Multitask Learning for Phase Source Separation in InSAR Burst Modes

The scanning synthetic aperture radar (ScanSAR) and Terrain Observation by Progressive Scans (TOPSs) burst acquisition modes are nowadays among the most widely used in synthetic aperture radar (SAR) satellite missions. Both allow for increased coverage at the expense of azimuth resolution. However, the intermittent nature of the burst acquisition results in an increased sensitivity toward burst edges to displacements in the along-track (AT) dimension. In the presence of azimuth motion in the scene, phase jumps between bursts occur. In this contribution, this increased sensitivity is considered as an opportunity to obtain information on the North-South displacement, in which current SAR systems are less sensitive due to their quasi-polar orbits. Specifically, we suggest the usage of a multitask learning (MTL) architecture trained in a supervised fashion to separate the phase contribution due to displacements in the zero-Doppler (ZD) direction from AT displacements and to further provide a first rough estimation for the along-track displacement. Through an ad hoc network architecture and loss functions, we inject information about the interferometric SAR system model into the learning process, following a machine learning approach. We apply our method to the estimation of inland glacier flow from Sentinel-1 interferometric wide (IW)-swath data. We show that we are able to estimate, with an excellent performance, AT surface displacements of a few centimeters to several tens of centimeters, providing an improvement in accuracy compared with speckle tracking, and in coverage compared with techniques that exploit the burst-overlap differential phase.