Coupling SWAT+ with LSTM for enhanced and interpretable streamflow estimation in arid and semi-arid watersheds, a case study of the Tagus Headwaters River Basin, Spain

Abstract

Accurate streamflow prediction is crucial for effective water resources management and flood risk assessment. The Tagus Headwaters River Basin (THRB), a semi-arid watershed, serves over 10 million residents in Peninsular Spain and diverts water to the Segura River Basin. As the THRB nears its water allocation limits, precise streamflow simulations are essential for sustainable management. This study is especially important for arid and semi-arid watersheds, where previous research has shown that the performance of rainfall-runoff modeling using the LSTM AI-based technique declines in more arid catchments. This research enhances streamflow simulations in the THRB by combining the Soil and Water Assessment Tool (SWAT+) with a Long Short-Term Memory (LSTM) model. Five scenarios were evaluated, using different combinations of meteorological data and SWAT+ model outputs as LSTM input data. Results showed that coupled models generally provided more accurate daily streamflow estimates than standalone SWAT+ or LSTM models. Coupled LSTM and calibrated SWAT+ models significantly outperformed coupled LSTM and default SWAT+ models when using SWAT+ estimated streamflow as the sole input. Additionally, coupled models using different SWAT+ hydrological outputs and meteorological data as LSTM input data outperformed those using only SWAT+ estimated streamflow. This improvement was more notable in scenarios combining LSTM and default SWAT+ models, highlighting the SWAT+ default model’s effectiveness in capturing basin characteristics, reflected in hydrological metrics like lateral flow, percolation and soil water content. SHapley Additive exPlanations (SHAP) analysis revealed that SWAT+ outputs, especially lateral flow and percolation, were the most influential factors, with global importance ranging from 34% to 40% and 23% to 36% across all stations in the default scenario, respectively. These advancements enhance decision-making with more precise coupled model forecasts, particularly in arid and semi-arid watersheds like the THRB.