Improving estimation capacity of a hybrid model of LSTM and SWAT by reducing parameter uncertainty

Hybrid models coupling process-based models with deep learning models have been widely used for estimating water quantity and quality. However, the impacts of uncertainty within process-based models on the performance of hybrid models remained largely unknown. This study focused on assessing the impact of uncertainties on a hybrid model that combines Soil and Water Assessment Tool (SWAT) with Long Short-Term Memory (LSTM) for estimating streamflow and suspended solids (SS). This study applied the output of SWAT as input for LSTM to make a hybrid model. By incorporating an additional constraint, remotely sensed evapotranspiration (RS-ET), the performance of hybrid models with and without considering RS-ET was evaluated at various temporal and spatial scales. Furthermore, various input settings (default and calibrated SWAT and inclusion/exclusion of precipitation) were considered. The results showed that the hybrid models tended to provide accurate estimations of monthly streamflow and SS compared to standalone SWAT and LSTM. The addition of precipitation to hybrid models did not make noticeable improvements for streamflow and SS. The hybrid models constrained by RS-ET tended to have improved estimations on streamflow and SS at daily and monthly scales compared to those unconstrained by RS-ET. Similar results were also observed at the sub-watershed level. These insights highlighted the potential of hybrid models to enhance hydrological estimations and underscore the importance of incorporating additional constraints to reduce uncertainties. Such advancements have practical implications for water resource management and decision-making processes, enabling more reliable and accurate estimations in applications of hybrid models.