YOLO-ET: A Machine Learning model for detecting, localising and classifying anthropogenic contaminants and extraterrestrial microparticles optimised for mobile processing systems

Abstract

Imminent robotic and human activities on the Moon and other planetary bodies would benefit from advanced in situ Computer Vision and Machine Learning capabilities to identify and quantify microparticle terrestrial contaminants, lunar regolith disturbances, the flux of interplanetary dust particles, possible interstellar dust, $\beta$-meteoroids, and secondary impact ejecta. The YOLO-ET (ExtraTerrestrial) algorithm, an innovation in this field, fine-tunes Tiny-YOLO to specifically address these challenges. Designed for coreML model transference to mobile devices, the algorithm facilitates edge computing in space environment conditions. YOLO-ET is deployable as an app on an iPhone with LabCam® optical enhancement, ready for space application ruggedisation. Training on images from the Tanpopo aerogel panels returned from Japan’s Kibo module of the International Space Station, YOLO-ET demonstrates a 90% detection rate for surface contaminant microparticles on the aerogels, and shows promising early results for detection of both microparticle contaminants on the Moon and for evaluating asteroid return samples. YOLO-ET’s application to identifying spacecraft-derived microparticles in lunar regolith simulant samples and SEM images of asteroid Ryugu samples returned by Hayabusa2 and curated by JAXA’s Institute of Space and Astronautical Sciences indicate strong model performance and transfer learning capabilities for future extraterrestrial applications.