Mining Area Ground Fissure Identification Method Based on AR-SE-ResUNet

The accurate identification of ground fissures in mining areas is crucial for geological disaster prevention and control and ecological restoration. To address the low efficiency and poor adaptability of traditional methods and the insufficient accuracy of the existing deep learning models in identifying ground fissures under complex scenarios, this letter proposes an AR-SE-ResUNet model that integrates adaptive rectangular convolution (ARConv) with the squeeze-and-excitation (SE) attention mechanism. Using uncrewed aerial vehicle (UAV) images from the Huipodi mining area in Shanxi Province, a sample dataset comprising 390 images with manually annotated ground fissures was constructed. The model employs ARConv to dynamically adjust convolution kernel sizes for enhanced multiscale feature extraction, utilizes the SE mechanism to perform channelwise reweighting for suppressing background interference, and combines the encoder-decoder architecture of ResUNet to achieve hierarchical feature fusion. Experiments demonstrate that the AR-SE-ResUNet model significantly outperforms comparative models in terms of IoU (87.84%), $F1$ -score (93.52%), and precision (88.30%), while exhibiting high robustness in detecting slender, discontinuous fissures under complex background scenarios. The model provides a novel approach for efficient identification of ground fissures in mining areas through adaptive multiscale feature extraction and noise suppression strategies, effectively supporting geological disaster monitoring and ecological restoration practices.