Methodologies for Improving a Watershed Water Quality Modeling Accuracy by Applying Deep Learning Models

Abstract

Watershed water quality modeling serves as a critical tool for preemptively identifying changes in water quality and formulating management strategies. Process-based Models (PBMs) are extensively employed to simulate the intricate physical, chemical, and biological processes occurring within a watershed. These models necessitate extensive datasets and rigorous calibration to reflect real-world conditions accurately. Nevertheless, the calibration and validation process of PBMs is labor-intensive and cannot entirely eliminate inherent uncertainties. This study proposes an enhancement to watershed water quality modeling by integrating PBMs with advanced Deep Learning (DL) models. Hybrid models are constructed by coupling the Soil and Water Assessment Tool (SWAT), a widely recognized PBM, with state-of-the-art DL models, including Long Short-Term Memory (LSTM) and Spectral Temporal Graph Neural Network (StemGNN). Initially, a SWAT-LSTM hybrid model is constructed at the watershed outlet using uncalibrated SWAT simulation outputs combined with meteorological data. This approach simplifies the modeling process while maintaining high predictive accuracy. Additionally, the simulation results for the entire watershed, integrated with meteorological data, are utilized to further refine water quality modeling via the StemGNN model. The hybrid modeling approach proposed herein not only enhances the convenience of watershed modeling but also markedly improves the accuracy rather than conventional PBMs. By incorporating comprehensive data from across the entire watershed, including tributaries, the proposed model achieves superior predictive accuracy and bolsters the explainability of water quality outcomes. The introduction of the SWAT-DL hybrid model in this study is anticipated to represent a significant advancement in the field of watershed water quality modeling.