An integrated watershed modeling approach using soil and water assessment tool and graph convolutional long short-term memory

Abstract

Watersheds are complex systems where upstream-downstream interactions play a critical role, necessitating an incorporated approach for effective watershed management. Previous studies have primarily relied on either process-based models or Machine learning (ML) and Deep learning (DL) models for watershed modeling. However, despite the strengths of each model, the inherent drawbacks underscore the need for a complementary integration of both models. This study proposes an approach that integrates the process-based Soil and Water Assessment Tool (SWAT) with the graph-based Graph Convolutional Long Short-Term Memory (GCLSTM) model to simulate streamflow and Total phosphorus (TP) load across multiple regions within a watershed. SWAT simulation results with default parameter values, meteorological data, and watershed information were utilized as input data for the GCLSTM model to perform incorporated watershed modeling. The study focused on the Yeongsan River watershed in Korea, with all simulations carried out over the 2017-2021 period. Compared to the calibrated SWAT, coupling uncalibrated SWAT with GCLSTM increased streamflow R2 from 0.22-0.74 to 0.40-0.88 and TP load R2 from 0.02-0.36 to 0.50-0.81. This performance gain reflected the ability of the GCLSTM to aggregate upstream hydrometeorological and land use signals across the river network, capturing nonlinear spatiotemporal dependencies. Network analysis revealed that upstream precipitation is the dominant driver of downstream streamflow, while upstream land use patterns govern TP load variability. By incorporating key factors through network analysis into the modeling framework, this approach underscores GCLSTM's potential as a decision-support tool for devising streamflow regulation and nutrient reduction measures that faithfully reflect actual watershed conditions.