Giacomo Nodjoumi

• The study of pits, skylights, and sinkholes on terrestrial bodies is vital for mapping subsurface voids, geological evolution, and potentially habitable zones. Feature identification remains constrained by sensor resolution and the complexities of radar analysis. This research explores deep learning computer vision to automate landform mapping and evaluates detecting subsurface voids using orbital radar. To address mapping challenges, this work introduces DeepLandforms, a toolkit developed to automate detection using You Only Look Once (YOLO), Detectron 2, and the Segment Anything Model (SAM). Built on Docker, it includes modules for data preparation, training, and inference, generating outputs compatible with GIS software. Validation against a dataset from the Mars Global Cave Candidate (MGC³) catalog demonstrates capacity for consistent, large-scale surveys. This research also introduces EchoTerraeTrace, an all-in-one toolkit providing workflows for SHARAD (Mars Reconnaissance Orbiter) and MARSIS (Mars Express) data. During validation North-West of Ascraeus Mons, three unmapped volcanic vents were identified, and the regional paleotopography was refined. Finally, this research assesses planetary software against FAIR principles. A web-based service built on a dockerized JupyterHub was developed as an all-in-one environment for standardized data processing. This service provides scalability without requiring local high-end resources, supporting reproducible research by moving the code to the data rather than the data to the code.