Accurate identification of coal mining subsidence areas based on DInSAR and deep learning

Junting Guo^1*Xiaolong WANG²Bin LI²Xiaobin LI¹Teng Teng¹

• ¹National Institute of Clean and Low-Carbon Energy (China), Beijing, China
• ²China Energy Shendong Coal Group Co., Ltd., Yulin, China

Coal mining subsidence poses a serious threat to the ecological environment and safety of mining areas, requiring long-term dynamic monitoring. Traditional methods such as leveling and GPS struggle to meet the demand due to their low resolution, high cost, and limited coverage. InSAR technology, with its advantages of all-weather operation, wide coverage, and millimeter-level accuracy, has become a vital tool for monitoring surface deformation. However, it still faces challenges such as abrupt deformation and phase decorrelation in subsidence areas. Deep learning technology provides a feasible approach for large-scale subsidence zone identification, and its integration with InSAR enables rapid and accurate detection of subsidence areas over extensive regions. Nevertheless, the deep learning process requires extensive labeled samples and massive InSAR data processing, which is time-consuming and inefficient. To improve identification accuracy and efficiency, a method combining simulated interferograms based on the probability integral model and two-dimensional Gaussian surface model with SBAS-InSAR results covering the entire Ningxia region from January 2020 to May 2022 is proposed. Comparative analysis between interferogram-based and deformation velocity map-based identification demonstrates significantly higher accuracy for the former, confirming the feasibility of the simulated interferogram strategy. By integrating interferograms and deformation velocity maps, 144 coal mining subsidence zones are identified in Ningxia, effectively enhancing detection accuracy. Additionally, systematic evaluation of the probability integral model and two-dimensional Gaussian surface model datasets reveals that the former is more suitable for characterizing subsidence features in regular mining areas, while the latter excels in simulating asymmetric deformations under complex geological conditions. The research outcomes provide a scientific methodology for efficient and accurate subsidence zone identification, with potential applications in ecological restoration planning, geohazard early warning, and sustainable mining practices, offering significant practical value for Precise management of mining subsidence areas and balancing resource development and environmental protection.