Impact craters are highly recognizable features on planetary surfaces. However, an AI system trained to identify craters from one set of images might not perform consistently when the surrounding terrain, image qualities, or observation context change. This issue, known as domain shift, causes models to work well on familiar data but less reliably on new environments. This patented method helps crater-detection AI adapt to such variations. Its feature-extraction network includes a low-level feature enhancement module and a circular-boundary enhancement module, both essential because crater recognition depends heavily on texture and circular boundary features. The approach uses progressive domain adaptation to create intermediate features between source and target domains, avoiding abrupt shifts. Detection results are then aligned at the instance level using similarity-based weighting. The goal isn’t just to detect more craters but to strengthen the recognition of consistent features in changing visual contexts. This has broader implications for AI, as models deployed outside controlled datasets often face new cameras, locations, lighting, and environments. The crater detection strategy exemplifies an adaptation approach: instead of assuming training and deployment data always match, it develops representations that adapt when they differ. In planetary exploration, this could improve automated image analysis across increasingly diverse terrains and datasets.
Finding craters when the landscape changes
Patent number: CN 116091905 B
Inventor(s): Z. Cai and S. Yang
Citation: Z. Cai and S. Yang, “一种域自适应撞击坑检测方法、系统、装置及存储介质 [Domain-adaptive impact-crater detection method, system, apparatus and storage medium],” China Patent CN 116091905 B, Apr. 21, 2026.
