Abstract: China′s underground coal mines are large and wide,and underground mining is highly concealed. The existing methods of manual investigation,remote sensing detection and field measurement are difficult to meet the requirements of automatic identification of large-scale mining subsidence areas,which is not conducive to the efficient supervision and dynamic monitoring. Therefore,this paper proposes an automatic identification method of mining subsidence area based on U2-Net wide-area synthetic aperture radar interferometry (InSAR). This method trains convolutional neural network (CNN)through simulation data sets of various deformation gradients and noise levels,so that it can output a binary matrix containing mining subsidence location information in one step from the differential interferogram. The test results show that the mean pixel accuracy (MPA) and mean intersection over union (MIoU) of U2-Net reach 0. 916 3 and 0. 911 9,respectively,which are higher than the other two models in the experiment. It can better suppress noise and highlight deformation signals. On the InSAR interferograms covering the Shendong mining area at different time intervals,U2-Net automatically identified interferograms covering an area of more than 54 600 km2,and detected multiple subsidence areas with clear and smooth boundary information. The average accuracy of recognition reached 92. 45%. The results show that compared with other networks,U2-Net can fuse multi-scale and multi-level features with less computation through a two-stage nested U-shaped structure,which has significant advantages in noise suppression and deformation region recognition. It can be indicated that joint deep learning can serve the detailed investi-gation of refined mining subsidence areas,promote the application of InSAR technique,and provide a new technical method for intelligent identification of wide-area mining subsidence areas.

Key words: mining subsidence,deep learning,InSAR,U2-Net,semantic segmentation