MAPUNet: Multi-scale attention for InSAR phase unwrapping in mining areas

Abstract

Due to the large deformation gradient caused by large-scale and high-intensity mining in the mining area, severe incoherence is prone to occur during the radar interferometry processing. The traditional phase PU method cannot obtain the correct phase unwrapping (PU) results, resulting in inaccurate monitoring results. To achieve precise monitoring of large gradient deformations in mining areas, a multi-scale phase PU method for mining area based on PUNet with multi-attention mechanism (named MAPUNet) was proposed in this paper. Firstly, the atrous spatial pyramid pooling (ASPP) is added to the original dilated block, and combined with dual attention network (DA-Net), which is used to extract the interferogram feature information at multiple scales. Then, an efficient channel attention network (ECA-Net) optimized residual module is designed to enable the network to extract important interferogram features at both branches and suppress phase noise. Three simulated interferograms with different signal-to-noise ratios (SNR) were selected for the experiment. Under conditions of large phase gradients changes and high noise (SNR = 1), compared with the optimal PUGAN method, SSIM increased by 0.0176, compared with the original PUNet model, SSIM increased by 0.298. The model in this paper has higher accuracy and better robustness, and achieves rewrapped results that are closer to the real situation in the real mining area interferogram. Finally, based on the MAPUNet model, the deformation inversion of the certain mining area in Shanxi has been achieved. The maximum settlement in this area within one month can reach −9.3 cm. Our method effectively solves the problems existing when large-scale subsidence in the mining area, providing a new technical reference for more accurate monitoring of surface deformation.