Automatic Detection of Subsidence Troughs in SAR Interferograms Based on Convolutional Neural Networks

In this letter, we present research on automatic detection of subsiding troughs caused by underground coal exploitation using deep convolutional neural networks. The problem differs from typical object detection tasks. Many troughs are hardly visible, and even a careful human annotator overlooks many of them in large and noisy synthetic aperture radar (SAR) images. For this reason, the training set in incomplete, and some troughs correctly found by the network are regarded as false detections, so training is ineffective. We proposed interactive completion of the data set in the training process, and this was crucial for proper training of the network. We developed two alternative systems. The first is based on single-shot detection (SSD) architecture with a VGG network, which achieved an area-under-curve (AUC) value of 0.89. The second, based on TinyYOLOv2, had an AUC value of 0.87 but was more than 10 times faster. Based on the related literature, the proposed systems are first detectors of subsiding troughs in SAR interferograms, the performance of which surpasses human ability of detection and is sufficient for fully automatic, unsupervised operation.