Small lunar crater detection using a few-shot object detection approach with Vision Transformer

Craters are among the most prominent features on the lunar surface, and their importance to the European Space Agency’s (ESA) planned lunar landings highlights the need for further investigation into the detection of small lunar craters. The rugged lunar terrain and varying lighting conditions make this task particularly challenging, and it is a major focus of current research. The application of deep learning methods is the most common approach in this field. In general, deep learning approaches rely on a sufficiently well-labeled dataset to achieve optimal performance. However, there is currently no common benchmark dataset specifically designed for small lunar craters, which presents a significant challenge for advancing research in this area. We present a novel method for detecting small lunar craters by utilizing and adapting the few-shot object detection capabilities of the OWLv2 model, based on a Vision Transformer (ViT), eliminating the need for a large labeled dataset. Our method is tested on high-resolution Lunar Reconnaissance Orbiter Camera (LROC) Calibrated Data Record (CDR) images, which offer resolutions as fine as 0.5 m/pixel, ensuring detailed analysis and evaluation. We do not fine-tune the OWLv2 model but instead modify the similarity score calculation method. We achieve promising visual results, along with recall and precision values of 0.83 and 0.64, respectively, tested on a sample from the labeled dataset of the IMPACT project.