Apprendimento Automatico Interpretabile per la Valutazione dei Fattori di Subsidenza del Suolo Utilizzando Dati Geospaziali e Idrogeologici in Quebec, Canada

Land subsidence (LS) is a critical environmental hazard with significant socio-economic and infrastructural impacts, particularly in regions reliant on groundwater resources. Southern Quebec, Canada, faces growing risks of LS due to increasing groundwater extraction, climate change, and population growth. However, the relationship between natural and anthropogenic drivers of LS in humid, groundwater-dependent regions like Quebec remains poorly understood. This study addresses this gap by developing an interpretable machine learning (ML) framework to assess LS vulnerability in Southern Quebec. The research introduces the ALPRIFTA framework, an enhanced version of the ALPRIFT model, incorporating eight geospatial and hydrogeological parameters: Aquifer media (A), Land use (L), Pumping of groundwater (P), Recharge (R), Impact of aquifer thickness (I), Fault distance (F), Water table decline (T), and Aquifer type (A). High-resolution datasets, including over 2,200 abstraction wells and 230 lithological logs, were integrated with Sentinel-1 InSAR data to map subsidence trends from 2014 to 2023. A Random Forest (RF) model, coupled with SHapley Additive exPlanations (SHAP), was employed to predict LS vulnerability and interpret the contribution of each driving factor. Key findings reveal that Land use/Land cover (LULC) is the most influential predictor (30.3%), highlighting the role of urbanization and agriculture in exacerbating LS. Aquifer thickness (14.9%), aquifer media (14.5%), and fault proximity (12.9%) also significantly contribute, while groundwater pumping (9.4%) and recharge rates (10.9%) further modulate subsidence risk. The regression-based RF model outperformed classification approaches, achieving a Pearson correlation of 0.46 with InSAR-derived subsidence data, demonstrating its robustness in capturing spatial variability. This study provides the first regional-scale, interpretable assessment of LS drivers in Quebec, offering actionable insights for sustainable groundwater management and urban planning. The ALPRIFTA framework and SHAP-based interpretability enhance transparency in ML applications for geohazard mitigation, serving as a model for other data-poor yet vulnerable regions globally.