What dominates the variation of Mining footprint ecological resilience based on explainable machine learning-Evidence from the typical resource-based city, China

Accurately assessing the ecological resilience of mining footprints (MER) and elucidating its driving mechanisms are crucial for enabling the sustainable transition of resource-based cities (RBC). However, conventional methods struggle to capture the complex nonlinear relationships inherent in MER dynamics. To address the limitations in parsing complex nonlinear relationships, the paper developed a multidimensional assessment framework for MER, termed the ARTS framework. Applied to typical RBC in China, this framework was combined with Gradient Boosting Decision Tree (GBDT) modeling and SHAP analysis to identify key drivers of MER and their nonlinear interactions. The results reveal three key findings: (1) The overall MER exhibited a fluctuating yet upward trend from 2000 to 2023, with its temporal evolution transitioning through stages of slow ascent, accelerated enhancement, and high-level fluctuation. Spatially, MER exhibited a clear stepwise decline from the southeastern coast toward the northwestern interior. Southern China had the highest mean MER value of 10.10, followed by Southwest China with 8.56 and Northeast China with 8.10. (2) The GBDT model significantly outperformed traditional linear models. AOD and LUCC were the most critical drivers of MER, with feature importance scores of 0.4323 and 0.2872, respectively. These factors exhibited pronounced threshold effects. Conditions most conducive to higher MER were associated with an NDVI above 0.6, a mean temperature near 20 °C, a GDP density exceeding 7000 units/km2, a population density between 1666 and 2797 persons/km2, and a forest-dominated land use pattern. (3) Furthermore, significant interaction effects were observed among drivers. Notably, the synergy between AOD and ecological baseline factors such as NPP, PRE, and NDVI generates a compound ecological pressure characteristic of resource-intensive landscapes. This study proposes a novel assessment framework that enhances the mechanistic understanding of MER in RBC. By integrating explainable machine learning with ecological analysis, this research offers both theoretical insights into the nonlinear dynamics of coupled human-natural systems and practical guidance for implementing threshold-based, differentiated restoration strategies. The framework and findings provide a transferable reference for MER assessment in other resource-dependent regions worldwide.