Hierarchical Learning for the Unsupervised Segmentation of Radar Sounder Data Acquired on the Cryosphere

In the radar sounder literature, extracting subsurface geological information relies on supervised deep learning with large labeled datasets. While some methods reduce the need for extensive labels through weak supervision, there remains a gap in the availability of unsupervised segmentation techniques. This paper proposes a novel method for unsupervised radargram segmentation based on incremental learning (IL). The method involves the prior geophysical modeling of the cryosphere subsurface targets into a class hierarchy. Through several IL steps, a network is trained to progressively extract semantically meaningful features that are analyzed to compute the segmentation map. Each step refines the segmentation map by considering the new targets from the following level of the class hierarchy that details the targets at the previous level. To enhance the training process, contrastive learning is incorporated, along with techniques for distilling information from prior iterations to recall the network the properties of previously seen classes. To validate the effectiveness of the proposed method, we conducted successful experiments on MCoRDS-3 data acquired in Greenland.