Advanced Deep Learning Methods for the Automatic Analysis of Radar Sounder Data

The growing integration of radar sounders (RS) into Earth and planetary exploration missions has generated vast archives of subsurface observational data, essential for monitoring the cryosphere and understanding planetary geological history. These instruments provide unique cross-sectional views of the subsurface. However, the automatic analysis of the resulting radargrams remains a challenge due to the complex nature of the data and the limited availability of labeled examples. While Deep Learning (DL) has revolutionized image analysis in many domains, its application to RS is limited by the scarcity of labeled data, the presence of different noise sources, and the uncommon characteristics of the data. This limits the effectiveness of standard supervised models both for detecting and delineating targets within the data, and to perform quantitative parameter estimation. This thesis addresses this critical challenge by developing a suite of novel methodologies for the analysis of RS data, focusing on data efficiency, expert interaction, and physical consistency. The thesis is structured around three primary contributions addressing the limitations of current automated interpretation methods. First, we tackle the scarcity of labeled data for semantic segmentation by introducing an interactive framework based on unsupervised random walks and user-guided label propagation. This approach formulates feature learning as a random walk process on the radar image graph, effectively leveraging the strong horizontal correlations in the data. It allows domain experts to drive large-scale, consistent segmentations using only sparse and intuitive annotations, significantly reducing the manual effort required for dataset creation. Second, we address the challenge of learning from weak supervision by designing a novel deep neural architecture incorporating a convolutional recurrent bottleneck and vertical nonlocal operations. This method explicitly models the specific structural priors of RS data, such as the continuity of layers along the flight track and the ordered vertical sequence of subsurface materials. By encoding these properties using ad-hoc mechanisms directly into the network, the model achieves robust segmentation performance even when trained with incomplete labels, and goes on par with fully supervised approaches. Third, we extend the analysis from segmentation to quantitative geophysical parameter inversion by introducing a Simulation-Based Inference (SBI) framework utilizing Neural Posterior Estimation (NPE). This approach integrates a GPU-based electromagnetic simulator into a probabilistic inversion pipeline, allowing for the estimation of terrain parameters with rigorous uncertainty quantification. Unlike traditional inversion techniques, this method approximates the full posterior distribution of the parameters, providing uncertainties and correlations among them, and effectively handles the lack of absolute calibration in orbital RS data through a relative power formulation. The effectiveness of the presented methodologies has been validated on real-world datasets acquired by the MCoRDS and SHARAD instruments. Experimental results consistently show superior performance in terms of accuracy, generalization capabilities, and robustness when compared to existing state-of-the-art techniques. Our SBI framework successfully retrieves geophysical parameters consistent with independent measurements and literature-based estimates, demonstrating its practical applicability. Overall, this work represents a significant step forward in the automatic interpretation of RS data, offering solutions that minimize the reliance on large-scale supervision and empowering the scientific community to leverage the growing archives of subsurface observations.