A Supervised Multi-Task Learning Architecture for Separating the Phase Contributions in InSAR Burst Modes

Abstract:
Multi-swath SAR systems are attractive solutions for monitoring the large-scale motions occurring over non-stationary areas. The main limitation of such interferometric systems is the variable sensitivity along the flight direction, which results in phase jumps between adjacent bursts in the interferograms. In this paper, we present a convolutional neural network that decouples the interferometric phase from the along-track phase contribution by simultaneously solving multiple tasks, (1) separating the phase due to displacements in the line-of-sight direction from that due to displacements in the along-track direction, and (2) predicting a proxy for the along-track displacement. The benefits of the proposed algorithm are verified using Sentinel-1 TOPS interferometric pairs over Greenland to track the inland glacier flow occurring within a time frame corresponding to the revisit time.