Comparative analysis of mask R-CNN and U-Net architectures using ResNet as backbone for lunar crater detection

Automated detection of lunar craters is crucial for advancing planetary science, enabling efficient geological mapping, surface age estimation, and resource identification. This study compares Mask R-CNN (instance segmentation) and U-Net (semantic segmentation) architectures using ResNet as the backbone for lunar crater detection. Key novelty is comparing model performance in both a Geospatial context (ArcGIS Pro environment) and non-Geospatial environment a method not heretofore attempted. Training and validation were conducted using Geocoded Chandrayaan-2 TMC-2 DEM data, employing a new strategy that facilitates accurate localization and precise detection of small, morphologically complex craters. Mask R-CNN achieved a precision of 91 %, a recall of 85 %, and an Intersection over Union (IoU) of 87 %, excelling in detecting intricate crater edges and identifying crater diameters with accurate geolocation information. However, it struggled to detect craters with less depth or degraded rims. Conversely, U-Net demonstrated superior recall (93 %) but moderate precision (85 %), making it efficient for broader crater localization tasks. U-Net excelled at identifying perfectly shaped craters but faced challenges in detecting larger and very small craters. Mask R-CNN identified previously uncatalogued craters, particularly those smaller than 1 km in diameter, while U-Net excelled at detecting a greater number of overlapping and nested craters, showcasing their complementary strengths. These findings underscore the potential of deep learning to enhance lunar research and future planetary exploration.