Small-Body Segmentation Based on Morphological Features with a U-Shaped Network Architecture

Small bodies such as asteroids and comets display great variability in terms of surface morphological features. These are often unknown beforehand but can be employed for hazard avoidance during landing, autonomous planning of scientific observations, and navigation purposes. Algorithms performing these tasks are often data driven, which means they require realistic, sizeable, and annotated datasets, which in turn may rely heavily on human intervention. This work develops a methodology to generate synthetic, automatically labeled datasets that are used in conjunction with real, manually labeled ones to train deep-learning architectures in the task of semantic segmentation. This functionality is achieved by designing U-shaped network architectures trained with different strategies. These show good generalization capabilities, implement uncertainty quantification estimates, and can be hybridized to exploit qualities from multiple networks.