Mining-induced surface cracks are critical indicators of overburden deformation and pose significant threats to ground stability and mine safety. However, their fine-scale geometry and complex background interference make automated extraction challenging. This study proposes an improved DRA-UNet model for high-precision crack detection from UAV orthophotos. The network integrates residual learning, a dual-attention mechanism (DAM), and an atrous spatial pyramid pooling (ASPP) module to enhance feature representation, suppress noise, and capture multi-scale contextual information. A complete analytical framework is established by coupling crack segmentation with skeleton extraction and quantitative geometric characterization, enabling fine-scale morphological analysis. Experimental results show that the proposed method outperforms representative segmentation models, achieving an F1-score of 71.60% and MIoU of 70.00% on a mining-area UAV dataset. Ablation studies verify the effectiveness of each module, while external validation on the Crack500 dataset demonstrates strong cross-scene generalization (F1 = 85.32%, MIoU = 83.69%). Geometric analysis reveals pronounced right-skewed distributions of crack length, width, and area, with rectangularity decreasing as crack length increases. Spatial results indicate higher crack density and complexity near working-face boundaries. Overall, the proposed framework provides a robust and fully automated solution for mining-induced crack detection and morphological analysis.
Zhou et al. (Thu,) studied this question.