Deep learning-based meteorite detection using YOLOv5 and faster R-CNN: dataset optimization and onboard drone inference
Aisha Al-Owais & Munya Al-Khalifa
Neural Computing and Applications, Volume 38, article number 325, Published: 21 April 2026
“To facilitate meteorite hunting and distinguish meteorites from terrestrial rocks, this study evaluates two object-detection architectures: YOLOv5 and Faster R-CNN. To date, no study has implemented a fully field-deployable system performing real-time meteorite detection using YOLOv5 on a single-board computer integrated with a smartphone camera mounted on a drone. This work addresses that gap by introducing a dual-device architecture in which the smartphone provides imaging while the Jetson Nano performs onboard inference, enabling autonomous detection.The objective is twofold: to demonstrate autonomous, drone-based meteorite detection and to compare YOLOv5 and Faster R-CNN to determine the best suited model for aerial real-time deployment. Images were collected from the Sharjah Academy for Astronomy, Space Sciences, and Technology (SAASST) collection and curated into five dataset variants to examine the effect of imaging conditions on performance. YOLOv5 was trained on all variants, whereas Faster R-CNN was trained on the most diverse dataset (Dataset 2).In offline inference, Models 2 and 3 achieved the strongest performance with a true positive rate of 1, and confidence scores ranging from 91% to 98% in real-time indoor testing, Model 2 detected 95% of the samples, while Model 6 (Faster R-CNN) detected all samples but with slower inference. Integrating offline and real-time results, Models 2 and 3 were selected for outdoor, drone-based testing using 20 samples, demonstrating the feasibility of the proposed autonomous detection framework. Although the testing sample size is limited, the system is presented as a proof of concept and will be validated on larger datasets in future work.”
