Hydrological and tectonic linkage to subsidence risk in the Rhineland coalfields, Germany: Integrating machine learning with remote sensing and geospatial data

Land subsidence is increasingly emerging as a widespread geohazard across many regions of the world, driven by unsustainable groundwater extraction, rapid urbanization, and resource overexploitation. However, in the Rhineland coalfields of Germany, large-scale and spatially extensive subsidence has been predominantly linked to open-pit coal mining operations. This study investigates the environmental and geological drivers of ground deformation in the coalfields, with a particular focus on identifying the most influential factors contributing to mining-induced subsidence. A range of geospatial variables including topography (Digital Elevation Model (DEM) derived slope and aspect), hydrological data (groundwater levels and river locations), geological data (lithological data, fault locations), Remote sensing data (InSAR measurements - ground motion data from European Ground Motion Service (EGMS)), Normalized Difference Vegetation Index (NDVI)), and anthropogenic data (mine locations) were integrated to train machine learning models aimed at understanding the relative influence of each factor on subsidence susceptibility. The EGMS serves as the reference data for training and testing the model. Among the three classifiers, Random Forest achieved the highest overall accuracy, with Inersection Over Union (IoU) scores of 0.95, 0.92, 0.91, and 0.99 across the Very High, High, Moderate and Low susceptibility classes, outperforming XGBoost and LightGBM. Its superior predictive stability justified its selection for Permutation Feature Importance (PFI) analysis to identify the dominant drivers of land subsidence. PFI analysis shows that distance from mines is the dominant predictor, reducing model accuracy by 25% when permuted, followed by groundwater level change with 21%. Geological factors show moderate influence (distance from faults 14%, lithology 8%), while hydrological factor (dist from river) accounts for 6%. In contrast, terrain and landcover variables contribute minimally (<5%). Areas classified under high risk were further examined using angular distortion analysis based on EGMS data, highlighting localized differential settlements that concentrate along mines and specific fault zones, amplifying the structural damage, as confirmed by field observations.