Modeling runoff, sediment, and nutrient dynamics in the Ganjiang River Basin, China: A hierarchical SWAT-BiLSTM framework

Rivers and lakes are vital to regional water security and ecosystem stability, yet the integrated simulation of hydrological, erosion and pollution processes remains constrained by the complex interplay between climate change and human activities. To address systematic biases arising from oversimplified mechanisms and strong non-linearities, this study proposes a hierarchical hybrid framework that sequentially feeds SWAT-simulated hydrological and water-quality processes into a BiLSTM network for residual correction, achieving coordinated, high-precision simulations of runoff, sediment and nutrient fluxes in the Ganjiang River Basin. Validated against multi-source data that hydrological data during 2011–2022 and water quality data during 2021–2022, the SWAT-BiLSTM outperformed standalone SWAT across all metrics, primarily driven by precision gains and residual mitigation rather than computational speedup. Specifically, the average R 2 increased by 6%–9% for runoff, 6%–19% for sediment and significantly by 9%–42% for nutrients, with a maximum P BIAS reduction of 67.4%. The hybrid model further revealed the cascade response mechanisms of the hydrological-erosion-pollution triad, highlighting concentrated rainfall and agricultural cycles as key drivers of seasonal nutrient peaks. These findings confirm the efficacy of deep learning-enhanced mechanistic modeling, offering robust decision support for basin-scale environmental management and pollution control.