Advanced Meteorite Identification through YOLO Object Detection Algorithms

Abstract

This study addresses the challenge of visually identifying meteorites from terrestrial rocks, traditionally a task for experts followed by chemical analysis. We propose a transformative approach using computer vision and machine learning, employing YOLO (You Only Look Once) object detection algorithms (versions 5, 6, 7, and 8), to overcome the bottleneck in expert availability for instantaneous classification. Leveraging a curated selection from the Sharjah Academy for Astronomy, Space Sciences, and Technology (SAASST) unique meteorite collection, we aim to differentiate meteorites from terrestrial rocks based on their surface features and characteristics. The collection comprises a diverse assemblage of approximately 8,000 objects such as iron meteorites, Martian meteorites, tektites, fulgurites, and more. Our methodology includes a comparative analysis of YOLO versions, focusing on precision, recall, and F1 scores to assess each algorithm's adaptability to the unique features of meteoritic material. Preliminary results indicate YOLOv5 as the most efficient compared to its previous versions, achieving a maximum mAP of 0.995 and correctly classifying 93% of test samples. This study aims to determine the optimal YOLO version for enhancing the accuracy and efficiency of meteorite classification. In addition, the selected optimal model will be deployed on a Jetson Nano processor aboard a drone, significantly enhancing onsite meteorite detection capabilities.