Deep Learning-Based InSAR Time-Series Deformation Prediction in Coal Mine Areas

The goafs left after coal mining can cause destructive surface deformations, such as surface subsidence and ground fissures. Monitoring and predicting surface deformation are essential for coal mine safety and urban sustainability. However, existing mining-induced deformation prediction models often lack effective attention mechanisms for critical time-series features and ignore potential relationships between deformation and external influencing factors. In this paper, we construct a multivariate deep learning model framework for precise surface deformation prediction. This framework integrates a Transformer-encoder module, a Bi-LSTM-decoder module, and an innovative convolutional attention feature extraction module. It can effectively capture both global and key temporal features and dynamically model the interactions among multimodal data. The Hunchun coal mining area is taken as a case study, where operational and closed mines coexist. First, Distributed Scatterer InSAR (DS-InSAR) and Multi-dimensional Small Baseline Subset InSAR (MSBAS-InSAR) methods were integrated to reveal the spatiotemporal distribution characteristics of surface deformation. The proposed model is then applied to predict future surface deformation. Main conclusions include: (1) Significant mining-induced surface subsidence was observed in Yingan, Baliancheng, and Banshi coal mines, while Chengxi coal mine experienced notable uplift possibly related to rising groundwater; (2) Comparisons with benchmark methods indicate that the proposed model achieves smaller errors and better predictive performance; (3) In the next two and a half years, surface deformation in the four coal mining areas is expected to worsen and further expand. The findings provide valuable guidance for risk warning and decision-making specifically in the Hunchun coal mining area.