Enhancing phase unwrapping for encoded InSAR interferograms

Abstract

Interferometric Synthetic Aperture Radar (InSAR) is a remote sensing technique for measurement of surface topography and ground deformation under all-weather, day- and-night acquisition conditions. Despite its effectiveness, the interferometric phase is wrapped within the interval (−π, π], and recovering the absolute phase required for geophysical interpretation remains a challenging ill-posed inverse problem. Phase un- wrapping is sensitive to noise, decorrelation, atmospheric disturbances, and geometric distortions, which can lead to error propagation in large-scale and high-resolution inter- ferometric datasets. This thesis aims to enhance the robustness, accuracy, and scala- bility of InSAR phase unwrapping by developing two complementary approaches capable of improving performance in noisy, low-coherence, and complex deformation scenarios. The first proposed approach introduces an optimization-based phase unwrapping frame- work using the Multi-Verse Optimizer (MVO). In this method, phase unwrapping is formulated as a global optimization problem, and an optimization-guided quality map is developed to adaptively guide the unwrapping path and reduce error propagation in un- reliable regions. The second proposed approach presents VOH-Net, a Vision-Optimized Hybrid Network designed for InSAR phase unwrapping. The architecture combines convolutional neural networks for local spatial feature extraction with enhanced multi- directional Long Short-Term Memory (LSTM) modules to capture long-range spatial dependencies and preserve global phase consistency across large interferograms. Exper- imental evaluations conducted on both synthetic and real InSAR datasets demonstrate that the proposed methods outperform classical, optimization-based, and learning-based phase unwrapping techniques. In particular, the enhanced VOH-Net achieved 78.4%, 94.7%, and 79.0% reductions in the Mean Absolute Error (MAE), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE), respectively, while improving the Struc- tural Similarity Index Measure (SSIM) by 3.8% compared with the original loss formula- tion. In addition, the proposed approaches exhibit robustness to noise and decorrelation, and improved reconstruction accuracy over large and highly deformed interferograms. The findings of this thesis demonstrate that integrating global optimization with deep learning provides an effective and reliable strategy for InSAR phase unwrapping. The MVO-based framework enhances robustness through adaptive optimization-guided path selection, whereas VOH-Net improves scalability and global phase consistency through hybrid spatial learning. Together, these complementary contributions advance the state of the art in InSAR phase unwrapping and provide a robust foundation for generating more accurate InSAR-derived geophysical products.