LDM-PUNet: A Lightweight Network for Denoising and Phase Unwrapping SAR Interferograms in Mining Deformation Monitoring

Interferometric synthetic aperture radar (InSAR) enables large-scale, all-weather, day-and-night monitoring of surface deformation, but phase unwrapping remains challenging in mining areas with large-gradient deformation. Most conventional phase unwrapping methods rely on the Itoh condition. In mining interferograms, dense fringes, low coherence and deformation-related noise can violate the Itoh condition, causing unwrapping errors to propagate into fragmented phase fields and unreliable deformation estimates. To address this problem, we propose a lightweight dilated multi-path phase unwrapping network, LDM-PUNet, for joint interferogram denoising and phase unwrapping in low-coherence mining environments. LDM-PUNet introduces multi-path parallel residual blocks with dilated depthwise separable convolutions to capture multi-scale fringe structures while reducing model complexity, and combines attention-based feature refinement with a phase-aware compound loss that integrates robust phase regression, wrapped-phase consistency and gradient consistency. To alleviate the shortage of labelled interferograms for mining deformation, we further develop a multi-effect deformation interferometric phase simulation strategy, M-DIPS, which generates training samples with controllable deformation, terrain, scattering, atmospheric and noise-related effects. Simulation tests were conducted on synthetic datasets with different deformation gradients and noise levels. LDM-PUNet improved RMSE accuracy by approximately 32.2–84.0% compared with the reference methods, while requiring only 0.02 s to process a single sample, demonstrating superior accuracy and efficiency. Real-data experiments in the Datong mining district and the 1071 working face of the Liangbei Coal Mine further demonstrate that, in long-term InSAR deformation monitoring, LDM-PUNet improves phase continuity and deformation inversion accuracy under dense fringes and decorrelation, producing highly consistent vertical displacement estimates. The proposed strategy and methods introduce deep learning into the time-series InSAR processing chain, providing an efficient and robust solution for rapid deformation monitoring in mining areas with large-gradient deformation.