The potential for improved SAR image coregistration using regularised speckle tracking

Abstract

Coregistration is the process of aligning multiple images with sub-pixel accuracy so that corresponding features in different acquisitions match spatially. This step is essential for a wide range of Synthetic Aperture Radar (SAR) applications, including interferometry, for digital elevation model generation, deformation monitoring of glaciers and change detection in environmental or disaster management. This report investigates whether the commonly used method of quantifying spatial distortions via cross-correlation can be improved by using a gradient-based optimisation technique. Developed at DLR's Microwaves and Radar Institute, the algorithm is supposed to be used for the coregistration process of SAR data. The algorithm improves upon the conventionally used speckle tracking algorithm by incorporating prior knowledge into the shift estimation process. The approach utilises the Python framework JAX for automatic differentiation for the optimisation technique. The proposed method was initially developed, tested, and refined in one dimension using simulated data. In the next step, it was validated on real SAR data acquired over the Aletsch glacier. The algorithm comprises several cost functions, an Image Pyramid and the ability to use multiple image pairs for the shift estimation. Results demonstrate that the proposed algorithm is capable of producing consistent results, has improved noise resistance, and is more flexible than its conventional counterpart. In particular, it achieved comparable or better correlation results while complying better with prior knowledge and offers greater adaptability to complex scenarios compared to speckle tracking.