Effects of input data accuracy, catchment threshold areas and calibration algorithms on model uncertainty reduction

Abstract

AbstractLow resolution of input data and equifinality in model calibration can lead to inaccuracy and insufficient reflection of spatial differences, thereby increasing model errors. However, the impact of input data accuracy, catchment threshold area, and calibration algorithm on model uncertainty reduction has not yet been well understood. The sequential uncertainty fitting version 2 (SUFI‐2) that is linked with the Soil and Water Assessment Tool (SWAT) in the package called SWAT Calibration Uncertainty Programs (SWAT‐CUP) was introduced to quantify the effects of different input data resolutions on parameter sensitivity and model uncertainty in the Jinghe River watershed, and the effects of different sub‐basin delineations and other two calibration algorithms on model uncertainty were also comparatively analysed. (i) USLE_C, EPCO, ALPHA_BNK, and CN2 are the most sensitive parameters among all SWAT projects. When the change of digital elevation model (DEM) resolution is small, the sensitivity of parameters does not change obviously. When the DEM resolution changes significantly, BIOMIX, LAT_SED, USLE_K, and CH_N1 become highly sensitive parameters by replacing OV_N, SMTMP, SURLAG, and USLE_P. However, the change in land use resolution has little impact on parameter sensitivity, with only a slight change in the sensitivity ranking of specific parameters. (ii) Model uncertainty responded to changes in the resolution of DEM more than land use. Most of the runoff simulations had smaller uncertainties (P factor, R factor, percentage of bias [PBIAS]) than sediment. High resolution DEM data reduced model uncertainty, but the models with 2000 m DEM resolution also achieved small uncertainty. Small catchment threshold area leads to high uncertainty of the model, and large catchment threshold areas decrease the model uncertainty. The model has relatively good simulation effects in runoff and sediment when the catchment threshold area was 2000 km2. (iii) The SWAT model has different simulation deviations and uncertainties in different calibration algorithms, the SUFI‐2 and generalized likelihood uncertainty estimation (GLUE) algorithms show better applicability than particle swarm optimization (PSO). The NSE indicators of the three algorithms are in the following order: SUFI‐2 > GLUE > PSO for runoff, and GLUE > SUFI‐2 > PSO for sediment. This study helps us understand the cause, knowledge of which moves from the particular to the general by the comprehension of essence, power, and nature in reducing model uncertainty.