Automatic Detection for Mining Subsidence Areas Using the CBAM-Enhanced VGG-UNet Model With Long Time Series InSAR Interferograms

Abstract

The technical theories of monitoring and preventing mining subsidence have long been key challenges and research priorities in the mining field. The rapid advancements in remote sensing technology and deep learning algorithms have enabled significant breakthroughs in monitoring and accurately identifying mining subsidence. In this article, a novel automatic detection method for mining subsidence is proposed using interferometric synthetic aperture radar (InSAR) wrapped interferograms. First, this study designs a VGG-UNet model enhanced by an attention mechanism module to learn and detect mining subsidence areas. This enhancement improves the feature representation and perception capabilities of the model. Second, to address the scarcity of real InSAR data in the training set, an efficient dataset simulation strategy is established. This strategy incorporates the realistic scenarios of monitoring the environment to improve the effect of model training. Finally, a complete workflow for model training and detection application is developed. The results demonstrate that the detection model achieves a precision of 92.55%, a recall of 90.43%, an accuracy of 93.37%, an F₁-score of 91.46%, and an intersection over union of 84.25% on the validation set. The model was applied to mining subsidence detection in the Huaibei–Yongcheng mining area, China, from June 2017 to July 2024. A total of 103 mining subsidence sites were identified, and their long-time series characteristics and the spatial distribution pattern of subsidence accumulation duration were analyzed. The findings offer critical technical support for sustainable mining management and land resource protection at the regional scale.