Explainable machine learning reveals drivers of ecological quality in a karst watershed: implications for sustainable management of the Mawei river basin, China

A significant research gap in quantitatively understanding the thresholds and interactions of ecological quality drivers hampers sustainable development in fragile karst watersheds. This study aims to develop an interpretable machine learning framework to identify complex drivers and their synergistic effects in a typical karst basin in China. We integrated the XGBoost-SHAP with trend analysis methods to analyze the drivers of the remote sensing ecological index (RSEI) in the Mawei River Basin (MRB) from 1990 to 2020. The key results indicate that land use/land cover (LULC) was identified as the dominant factor affecting spatial heterogeneity, with critical thresholds observed on slopes <20° and elevations <1200 m, which amplified anthropogenic degradation. Southeastern aspects (50°–250°) presented the lowest RSEI values, primarily because of solar-induced aridity and agricultural expansion. Furthermore, interactions between carbonate lithology and low elevation exacerbated soil erosion (>100 t·ha–1 ·yr–1 ). These findings directly inform sustainable management strategies; for example, extending the slope threshold for China’s Grain for Green Program from 25° to 20° in food-secure areas could improve ecological quality across 43.82 km². This study provides a transferable framework for ecological management in vulnerable karst regions, demonstrating how artificial intelligence can bridge scientific insights and policy innovation.