.193 
0.272 
ALPHA_BF 
0.240 
0.171 
0.297 
GW_DELAY 
130.159 
121.893 
167.1877 
GWQMN 
0.734 
0.783 
0.238 
CH_N2 
0.222 
0.204 
0.245 
CH_K2 
159.381 
133.678 
165.119 
ALPHA_BNK 
0.763 
0.609 
0.910 
SOL_ZMX 
108.031 
66.609 
162.382 
SOL_Z 
365.522 
315.512 
506.027 
SOL_AWC 
0.074 
0.010 
0.089 
SOL_K 
0.440 
0.434 
0.884 
SOL_BD 
0.431 
0.333 
0.671 
HRU_SLP 
0.066 
0.044 
0.093 
OV_N 
0.196 
0.199 
0.168 
Table 3. Statistical Analysis of runoff simulation.
Variable 
Pfactor 
Rfactor 
NS 

Before Calibration 
FLOW_OUT 
0.35 
0.84 
0.53 
0.28 
After Calibration 
FLOW_OUT 
0.51 
0.54 
0.86 
0.77 
Validation 
FLOW_OUT 
0.63 
0.49 
0.85 
0.74 
CONCLUSION
Many hydrologic studies and applications has been used SWAT model and SWATCUP program for calibration and validation data with decreasing uncertainty (Schuol & Abasspour 2006; Stedinger et al. 2008; Alibuyog et al. 2009; Li et al. 2010; Gosling et al. 2011; Du et al. 2013; Lin et al. 2015; Nyeko 2015).
Hosseini et al. (2011) applied SUFI2 algorithm to simulate streamflow in Taleghan basin with an area of 800 . Fukunaga et al. (2015) investigated runoff simulation in the tropical watershed at Brazil using SUFI2 algorithm. Their results revealed SUFI2 algorithm performance was satisfactory in hydrology modeling. Vilaysane et al. (2015) applied SWAT model to test the capability of the model for predicting stream flow and also used SUFI2 algorithm for calibration and uncertainty analysis in Xedone river basin. Pagliero et al. (2011) used SWAT model to predict surface water flow and nutrient loads in the Danube basin with an area of 803000 . They applied SUFI2 algorithm to reduce parameters uncertainty.
In SUFI2 algorithm, all the uncertainties are combined and expressed through the Pfactor, which is the percentage of measured data bracketed by the 95% prediction uncertainty (95PPU) with ranges from 0 to 1. Also, in uncertainty analysis used the Rfactor, which is the average thickness of the 95PPU band divided by the standard deviation of the measured data (Yang et al., 2008; Abbaspour 2011; Narsimlu et al., 2015). In this study, It is calibrated fourteen parameters (e.g. CN2, ALPHA_BF, GW_DELAY, GWQMN, CH_N2, CH_K2, ALPHA_BNK, SOL_ZMX, SOL_Z, SOL_AWC, SOL_K, SOL_BD, HRU_SLP and OV_N) and tried to finding the best range of parameters with the most appropriate values of Pfactor and Rfactor (Table 3) that shown successfully efforts for decreasing uncertainty. In SWATCUP program, the Sequential Uncertainty Fitting (SUFI2) algorithm try to predict all uncertainties (input data, parameters, model structure, output data) by finding the best amount of parameters uncertainty (Abbaspour et al. 2004, 2007b). The measured data uncertainty should be considered and the repeat of performance calibration can be obtained the best goodness fit if the rating of Pfactor, Rfactor, and NS are relaxed (Abbaspour et al. 2007a). Table 3 illustrates the values of P and R factors, and NS in calibration (2001 to 2007) and validation (2008 to 2011) periods.
The Pfactor values close to 1 indicating a very high model performance, while the Rfactor is the average width of the 95PPU band (Abbaspour et al. 2007b; Yang et al. 2008). According to Table 3, after calibration and validation periods the Pfactor was obtained close to 1 with 0.51 and 0.63 respectively and thickness of the 95PPU band (Rfactor) was lower than prior. These values confirm the accuracy of runoff simulation processes to decreasing data uncertainty. In other hand, according to Moriasi et al. (2007) classification, who defined a ‘‘good model simulation’’ with NS values from 0.65 to 0.75 and a ‘‘best model simulation’’ with NS values greater than 0.75, the calibration and validation model show better performance of model with Nash and Sutcliffe efficiency (NS) value of 0.77 and 0.74 respectively. Also coefficient of determination ( ) value of 0.86 for calibration and 0.85 for validation period showed a good correlation between observed and simulated values (see Fig. 5).
These results to confirm the potential of SUFI2 algorithm of SWATCUP program for simulating runoff data in Shafaroud watershed and matched well with those of the other authors (Tang et al. 2012; Rostamian et al. 2013; Singh et al. 2013; Vilaysane et al. 2015; Narsimlu et al. 2015). It is suggested in future studies, to use SUFI2 algorithm in model parameters sensitivity and uncertainty analysis.
Also this algorithm can be used in further evaluation of land use change, sediment, climate change, water quality and evapotranspiration effect assessment on water resources.