A Mask R-CNN Network for Wide-Area Mining Subsidence Automatic Detection With InSAR Observations

Abstract

Land subsidence caused by mining activity is one of the most serious anthropogenic geohazards. The rapid detection and continuous monitoring of mining subsidence facilitate the swift detection of geohazards. Traditional methods of monitoring mining subsidence have shortcomings, such as offering only a limited coverage and being time consuming. Interferometric Synthetic Aperture Radar (InSAR) has been proven to be a powerful tool to identify mining subsidence hazards from unwrapped interferograms but this method can be complex and inefficient, particularly for wide areas. In this paper, a Mask R-CNN model is presented to automatically detect mining subsidence and monitor the surface activity over wide areas using original SAR interferograms to avoid the time-consuming and error-prone phase unwrapping procedure. Using Sentinel-1 wrapped interferograms as the real dataset and simulated wrapped interferograms generated with the Gaussian surface function combined with Generic Atmospheric Correction Online Service for InSAR (GACOS) as the simulated dataset, the Mask R-CNN deep neural network was used to train the mining subsidence detection model. It turned out that the accuracy of the detection model was 91.48%, while the precision was 96.44%, the recall rate was 93.88% and the F1 index was 0.949. The detection model was then utilized to detect mining subsidence in south-western Shanxi Province, China between 2016 and 2022. A total of 152 land subsidence points were detected and long-term monitoring results were also obtained. An analysis of the state of land subsidence in the mining areas was carried out to obtain land subsidence activity during the monitoring period.