Automated crater detection on Mars using deep learning

Impact crater cataloging is an important tool in the study of the geological history of planetary bodies in the Solar System, including dating of surface features and geologic mapping of surface processes. Catalogs of impact craters have been created by a diverse set of methods over many decades, including using visible or near infra-red imagery and digital terrain models. I present an automated system for crater detection and cataloging using a digital terrain model (DTM) of Mars - In the algorithm craters are first identified as rings or disks on samples of the DTM image using a convolutional neural network with a UNET architecture, and the location and size of the features are determined using a circle matching algorithm. I describe the crater detection algorithm (CDA) and compare its performance relative to an existing crater dataset. I further examine craters missed by the CDA as well as potential new craters found by the algorithm. I show that the CDA can find three-quarters of the resolvable craters in the Mars DTMs, with a median difference of 5-10% in crater diameter compared to an existing database. A version of this CDA has been used to process DTM data from the Moon and Mercury (Silburt et al., 2019). The source code for the complete CDA is available at https://github.com/silburt/DeepMoon, and Martian crater datasets generated using this CDA are available at https://doi.org/10.5683/SP2/MDKPC8.