Assessment of Input Uncertainty in SWAT Using Latent Variables

Applications of the Soil and Water Assessment Tool (SWAT) require a large amount of input data to perform model simulations. Consequently, uncertainty in input data tends to influence the accuracy of SWAT hydrologic and water quality outputs. It has been shown that input uncertainty can be quantified explicitly during model calibration with latent variables. In this study, latent variables were explored to examine their sensitivity to SWAT outputs and further the potential impact of input uncertainty to model predictions. Results show that the increases in the range of latent variables pose a significant influence to streamflow and ammonia predictions while the impact was less significant in sediment responses. The performance of SWAT in predicting streamflow and ammonia declined with wider ranges of latent variables. In addition, the increase in the range of latent variables did not present noticeable effect on the corresponding predictive uncertainty in sediment predictions. In this study, the calibration results did not improve significantly with the applications of wider ranges of latent variables which are different from the findings in previous research work. The use of latent variables to incorporate input uncertainty may not be the proper alternative choice in terms of generating better results and should be carefully evaluated in the implementations of complex watershed simulation models.