Use of interpretable machine learning for understanding ecosystem service trade-offs and their driving mechanisms in karst peak-cluster depression basin, China

• We revised the carbon fixation and soil erosion models in karst areas. • We used interpretable ML methods to identify key drivers of ecosystem service trade-offs. • The carbon fixation and soil erosion mainly exhibit a trade-off relationship. • NDVI is a key controlling factor for ecosystem service trade-offs. The peak-cluster depression is one of China’s most ecologically fragile areas, with extensive karst development. However, existing models for assessing karst ecosystem services often fail to consider the unique geological context of karst peak-cluster depressions, making it challenging to apply general international models to this area. To address these challenges, this study focused on the karst basin of Southwest China and evaluated ecosystem service in 2000 and 2020 by revising carbon fixation and soil erosion models. Using interpretable machine learning model (XGBoost-SHAP, eXtreme Gradient Boosting and SHapley Additive exPlanations), we quantified the nonlinear characteristics and threshold effects of ecosystem service trade-offs and synergies. Our findings include the following: (1) Carbon fixation increased from 753.99 tCO 2 ∙km −2 ∙a -1 in 2000 to 756.70 tCO 2 ∙km −2 ∙a -1 in 2020; however, soil erosion decreased from 16.56 t∙hm −2 ∙a -1 to 15.12 t∙hm −2 ∙a -1 . (2) At the basin scale, carbon fixation and soil erosion exhibited both trade-offs and synergistic relationships, with 63.3 % of the area showing a trade-off and 36.7 % showing a synergistic relationship. Trade-off relationships were prevalent in the upper and lower reaches, while the middle reaches demonstrated synergistic relationships. (3) Normalized Difference Vegetation Index (NDVI) emerged as the primary driver of changes in ecosystem service trade-offs, with NDVI, precipitation, temperature, evapotranspiration, elevation, and lithology as the most significant explanatory factors. These factors impact ecosystem service trade-offs in a nonlinear manner and exhibit pronounced threshold effects. (4) Climate factors contributed 31.65 % to ecosystem service trade-offs, geomorphic factors contributed 14.81 %, soil factors contributed 5.72 %, and human activities contributed 5.39 %. (5) Local interpretability SHAP values indicated substantial differences in the contributions of drivers at different scales to ecosystem service trade-offs. The methodology implemented in this study offers a practical approach for the sustainable and differentiated management of karst ecosystem services by integrating karst ecosystem service assessment models with interpretable machine learning methods.