Automated monitoring and geometric quantification of mining-induced ground fissures via UAV imagery based on MDE-UNet

Ground fissures caused by intensive underground coal mining threaten both ecological environments and mine safety. However, complex surface conditions in mining areas make fissure detection labor-intensive and inefficient. This study innovatively proposes MDE-UNet, a high-precision segmentation and accurate quantification model for ground fissures in mining areas. The model integrates deformable convolutions with a dual-attention mechanism through attention convolution blocks, enhancing adaptive perception of irregular boundaries and complex topologies. A deep multi-scale attention interaction module captures cross-resolution contextual semantics via dynamic scale selection, while attention-guided dense skip connections filter redundant shallow details and facilitate efficient deep-shallow feature fusion. Furthermore, skeletonization and Euclidean distance transformation are employed to automatically extract fissure geometric parameters, including length, width, and area. Experimental results demonstrate that the proposed method outperforms representative segmentation models on the UAV mining-area dataset, achieving an F1-score of 86.15% and an IoU of 76.11%. The quantified geometric parameters of the fissures show high consistency with manually measured field data. Its reliability and generalization capability were further validated on additional datasets, including Crack500, DeepCrack, and CrackForest. This study provides a new approach for automated, high-precision monitoring of mining-induced ground fissures and supports geological hazard warning and ecological restoration.