Combined use of crop yield statistics and remotely sensed products for enhanced simulations of evapotranspiration within an agricultural watershed

Abstract

Water cycling within agricultural watersheds includes high uncertainty because of natural and anthropogenic factors (e.g., cultivation practices). Remotely sensed evapotranspiration products (RS-ET) have been adopted as an additional constraint on watershed modeling to enhance the accuracy of water cycling predictions while reducing uncertainty. However, plant parameters affecting evapotranspiration (ET) in watershed models are poorly calibrated without the use of appropriate constraints. The goal of this study is to assess the predictive uncertainty of the Soil and Water Assessment Tool (SWAT), depending on the inclusion or exclusion of annual crop yield as an additional constraint for an agricultural watershed. We analyzed the simulated results with acceptable performance measures depending on a varying degree of model constraints: one constraint (streamflow), two constraints (streamflow and RS-ET) and three constraints (streamflow, RS-ET, and crop yield). The three performance measures used were Nash-Sutcliffe Efficiency (NSE), Percent bias (P-bias), and Kling-Gupta Efficiency (KGE). As the number of model’s constraints increased, the number of acceptable parameter sets were substantially reduced from 180 (acceptable for streamflow) to 116 (acceptable for streamflow and RS-ET) and 2 (acceptable for streamflow, RS-ET, and crop yield). In addition, overall model performance measures for ET were greatest in the simulation results with three constraints representing 0.02–0.2 and 0.04–0.05 greater NSE and KGE values than those of one constraint and two constraints, respectively. The parameter set with the best ET performance measures was also acceptable for predicting crop yield. Based on these results, we conclude that this crop yield data can be adopted as a model constraint for agricultural watersheds to reduce model uncertainty in ET simulations and to increase model prediction accuracy.