Simplicity Beats Complexity in Physics-Constrained InSAR Phase Unwrapping

A recent investigation published on arXiv questions the prevailing preference for complex computer vision architectures in the context of physics-constrained InSAR phase unwrapping. This research represents the first extensive architectural ablation analysis on a global LiCSAR benchmark, which includes 20 frames, 39,724 patches, and a total of 651 million pixels. The findings indicate that a standard U-Net with 7.76 million parameters surpassed attention-based models, which have 11.37 million parameters, achieving an R² of 0.834 and an RMSE of 1.01 cm—showing improvements of 34% in R² and 51% in RMSE. Additionally, Power Spectral Density analysis showed that attention mechanisms introduce unphysical high-frequency artifacts, breaching the smoothness constraints essential for geophysical fields. This research highlights the potential effectiveness of simpler models for operational phase unwrapping in monitoring volcanic and seismic activity.