Sensitivity analysis of SWAT model in the Yarra River catchment

Catchment-scale hydrologic and diffuse source pollution models simulating a catchment are useful analysis tools to understand problems and find solutions through simulation of BMPs for particular catchment and agronomic settings. However, developing reliable catchment model and validating them on real-world catchment with monitored data is challenging. In this regard, model calibration and uncertainty analysis help to evaluate the ability of the model to sufficiently predict streamflow and constituent yields for specific applications. Complex physics-based distributed models contain many parameters that can complicate calibration process. In addition, the model when includes multi-variable at multi-site with multi-objective functions introduces more complexity to the calibration process. Over-parameterization is a well-known problem in such distributed model. Sensitivity analysis methods reducing the number of parameters to be adjusted during calibration are important for simplifying the use of these models. The objective of this paper is to perform a sensitivity analysis for multiple variables (streamflow, sediment and nutrients) at three sites on a SWAT model developed in the agricultural part of the Yarra River catchment, Victoria (Australia) so that the model can be calibrated efficiently for water quality analysis purposes. SWAT is a continuous physics-based distributed model that operates on a daily time-step. The SWAT model requires the following data types: digital elevation model (DEM), land use, soil, land use management, daily climate, streamflow and water quality data. Australian catchments are data-rich in terms of hydroclimatic data, but data-poor especially for water quality and land use management. For this study, all the data were collected from local organizations except DEM. Water quality and land use management data were most sparse. All input files for the model were organized and assembled following the guidelines of ArcSWAT interface of the SWAT 2005 version. The study area was delineated into 51 sub-catchments and 431 hydrological response units (HRU), which are unique combinations of land use, soil type and slope. The main methods used in modeling the hydrologic processes in SWAT were curve number method for runoff estimating, Penman-Monteith method for PET and Muskingum method for channel routing. SWAT has an embedded automatic sensitivity, and calibration and uncertainty analysis tool. The sensitivity analysis method is a combination of Latin-Hypercube and One-factor-At-a-Time (LH-OAT) sampling that allows a global sensitivity analysis for a long list of parameters with only a limited number of model runs. SWAT has 26 streamflow, 6 sediment and 9 nutrient parameters. The LH-OAT sensitivity analysis was applied for streamflow (Q), Total Nitrogen (TN), Total Phosphorus (TP) and Total Suspended Solid (TSS) output variables at three sites in the study area for 1998-2008 periods. The LH-OAT sensitivity analysis provides a simple and quick way to assess parameter sensitivity for multiple variables across the study area. The output variables found to be most sensitive to 15 hydrologic parameters, and 13 sediment and nutrients parameters in the SWAT model. The results show that the hydrologic parameters dominate the highest parameter ranks. The results also show that water quality variables are potentially capable of contributing to the identification of water quantity parameters within the SWAT model, and a single parameter is correlated to multiple variables. Moreover, there were clear differences in ranking of a parameter among the three sites. This result has evidenced how the parameter importance depends on land use, topography and soil types, meaning that a generalization within a catchment is limited. Hence, justify the importance of multi-site parameterization.