Cross-attention induced multilayer domain adaptation network for extraction of sub-kilometer craters from HiRIC images

Deep learning-based crater detection algorithms have made remarkable progress in the fields of space science and engineering, establishing themselves as the leading solutions for Mars geology exploration. However, achieving outstanding performance with deep networks requires extensive annotated data, resulting in timeconsuming and laborintensive labeling processes. In response to this challenge, we propose a novel approach, the Cross-Attention-Induced Multilayer Domain Adaptation Network (CAMDA-Net), specifically designed for crater detection. CAMDA-Net leverages domain adaptation techniques to transfer valuable crater knowledge from the Moon to Mars, effectively bypassing the need for manual labeling. In our network, we integrate CenterNet*, an anchor-free detector, to expedite and enhance the accuracy of crater detection. Furthermore, we implement crater feature alignment through adversarial training at both the image and instance levels. The dual cross-attention mechanism at the image level directs the network's focus towards transferable crater features while suppressing noise effects. At the instance level, the foreground and background alignment model streamlines feature alignment and eliminates negative proposals. To assess the efficacy of our proposed method, we build the interplanetary domain-adaptive crater dataset (IDCD), which encompasses three subsets: the LRO WAC crater dataset, THEMIS crater dataset, and HiRIC crater dataset. This diverse dataset incorporates various shapes, lighting conditions, scales, and resolutions, making it an ideal resource for conducting transfer learning studies. We have made the IDCD publicly accessible on Github at https://github.com/PlanetaryScience3510. The experimental results on the IDCD subset from the LRO WAC to HiRIC demonstrate that our CAMDA-Net achieves precision and recall values of 0.77 and 0.79, respectively, slightly surpassing classic domain-adaptive detectors, DAF and SCDA. With encouraging performance, CAMDA-Net shows great promise for advancing crater detection and analysis on Mars, presenting exciting possibilities for future research and applications.

Automated summary

This article introduces a deep learning-based crater detection algorithm, the Cross-Attention-Induced Multilayer Domain Adaptation Network (CAMDA-Net), which transfers valuable crater knowledge from the Moon to Mars without the need for manual labeling. CAMDA-Net aligns crater features at both the image and instance levels and demonstrates excellent performance in crater detection on Mars.