Coupling the remote sensing data-enhanced SWAT model with the bidirectional long short-term memory model to improve daily streamflow simulations

Abstract

Global climate change has led to an increase in the frequency and scale of extreme weather events worldwide, and there is an urgent need to develop better-performing hydrological models to improve the accuracy of streamflow simulations and to facilitate water resource planning and management. The Soil and Water Assessment Tool (SWAT) has a notable physical foundation and is widely used in hydrological research. However, it uses a simplified vegetation growth model, introducing uncertainty into the simulation results. This study focused on improving the model based on remotely sensed phenological and leaf area index (LAI) data. Phenological data were used to define vegetation dormancy, and the LAI data replaced the corresponding data simulated by the original model. This approach improved the accuracy of the model in describing vegetation dynamics. Then, the enhanced SWAT model was coupled with the bidirectional long short-term memory (BiLSTM) model to validate the simulation of hydrological processes upstream of the Hei River. During model validation, the performance of the enhanced SWAT model in simulating streamflow (R2 = 0.835, NSE = 0.819) was better than that of the original SWAT model (R2 = 0.821, NSE = 0.805). In terms of simulating evapotranspiration, the enhanced SWAT model demonstrated even greater advantages. During the verification period, compared to those of the SWAT model, the R2 and NSE values of the enhanced SWAT model for daily-scale simulations increased from 0.196 and −0.269 to 0.777 and 0.732, respectively. The R2 and NSE values for monthly-scale simulations increased from 0.782 and 0.678 to 0.906 and 0.851, respectively. Simultaneously, the performance levels of two coupling approaches in streamflow prediction were compared, i.e., direct coupling of the original SWAT and BiLSTM models (SWAT-BiLSTM) and coupling of the enhanced SWAT and BiLSTM models (enhanced SWAT-BiLSTM). The results showed that the enhanced SWAT-BiLSTM model always performed better than the SWAT-BiLSTM model during the entire simulation period, especially the enhanced SWAT-BiLSTM model, which could more accurately predict peak streamflow changes. This study demonstrated that coupling an improved physical model with deep learning models could improve the streamflow prediction accuracy.