Analyzing dominant driving factors of ground deformation in mining area via multimodal feature fusion and explainable modeling

Ground deformation in areas of natural resource extraction results from the interplay of multiple anthropogenic and hydrogeological factors. Identifying the main factors and their relationship to deformation is crucial for mitigating disasters when extreme events happen. Here, we show how multimodal artificial intelligence models can disentangle the effects of dynamic and static features related to deformation in the Rhineland coalfield in Germany. We systematically investigate the synergistic effects of multi-source data by integrating static environmental features, including mining areas, urban areas, fault lines, farmland, and geological distributions, with dynamic features, such as surface deformation time-series and annual groundwater-level observations, and organising them into 12 nested multimodal subsets. Using mainstream deep learning modules, including Convolutional Neural Networks, Graph Convolutional Networks, and Transformer, we build six distinct network architectures and evaluate them under comparable settings through progressive feature fusion and feature ablation. The results show that multimodal fusion consistently improves predictive accuracy compared to a model driven only by deformation data, indicating that static spatial distribution information enhances surface deformation prediction when incorporated as contextual background. Within the controlled feature-space design, feature ablation provides stable channel-level comparisons, whereas SHAP remains a useful complementary post hoc attribution method. At the regional scale, geology and groundwater emerge as the most important predictors of the overall deformation pattern, which is consistent with the interpretation that geological conditions shape the hydrogeological setting and the spatial expression of groundwater fluctuations. At the local scale, fault information becomes the most important predictor within spatially confined high-impact zones, because structural discontinuities concentrate deformation gradients and groundwater-related responses. This distinction between regional background control and local structural triggering provides a data-driven basis for assessing deformation susceptibility and mitigating hazards in mining areas.