KG-SubNet: A Knowledge-Guided Deep Learning Framework for 3-D Mining Subsidence Retrieval From InSAR Interferograms

Abstract

High-precision mapping of 3-D surface displacement in mining areas is crucial for geological hazard prevention and control, as well as for ecological environment assessment. Most existing interferometric synthetic aperture radar (InSAR)-based methods require phase unwrapping (PU), which is often compromised by complex surface environments in mining areas, including dense vegetation and large deformation gradients, thereby reducing the accuracy and reliability of the final results. To address these issues, this study proposes a novel deep learning framework that retrieves 3-D surface displacement end-to-end directly from wrapped interferograms, completely bypassing PU. The core of the framework is the knowledge-guided subsidence network (KG-SubNet), which comprises three key components: 1) knowledge-guided training sample generation, which uses mechanistic subsidence model to synthesize large-scale training datasets and address the problem of training sample scarcity; 2) an innovative architecture integrating U-Net and a high-resolution network, which aggregates multiscale features from all encoder stages to preserve the high-resolution spatial details of deformation fringes; and 3) a knowledge-guided loss function that incorporates phase gradient priors and subsidence patterns to improve model reliability in low-coherence regions and enhance generalizability. Simulation experiments and engineering case studies demonstrate that the proposed method accurately retrieves 3-D surface displacement patterns even under low-coherence conditions. Compared with existing PU-based methods, this approach reduces root-mean-square error (RMSE) by a minimum of 45.3% and processes a single interferogram in under one second. This study establishes a PU-free deep learning framework for 3-D surface displacement retrieval in coal mining areas, offering an efficient and reliable solution for geohazard monitoring.