Interferometric Synthetic Aperture Radar (InSAR) is a key remote sensing technique for monitoring Earth's surface changes, such as topography and ground deformation. InSAR provides phase information that is essential for these measurements; however, this phase data is "wrapped" within a limited range (−π,π], necessitating a process called phase unwrapping to recover the true phase values. Phase unwrapping allows estimating relative distances, thus accurate heights in remote sensing, but this is a challenging ill-posed problem due to noise, phase discontinuities, and rapid phase changes.
Left: raw image acquired from InSAR. Right: Unfolded image representing a digital elevation.
While recent deep learning approaches have demonstrated improved performance in phase unwrapping, many methods—such as those employing Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks—struggle to capture long-range spatial relationships that are critical for accurate unwrapping [1]. Moreover, these approaches may require large datasets and do not fully utilise additional data structures that could enhance performance.
This thesis addresses the design of a novel DL framework that integrates CNNs and GNNs to perform phase unwrapping. Our idea is to process the InSAR image data using a CNN, while the GNN will handle an additional graph-based data structure that captures long-range spatial relationships, which are critical for accurate phase unwrapping. This hybrid approach aims to overcome the limitations of existing methods by leveraging the strengths of both CNNs in local feature extraction and GNNs in modeling complex spatial dependencies.