Query Refined and Spatially Enhanced Algorithm for Small- and Medium-Sized Crater Detection and Cataloging

Small- and medium-sized craters (SMCs) play a key role in the field of deep space exploration, whose detection and measurement face problems such as fringe blurring and low contrast due to their size. This study proposes a new deep learning-based SMC detection algorithm named small- and medium-sized crater detection Transformer (SMC-DETR), which improves crater detection performance in three aspects: query selection, deformable attention mechanism, and dense supervision strategy. First, a valid query selection algorithm adopted can dynamically pick more distinguishable embeddings to remove redundancy and enhance the efficiency of the decoder. Pseudo-labels produced by the nonmaximum suppression (NMS) algorithm work as guides for the valid query head to converge with the help of rank loss. Second, the spatial information is explicitly injected into the deformable attention weight, leading to a stronger robustness to perceive spatial locations and facilitate the detection of SMC. Finally, a densely supervised (DS) training strategy is implemented to increase the amount of detection output to be assigned, ensuring more intensified gradient for SMC-DETR to converge. The model proposed in this article achieves 14.3% average precision (AP) on the SMC dataset while reaching 36.9% AP on MDCD, which defeats other crater detection baselines. In addition, a digital elevation map (DEM) helps to filter false detection results and candidate craters are manually verified so that an elaborate catalog that contains 15 595 SMC with diameters around 10 m near the landing site of Chang’E-5 could be built with high accuracy and precision. Detailed topographic information is provided, and statistical analysis is conducted as well. The crater detection algorithm (CDA) model, SMC dataset, and completed crater catalog are accessible in https://github.com/BugBubbles/SMC-DETR