Stage-transition mechanism of forest-grass vegetation coverage in machine-learning simulation of event-scale...

Vegetation restoration is widely regarded as a key measure for mitigating soil erosion in the middle reaches of the Yellow River. However, the regulation of flood sediment transport by forest-grass vegetation coverage (Ve) shows strong nonlinear characteristics and threshold effects. This makes it difficult for traditional physical models to accurately describe sediment production responses across different vegetation stages. To address this issue, this study proposes a stage-specific water-sediment simulation framework that integrates Ve threshold identification with machine learning, namely a threshold-aware modeling framework. Four typical basins in the middle reaches of the Yellow River, namely the Kuye, Wuding, Fen, and Wei River basins, were selected as the study areas. Based on multi-source data from 1979 to 2025, including flood events, rainfall, Ve, and land use, random forest was first used to identify the dominant factors controlling sediment production. The exponential function, piecewise linear regression, and Copula model were then combined to identify Ve threshold bands from three perspectives: functional form, structural breakpoint, and probability dependence. Based on the identified thresholds, threshold information was embedded into LSTM, XGBoost, and SVR models to construct overall and stage-specific simulation scenarios.The results show that: (1) Ve and rainfall are the key factors controlling sediment production; (2) clear Ve threshold bands exist in all basins, with recommended ranges of 31%∼35% for the Kuye River Basin, 27%∼32% for the Wuding River Basin, 23%∼28% for the Fen River Basin, and 20%∼23% for the Wei River Basin; (3) model performance improved significantly after introducing threshold information. Taking the Kuye River Basin as an example, the NSE values of the LSTM, XGBoost, and SVR models increased to 0.858, 0.861, and 0.830, respectively; and (4) vegetation exerted a strong regulatory effect on suspended sediment concentration in the pre-threshold stage, whereas the system gradually shifted to a sediment-supply-limited state in the post-threshold stage, with a markedly weakened marginal vegetation effect. This study reveals the stage-specific regulatory mechanism of vegetation restoration on water-sediment processes. It also provides new theoretical and methodological support for intelligent modeling of complex water-sediment systems and ecological management of river basins.