Artificial-to-Real Domain Adaptation for Lunar Landscape Semantic Segmentation

From Apollo 11 to India’s Chandrayaan-3, humans have put a vast amount of effort into exploring the moon. While analysis on the lunar surface provides valuable information for humans, costly and risk-taking problems of lunar exploration resulted in such scant amount of data accessible to the lunar landscape to this day. Nonetheless, study on lunar landscape remains pivotal for probing resources, detecting hazards, and studies on moon evolution. In addition, analysis on the lunar surface is the ground for future development on the moon as a potential site of resources, identifying safe landing sites, and further civilization. In this study, I propose a machine learning-based lunar landscape image semantic segmentation system. Given lunar landscape images, the proposed method outputs semantic segmentation maps that separate different types of objects such as the ground, sky, and rocks. These segmented objects potentially provide valuable insights for guiding autonomous investigation rovers. The proposed method is trained on a synthetic lunar landscape dataset and evaluated on both synthetic and real lunar landscape samples. Through comprehensive experiments, it is demonstrated that the proposed method exhibits domain adaptation capabilities, achieving state-of-the-art performance on real lunar landscape images.