Deep Learning Applied to Lunar Crater Counting

Abstract:

Lunar crater counting is a powerful method for estimating lunar age. It has been carried out by human eyes, but this method has potential errors. Recently, Chirstopher(2019) and Silburt(2016) tried to count craters using deep-learning at Mars and Moon. The results are about up to 10% error, and most of the errors have occurred in the process of creating labeling data. We tried to detect lunar craters applying deep-learning with LROC images. We used the crater’s shapefiles and the lunar surface images with 100m/px provided by LROC data archive. The images were obtained by LROC/WAC. We proceeded in the following; 1) using LROC data and crater’s shapefile, make patch image and training dataset, validation dataset, and test label set. 2) YOLOv3 and RetinaNet, the object detection models are trained with training-dataset and validation dataset. 3) the model has carried out the evaluation with a test-dataset. We will try to apply to the images with under 5km and upper 20km craters, and then will apply to TC/Kaguya data. We are planning to apply to the images by ShadowCam, finally. Although the PSR images look different from the others, we expect that the model in this work will be able to used for PSR images taken by ShadowCam. This research was conducted by NRF (2018M1A3A3A02065832) support.