Advanced Radar Sounder Data Analysis Methods under Limited labeled Data Constraints

Radar sounders are active remote sensing instruments that transmit electromagnetic waves and record the echoes reflected from the surface and the subsurface layers. They generally operate at low frequencies, typically within the High-Frequency (HF) to Very- High-Frequency (VHF) range, and use relatively wide bandwidths. This frequencies enables penetration into ice, soil, and rock, making radar sounders highly effective for investigating subsurface environments that are otherwise inaccessible. These instruments have been extensively used in Earth observation to study ice-sheet dynamics, bedrock topography, and subglacial hydrology, as well as in planetary exploration missions to reveal hidden geological structures. However, the vast archives of radargram data produced by these instruments remain largely underexplored. Traditional manual analysis is impractical due to the scale of the data, leaving much of the information contained in these archives unexploited. To address this challenge, several automatic classification methods have been developed. However, most of them depend on large labeled datasets and work only when trained and tested on data from the same campaign, which limits their ability to generalize to new environments. This creates the need for methods that can operate effectively under weak supervision and limited labeled data, while also generalizing across different environments. This thesis addresses these challenges through a progression of frameworks that advance from classical deep learning to foundation models, with an emphasis on data efficiency and limited supervision. The first contribution presents a spatially aware few-shot learning framework for pixel-level classification of radargrams. By integrating spatial and sequential context, this approach enables robust classification with only a small number of labeled examples, addressing the problem of the extreme scarcity of annotations in radar sounder archives. The second contribution introduces scribble driven semi-supervised semantic segmentation using an efficient u2net architecture. First, we develop a parameterefficient u2net architecture that achieves high accuracy even with limited labeled data. Building on this, we introduced a scribble-driven semi-supervised semantic segmentation framework that propagates sparse scribble annotations across large radargram collections while leveraging semi-supervised consistency training. This approach significantly reduces annotation costs and accelerates large-scale subsurface mapping. The third contribution focuses on introducing the use of foundation models for the analysis of radar sounder data to improve generalization across different datasets. We first develop a weakly supervised framework for bedrock detection that combines text and image inputs using a vision–language model with gradient-based explainability to automatically generate prompts. This approach allows accurate segmentation with minimal superviision and strong generalization across domains. We then extend this work by fine-tuning Segment anything (SAM) the vision foundation model for multi-class segmentation of radar sounder data. To improve efficiency, we employ Low-Rank Adaptation (LoRA) together with an Atrous Spatial Pyramid Pooling (ASPP) based multi-class decoder, which substantially reduces the number of trainable parameters while maintaining high segmentation performance. This design enables effective model adaptation without the need for full retraining. Collectively, these contributions establish scalable and transferable solutions for radargram analysis, enhancing the scientific value of radar sounders for climate, cryosphere research, and planetary research. Comprehensive qualitative and quantitative experiments on both terrestrial and planetary radargrams confirm the effectiveness of the proposed methods compared to state-of-the-art approaches.