MoDL-PU: Model-Based Deep Learning for InSAR Phase Unwrapping

Phase unwrapping (PU) is a critical process for numerous synthetic aperture radar interferometry (InSAR) applications. The advent of deep learning (DL) has revolutionized PU, with various DL-based methods emerging over the past four years that consistently achieve state-of-the-art results. However, the generalization capability of these methods is impeded by the lack of a large-volume training dataset for InSAR PU. In contrast, model-based PU methods, both single-baseline (SB) and multibaseline (MB), utilize mathematical formulations that represent the PU knowledge, often resulting in robust generalizability in their respective domains. This article introduces MoDL-PU, a hybrid model-based/DL-based PU method. MoDL-PU, leveraging the Res-UNet-Inception architecture, employs a two-stage training approach: initial dataset-based training to extract low-frequency unwrapped phase information, followed by retraining with three specialized PU knowledge to enhance high-frequency details for targeted PU tasks: SB PU (MoDL-SBPU), MB PU for Digital Elevation Model (DEM) reconstruction (MoDL-MBPUT), and MB PU for deformation monitoring (MoDL-MBPUD). The proposed method exhibits robust generalizability, interpretability, and noise resilience, effectively combining the advantages of model-based and DL-based PU methodologies. The experimental outcomes demonstrate that MoDL-PU not only exceeds the performance of conventional model-based PU methods but also competes favorably with state-of-the-art supervised and self-supervised DL-based approaches.