Detection and analysis of potential planetary cave entrances in remote-sensing data using machine and deep learning

Machine and deep learning (ML/DL) techniques have been successfully applied to imagery across numerous fields, including Earth observation. However, prior to this thesis, the usage of ML/DL in the domain of remote-sensing data taken of the Moon and Mars was dominated by impact crater detection algorithms. This is in spite of the numerous other interesting and relevant features on these surfaces, which hadn’t received the same treatment. This is addressed within this thesis through three separate projects that employ ML or DL techniques upon planetary surface features. The features in question are pits, skylights, and impact melt fractures, which are significant for astrobiological investigation and space exploration for their potential to be entrances to sub-surface cavities.

In the first project, the Pit Topography from Shadows (PITS) tool is described, which automatically derives apparent depth profiles of Lunar and Martian pits and skylights using only a single visual-band satellite image. PITS does this by utilising k-means clustering and silhouette analysis to detect shadow pixels (achieving precision and recall rates of 99.6 and 94.8% upon shadow-labelled HiRISE imagery) under a range of illumination conditions. The known image resolution and positions of the Sun and satellite are used to produce apparent depth measurements (corrected for non-nadir observations) along the entire length of the shadow. Thanks to the PITS tool, 10 Mars Global Cave Candidate Catalog features were discovered to exhibit signs of possible cave entrances in their depth profiles. The second project trains Mask R-CNN (Region-based convolutional neural network) instance segmentation DL models to detect pits and skylights on the Moon, with the objective of bolstering the existing Lunar Pit Atlas. During testing, the best model (which was trained on Lunar, Martian, and synthetic Lunar data with a ResNet50 backbone) achieved F1-scores (F1) of 82.4 and 93.7% for the bounding boxes and predicted masks, respectively. Despite having only been applied to ≈1.9% of the total Lunar maria, this model—named ESSA (Entrances to Sub-Surface Areas)—has found two previously uncatalogued skylights on the Moon, with one situated ≈58.7◦ in latitude and the other being found in the previously mapped Marius Hills. ESSA has also been applied to the January 2025 release of HiRISE RDRV11 images. The third project presents the results of training a DeepLabV3 semantic segmentation DL model to detect cooling fractures found within the melt deposits of impact craters on the Moon. In testing upon impact melt fractures (IMFs) found within the Copernicus and Virtanen F craters, the highest F1 of 55.9% was achieved by the DeepLabV3 trained with a ResNet50 backbone and a batch size of 32—the combination of which is named IMFMapper. IMFMapper has been deployed to map the IMFs within the previously surveyed Crookes crater, where it found new candidate melt deposits within the crater’s western and southern walls. Moreover, IMFMapper has produced the first mapping of any kind for Schomberger A crater, where IMFs may act as permanently shadowed regions due to its proximity to the Lunar south pole. This project also represents an avenue for future work due to the prospect of BepiColombo imaging IMFs on Mercury with greater detail.