A Novel Phase Unwrapping Method for Low Coherence Interferograms in Coal Mining Areas Based on a Fully Convolutional Neural Network

Subsidence caused by underground coal mining activities seriously threatens the safety of surface buildings, and interferometric synthetic aperture radar has proven to be one effective tool for subsidence monitoring in mining areas. However, the environmental characteristics of mining areas and the deformation behavior of mining subsidence lead to low coherence of interferogram. In this case, traditional phase unwrapping methods have problems, such as low accuracy, and often fail to obtain correct deformation information. Therefore, a novel phase unwrapping method is proposed using a channel-attention-based fully convolutional neural network (FCNet-CA) for low coherence mining areas, which integrates multiscale feature extraction block, bottleneck block, and can better extract interferometric phase features from the noise. In addition, based on the mining subsidence prediction model and transfer learning method, a new sample generation strategy is proposed, making the training dataset feature information more diverse and closer to the actual scene. Simulation experiment results demonstrate that FCNet-CA can restore the deformation pattern and magnitude in scenarios with high noise and fringe density (even if the phase gradient exceeds
π
). FCNet-CA was also applied to the Shilawusu coal mining area in Inner Mongolia Autonomous Region, China. The experimental results show that, compared with the root mean square error (RMSE) of phase unwrapping network and minimum cost flow, the RMSE of FCNet-CA in the strike direction is reduced by 67.9% and 29.5%, respectively, and by 72.4% and 50.9% in the dip direction, respectively. The actual experimental results further verify the feasibility and effectiveness of FCNet-CA.