Automatic Detection and Calculation of Mining Subsidence in Large-Scale Interferograms With Transformer-CNN Model

Coal mining operations trigger geological hazards including ground subsidence, surface fracturing, and landslides, endangering resident safety and infrastructure in mining regions. To address this, rapid large-scale detection and precise quantification of mining-induced subsidence are critical. Therefore, this study develops a wide-area subsidence calculation model integrating convolutional neural network and transformer. Initially, a training sample set for mining subsidence interferometric synthetic aperture radar was constructed by leveraging the probability integral method, Generic Atmospheric Correction Online Service, coherence maps, and zero-mean Gaussian noise. Subsequently, the Residual Convolution Block and Swin Transformer Block were adopted as the fundamental building blocks of the model architecture to develop RAUNet, a synergistic network designed for detecting and calculating wide-area mining subsidence zones. Simulation experiments revealed that the RAUNet recognition branch achieved a mean pixel accuracy and mean intersection over union of 0.928 and 0.867, respectively, with the calculation branch attaining a root mean square error of 0.38. These metrics significantly outperformed conventional methods and alternative deep learning approaches. Validation experiments utilizing Sentinel-1, ALOS-2, and LT-1 satellite data demonstrated that: 1) RAUNet enables precise detection of subsidence zones followed by calculation applied exclusively to the interferogram tiles within those zones, significantly improving data-processing efficiency and more intuitively revealing surface deformation in mining areas; 2) RAUNet can adapt to the differences between C-band and L-band data, yielding superior accuracy compared with traditional methods across multiple interferometric pairs. Therefore, the proposed method shows significant potential for studying and preventing geologic hazards in mining areas.