Multidimensional spatiotemporal differentiation and driving mechanisms of soil erosion in a mountainous small watershed: an ST-TransUNet-R framework

Soil erosion threatens ecological security and water conservation in mountainous small watersheds. To improve pixel-level assessment under complex terrain and incomplete multi-source raster inputs, we developed ST-TransUNet-R, a deep-learning framework that integrates a ResNet-50 encoder, a Transformer module, a joint-validity mask, and masked mean squared error loss. The framework was applied to the Xikouwai Small Watershed upstream of the Miyun Reservoir for five assessment years between 2005 and 2020 and achieved R² = 0.87, RMSE = 2.08, MAE = 1.15, and PBIAS = − 3.13%. Multidimensional analysis showed that the 600–900 m elevation band contained 48.0% of the eroded area, whereas zones above 1100 m produced disproportionately high soil loss per unit area. Slopes > 35°, vegetation cover < 30%, unused land, and skeletal soils were consistently erosion-sensitive. SHAP analysis identified slope as the dominant predictor, followed by slope length and vegetation cover, indicating that terrain-controlled runoff concentration and surface protection jointly shaped spatial erosion heterogeneity. These findings provide a high-resolution basis for prioritizing erosion-control measures in mountainous water-source protection areas.