COMPARATIVE ANALYSIS OF SWAT CUP AND SWATSHARE FOR CALIBRATING SWAT MODELS
In order to distinguish essays and pre-prints from academic theses, we have a separate category. These are often much longer text based documents than a paper.
Soil and water assessment tool model (SWAT model) is a widely used model when dealing with large and complex watershed simulations. To correctly predict runoff of a watershed, auto-calibration methods are applied. Among all the platforms, SWAT CUP is widely used in the SWAT model community. The new web-based calibration platform: SWATShare is also gaining its popularity due to the benefits of user-friendly interface, access to high-performance computing resources, and collaborative interface. While the algorithm implemented in SWAT CUP is Sequential Uncertainty Fitting version 2 (SUFI2), Sorting Genetic Algorithm II (NSGA-II) is the algorithm employed by SWATShare. There is a limited amount of research comparing the model performance between these two calibration algorithms and platforms.
This study aims to examine whether the performances of calibrated models are providing equally reliable results. Thirty US watersheds are studied in this research, SWAT models were calibrated using seven years of rainfall data and outflow observations from 2001 to 2007, and then the models were validated using three years of historical records from 2008 to 2010. Inconsistency exists between different algorithms calibrated parameter sets, and the percentage difference between parameter values ranges from 8.7% to 331.5%. However, in two-thirds of the study basins, there is no significant difference between objective function values in two algorithms calibrated models. Correlations are examined using values of parameters and watershed features. Among all the features and parameters, Length of reach and GW_DELAY, CH_N2 and ALPHA_BF, climate zone and GWQMN, SFTMP and NSE have medium correlation exist in both SWATShare and SWAT CUP calibrated models among 30 watersheds. The correlation coefficient difference between them are less than 0.1. When visualizing results by Ecoregions, KGE and NSE are similar in calibrated models from both tools.
The initial parameter range used for SWAT CUP calibration could lead to satisfactory results with greater than 0.5 objective function values. However, the parameter values of the calibrated model might not be presenting a real physical condition since they are out of the realistic range. The inaccurate parameter values might lead to lower objective function values in the validation. The objective function values can be improved by setting the range of parameter values to match the realistic values.
By comparing two tools, SWATShare accurately calibrates parameter values to a realistic range using default range in most cases. For those models with an unsatisfactory result from SWATShare, the objective function values could be improved after specifying the parameters to the best-fit range given by SWAT CUP results. Also, for those watersheds which have similar satisfactory calibrated objective values from both tools, constraining the parameter to a reasonable range could generate a new calibrated model that performs as well as the original one. Using the approach to constrain parameter values to a realistic range gradually can exclude some statistically satisfactory but physically meaningless models. Comparing two auto-calibration software, SWATShare accurately calibrates parameter values to a realistic range using default range in most cases. Also, in some of the ecoregions, the best parameter sets in SWATShare fall in a more physically meaningful range. Overall, the newly emerged platform, SWATShare, is found to have the capability of conducting good SWAT model calibration.