InSAR Phase Unwrapping Methods Based on Integrated Multi‐Model Deep Learning

ABSTRACT

Interferometric Synthetic Aperture Radar (InSAR) serves as a fundamental tool for deformation analysis within Geographic Information System (GIS)‐based geohazard monitoring. Phase unwrapping (PU) remains a critical stage for retrieving high‐precision topographic and displacement data. The accuracy and efficiency of PU directly dictate the reliability of spatial modeling and emergency responses to geological hazards including landslides and land subsidence. Conventional algorithms typically lack robustness under conditions of intense noise, low‐coherence, and steep topography, thereby undermining hazard assessments. To enhance geospatial fidelity and topological consistency, this research introduces and evaluates three independent deep learning (DL) architectures: UA, GL, and CUA. These models leverage U‐Net's multi‐scale spatial fusion, generative adversarial networks for texture synthesis, and channel attention mechanisms for adaptive feature recalibration. Evaluations on a large‐scale InSAR benchmark dataset reveal that these architectures outperform traditional methods in structural fidelity across challenging interferometric conditions. Moreover, they achieve temporal consistency comparable to network flow algorithms in active tectonic scenarios while isolating localized phase errors to halt global propagation, consistently yielding a peak signal‐to‐noise ratio exceeding 28 dB and a structural similarity index measure above 0.85. These findings provide systematic guidelines for architecture selection in DL‐based PU workflows for geohazard monitoring applications.