Assessing the applicability of the soil and water assessment tool-deep learning hybrid model for predicting total nitrogen loads in a mixed agricultural watershed

Abstract

Soil and Water Assessment Tool (SWAT) is a widely used process-based watershed model for simulating hydrology and water quality under varying land use and climate conditions. Its performance relies heavily on effective calibration and validation to achieve accurate parameterization. However, these processes are often time-consuming and subject to considerable uncertainty. Multi-site calibration has been introduced to address the spatial limitations of conventional single-site calibration, yet it remains resource-intensive and may introduce inconsistencies among subbasins. To overcome these challenges, this study proposes SWAT-deep learning (DL) hybrid models for predicting total nitrogen (TN) loads in a mixed agricultural watershed. Specifically, the objective was to evaluate whether DL models trained at the watershed outlet using uncalibrated SWAT outputs could generalize effectively to upstream subbasins, thereby bypassing the need for calibration. Two hybrid models, SWAT-Long Short-Term Memory (LSTM) and SWAT-Gated Recurrent Unit (GRU), were constructed using uncalibrated SWAT simulations and precipitation data. Both hybrid models consistently outperformed the multi-site calibrated SWAT model. The SWAT-LSTM model demonstrated higher sensitivity in capturing sharp TN peaks during rainfall events, whereas the SWAT-GRU model provided more stable predictions across postpeak and recovery periods. Feature importance analysis further revealed distinct dependencies on hydrological and water quality variables. In addition, the SWAT-DL hybrid framework yielded a substantial practical advantage, achieving a more than a tenfold gain in computational efficiency over multi-site SWAT calibration while sustaining high accuracy. By reducing calibration demands without compromising accuracy and transferability, this hybrid approach represents a scalable and resource-efficient alternative for watershed-scale water quality modeling.