Deep Learning for the Detection and Phase Unwrapping of Mining-Induced Deformation in Large-Scale Interferograms

Abstract

This article proposes deep convolutional neural networks to detect and map localized, rapid subsidence caused by mining activities using time-series Sentinel-1 synthetic aperture radar (SAR) images. A deformation detection network (DDNet) is developed to automatically identify rapidly subsiding areas from wrapped interferograms, and a phase unwrapping network (PUNet) is designed to unwrap the cropped interferogram patches centered on the detected subsiding locations. To train the two networks, interferogram simulation strategies are developed to generate various training samples using the distorted 2-D Gaussian surface and fractal Perlin noises. The performance of the DDNet is verified by simulations on a synthetic dataset with 13 large interferograms, while the PUNet is evaluated by simulations using synthetic datasets with different levels of deformation gradients and noises. Compared with the traditional and deep-learning methods, the PUNet exhibits excellent performance and efficiency in unwrapping interferograms with rapid mining-induced deformation. The proposed networks are further verified by applying them to Shanxi province, China, which is characterized by serious ground subsidence hazards caused by long-term coal mining activities. The time-series deformations of 1344 detected subsidence areas are calculated with the vertical velocities ranging from −19.7 to −254.8 cm/year. The results are validated using the ascending and descending Sentinel-1 Interferograms and an L-band ALOS-2 interferogram covering the same area within the acquisition period, showing highly consistent vertical deformation rates. The proposed strategy and methods introduce deep learning to the time-series interferometric SAR (InSAR) processing chain and may have profound implications on the detection and monitoring of localized mining-induced deformation using InSAR.