Explainable artificial intelligence reveals heterogeneous erosion responses to extreme rainfall: A new framework for conservation prioritization

• A novel framework integrates Explainable AI to quantify erosion drivers. • Short-duration intense rainfall is the primary trigger of regional soil erosion. • Erosion responses to extreme rainfall vary across different conservation zones. • LightGBM–SHAP effectively decodes spatial heterogeneity in erosion risk. • A replicable AI-based framework supports SWC resource optimization. Effectively combating soil erosion under intensifying extreme precipitation requires moving beyond broad-scale assessments to precision-targeted interventions. Current approaches often fail to capture the differentiated erosion responses across diverse landscapes, limiting the efficacy of conservation investments. Here, we introduce a novel framework that combines modeling (RUSLE) with Explainable artificial intelligence (LightGBM model with SHAP method) and multi-criteria decision analysis to prioritize soil and water conservation (SWC) efforts. Applied to 20 SWC zones in the ecologically critical Yunnan Province, China, our analysis reveals that maximum 5-day precipitation (RX5) is the paramount driver of erosion, superseding other precipitation metrics. We uncover distinct regional response mechanisms: erosion in northwestern cold alpine zones is governed by cumulative rainfall, while erosion in temperate and tropical zones is triggered by short-duration, high-intensity events. This mechanistic understanding enabled the robust identification of the Northwest Alpine Gorge and Western Broad Valley zones as highest-priority areas demanding urgent action. By systematically diagnosing the primary climatic drivers and identifying the most vulnerable regions, our framework provides a powerful, replicable blueprint for optimizing conservation resources and enhancing climate resilience in mountainous ecosystems globally.