Self-Supervised Network-Based Atmospheric Phase Compensation Method for GB-InSAR

In ground-based interferometric synthetic aperture radar (GB-InSAR), the atmospheric phase (AP) can exhibit complex spatial variation due to rapid weather changes and steep topography. Conventional parametric-model-based compensation methods become inapplicable. Existing deep learning methods mostly rely on full supervision, which limits their interpretability and generalization performance. To address these issues, this article, based on the integral relationship between AP and refractivity variation, proposes a self-supervised network-based AP compensation (APC) method for GB-InSAR, which accomplishes APC by estimating gridded refractivity variation. First, the method constructs a deep residual network that takes the sequence of unwrapped phase images and other auxiliary images as input and outputs the gridded refractivity variation for each unwrapped phase image. Second, two main loss terms for self-supervision are constructed. For nondeformation persistent scatterers (PSs), the main loss term 1 is built using their unwrapped phases. For deformation PSs, the main loss term 2 is built by constraining the temporal continuity of the phases after APC. Third, prior information, such as spatial continuity and value ranges, is embedded into the self-supervised iteration process in the form of loss terms and network modules. Finally, the gridded refractivity variation is solved via gradient backpropagation and iteration in a self-supervised manner, thus achieving APC. Results on both simulated and measured data demonstrate that the method effectively compensates for spatially varying AP, provides high stability, and overcomes the edge effect.