Enhancing mid- to long-term runoff simulation in human-impacted basins through coupled SWAT-LUT and LSTM modelin

Study Region: This study focuses on the Dawen River Basin in China, a representative human-impacted watershed experiencing rapid urbanization, land-use change, and strong reservoir regulation. The basin is characterized by pronounced seasonal hydrological variability, making it a suitable case for evaluating the combined impacts of land-use dynamics and human interventions on runoff processes. Study Focus: An integrated SWAT-LUT-LSTM model was developed by coupling a dynamic land-use-based SWAT (SWAT-LUT) with a long short-term memory (LSTM) network to improve mid- to long-term runoff simulations. Its performance was compared against conventional SWAT and SWAT-LUT models. Two indices—the land-use inconsistency coefficient (IC) and the pervious change coefficient (PCC)—were introduced to quantify spatiotemporal impacts of land-use change. Additionally, the monthly-seasonal pattern (SP) of SWAT-LUT errors was analyzed to evaluate reservoir regulation effects. New Hydrological Insights for the Region: Runoff dynamics in the Dawen River Basin have shifted markedly over the past four decades due to intensive human activities. Urbanization reduced surface permeability (PCC 0.08–0.30) and increased the share of surface runoff from 46 % to 64 %, altering the basin’s water balance. A critical IC threshold of 0.25 was identified; with the basin now at IC ≈ 0.30, dynamic land-use inputs are essential for accurate simulation. Reservoir operations further reshaped intra-annual runoff, producing a 12-month periodic error pattern strongly linked to water-level fluctuations.