Scribble Driven Semi-Supervised Semantic Segmentation of Radar Sounder Data

Radar sounder (RS) data play a crucial role in exploring subsurface geological features in various terrains, including icy and arid regions both on Earth and other celestial bodies. Existing semantic segmentation methods for RS data analysis mainly rely heavily on dense pixel-wise annotations, which are labor-intensive to produce. This paper addresses this limitation by proposing a novel cost-effective and efficient scribble annotations approach. The proposed approach is based on two stages. The first stage introduces a scribble-guided K-Nearest Neighbors (KNN) method to propagate sparse scribble annotations into dense labels. The second stage utilizes a hierarchical multi-task efficient u2net architecture, which integrates the classification of radargrams into broad geographical classes (like coastal or inland regions) and then perform its segmentation of subsurface features within these classes. Experiments on radargrams acquired in Antarctica demonstrate that the proposed approach provides a cost-effective solution, reducing the need for dense labels while maintaining competitive accuracy in subsurface feature segmentation.