Reliability-adaptive graph-cut phase unwrapping for InSAR geohazard deformation monitoring

Phase unwrapping is critical for InSAR deformation retrieval, yet decorrelation, dense residues, and Itoh-condition violations reduce reliability in high-gradient geohazard scenes. To address these challenges, this study proposes EGCPU, a training-free, reliability-adaptive graph-cut method that estimates integer phase ambiguities on a pruned weighted graph constructed from coherence, wrapped gradients, and local circular dispersion. A Huber-penalized differential energy is minimized through bidirectional unit expansion – cut moves solved with Incremental Breadth-First Search max-flow. Evaluations using controlled simulations, two station-validated real-data cases, and an additional Kumamoto coseismic generalization test show that EGCPU improves phase-unwrapping stability in heterogeneous geohazard scenes. EGCPU achieves 5.44 rad RMSE and 0.80 SSIM in the highest-noise simulation, reduces Longs Peak RMSE by about 65% relative to SNAPHU and PUMA, and obtains 1.81 cm LOS RMSE against LOS-projected GNSS observations and 1.10 mm LOS MAE against LOS-projected extensometer measurements, supporting stable phase unwrapping in heterogeneous geohazard scenes.