Learning a versatile representation of SAR data for regression and segmentation by leveraging self-supervised despeckling with MERLIN

Abstract:
Synthetic Aperture Radar (SAR) images are abundantly available, yet labels are often missing. Thus, training a neural network in a fully supervised manner is arduous. In this work, we leverage MERLIN, a self-supervised despeckling algorithm, to learn a mapping of SAR images into a representation space shared among despeckling, building segmentation and height regression. Our experiments demonstrate that the joint training of a neural network for these three tasks reduces considerably the need for labeled data to solve the supervised tasks: positive results are obtained even when only 1% of the dataset is annotated.