Spatial heterogeneity of land subsidence in the california central valley: a model-based attribution perspective

The California Central Valley (CCV), a globally important agricultural region, has long experienced severe land subsidence (LS) associated with groundwater overextraction, thereby threatening infrastructure safety and aquifer storage capacity. Although the spatial pattern of subsidence in this region is strongly heterogeneous, existing approaches still have limitations in fully quantifying the spatial variability of internal statistical associations and in systematically summarizing differentiated zones. To address this issue, this study develops the Data-Ensemble-Explanation-Cluster-Threshold (DEECT) framework, which integrates stacking-based prediction, SHAP-based explanation, clustering, and GAM-based analytical transition analysis within a single exploratory workflow. Under a conventional random split benchmark, the stacking ensemble achieved strong fit (R2 = 0.9756, RMSE = 5.3997); however, this result is interpreted only as an upper-bound reference for within-study-area exploratory analysis rather than as evidence of spatial generalizability. Under a conventional random-split setting, the stacking ensemble achieved strong fit (R² = 0.9756, RMSE = 5.3997). To assess model robustness under spatial separation, we also applied a 10 km spatial block cross-validation, under which the stacking model achieved R² = 0.9122 and RMSE = 10.1306. Compressible layer thickness, precipitation, and non-ponded agricultural pumping emerged as common high-ranking model-based attribution factors. The study area was further partitioned into four typical subsidence zones with distinct statistical association patterns, and cluster-specific nonlinear analytical transition points were identified for selected variables. Overall, the DEECT framework provides an exploratory basis for organizing within-region heterogeneity in subsidence associations and for informing subsequent zonal hydrogeological validation.