Enhancing lunar DEM data using super-resolution techniques and optimizing the faster R-CNN network for sub-kilometer crater detection

Abstract

Meteorite impact craters are the most prominent geomorphological features on the lunar surface, playing a crucial role in studying lunar geological characteristics, evolution, and the selection of landing sites for spacecraft. However, due to the resolution limitations of global lunar Digital Elevation Models (DEM), existing DEM-based automatic crater detection methods struggle to identify small craters effectively. This paper proposes an SR-Net method based on super-resolution techniques to address this challenge to enhance lunar DEM image features. We combine this method with an improved Faster R-CNN algorithm to efficiently detect small craters. We used the Robbins database for training and testing, and we evaluated the model's performance for detecting craters with diameters greater than 1 km. Experimental results showed that the improved model achieved a recall rate of 81.2 %, a precision of 76 %, and an F1 score of 78.5 % on the test set. In the 4× super-resolved DEM images, we detected some small craters, less than 1 km in diameter, in the Change-5 landing area that were not recorded in the Robbins database. After manual inspection by four experts, the average false positive rate(FPR) for small craters that were more prominent than 500 m in diameter was 10 %, while the false positive rate for craters more prominent than 700 m dropped to 5 %. The results demonstrate that the proposed method significantly improves the detection of small craters, particularly in low-resolution scenarios, providing a viable solution for small crater detection tasks.