A Convolutional Recurrent Bottleneck and Vertical Nonlocal Operations for the Weakly Supervised Semantic Segmentation of Radar Sounder Data

The increasing number of operating radar sounders (RS) requires fast and reliable automatic data processing methods. In this context, the semantic segmentation of radargrams enables the identification of key subsurface features. While deep learning (DL) is increasingly employed to enhance generalization, few approaches treat radargrams as sequences of correlated columns, despite the natural continuity between adjacent rangelines. Previous sequential methods model this correlation statistically, using techniques such as regression or Markov chains. Moreover, only a few models aim to reduce the number of labeled samples required during training. In this paper, we propose a methodology for the weakly-supervised semantic segmentation of radar sounder data, introducing a novel convolutional recurrent bottleneck to process radargrams. This recurrent unit allows a U-shaped architecture to capture horizontal correlations across the radargram, which is represented as a sequence of columns. Crucially, it also enables the application of newly introduced cycle-consistency-based loss functions, allowing training with just one labeled column per sequence, significantly reducing labeling requirements. We further enhance our architecture with a fast non-local operation to encode the vertical relationships between semantic classes. We validate our approach through extensive experiments on two datasets of radargrams acquired in Antarctica. Results demonstrate the effectiveness of the proposed methodology in weakly-supervised settings and confirm the competitiveness of our architecture even under full supervision.