Comparative Analysis of Sinkhole Susceptibility Models Using Ensemble Machine Learning and Local Interpretable Model‐Agnostic Explanations ( <scp>LIME</scp> )

ABSTRACT The accelerating rate of cover‐collapse sinkhole formation in the Konya Closed Basin (KCB) poses a critical threat to infrastructure and agriculture. However, standard machine learning susceptibility models often yield unreliable results by failing to account for the strong spatial clustering inherent in such geohazards. This study addresses this methodological gap by developing spatially robust and explainable Sinkhole susceptibility maps (SSMs). A comprehensive inventory of 495 sinkholes (2008–2025) was analyzed using a temporally consistent dataset. Unlike traditional studies relying on random partitioning, this study implemented a rigorous Spatial Block Cross‐Validation (SBCV) strategy to mitigate spatial leakage. Five ensemble algorithms were evaluated; results indicated that XGBoost and CatBoost demonstrated superior spatial discrimination with mean AUC scores of 0.88 and 0.85, respectively, confirming their robustness against spatial heterogeneity. To transcend “black‐box” opacity, the LIME technique was employed, revealing novel causal mechanisms. Crucially, the analysis empirically validated the “hypogenic karstification” theory, identifying a 20 km influence zone around volcanic forms. Furthermore, the strong predictive power of well density and groundwater depletion exposed the direct link between unsustainable irrigation and sinkhole formation. This study provides a statistically robust baseline for land‐use planning and advocates for a paradigm shift from reactive response to proactive aquifer management.