Topographic knowledge-aware network for automatic small-scale impact crater detection from lunar digital elevation models

Impact craters represent the most prevalent and prominent topographical features on the surfaces of planets. They provide crucial insights into the internal and surface-level geological activities of planets but are difficult to identify from digital topographic data due to heterogeneous planetary surfaces and lack of distinguishing features. Previous studies, which implemented convolutional neural networks to automatically detect the planetary impact craters, focus on designing the state-of-the-art architectures to achieve optimal detection performance, without taking topographic knowledge into consideration. Therefore, we construct a topographic knowledge-aware network aimed at the automatic detection of small-scale impact craters utilizing lunar digital elevation models. In our work, we investigate the transfer-learning performances of one detector and conduct experiments on transferability from the perspectives of the geographic or topographic differences, which shows the knowledge transfer from source domain to target domain works well. Moreover, we calculate some quantitative metrics to describe the complexity of terrains and further find that the selected representative samples provide more valuable source of information for the generalization of one detector. For evaluative purposes, both qualitative and quantitative, each detector is comprehensively trained and then utilized for the detection of impact craters, drawing from the integrated digital elevation model of the Lunar Reconnaissance Orbiter (LRO) and Kaguya that extends over latitudes of ± 60° and the complete longitudinal spectrum, as well as its detection outcomes are benchmarked with the pre-existing crater catalog LU1319373. Experimental results that the topographic knowledge-aware detection outperforms traditional detection, with average differences of approximately 5 % in mAP and approximately 4 % in F1-score. The benefits of topographic knowledge-aware progressive detector training solution primarily derive from discrepancies in visual characteristics, heterogeneous lunar surfaces around the impact craters for generalization purposes. Our findings would be significantly applied to the discovery of scientific insights in both existing and new planetary datasets through the use of machine learning.