Meteor detection and localization using YOLOv3 and YOLOv4

Meteors in the United Arab Emirates are observed daily through the U.A.E. Meteor Monitoring Network (UAEMMN). As of September 2022, more than 40,000 meteors have been observed. However, the high sensitivity of the network also captures non-meteor objects such as airplanes, birds, insects, and space debris appearing in the atmosphere. To accurately identify and label meteors, this study employs object detection algorithms to reduce data and accurately detect meteor and non-meteor objects. The YOLOv3 and YOLOv4 object detection algorithms, utilizing convolutional neural networks, were utilized in this research. The models were trained on both an imbalanced and a balanced dataset that consisted of thousands of images. The imbalanced YOLOv4 model yielded the highest recall score of 98.5% followed by the imbalanced YOLOv3 model with a recall score of 98%. The highest accuracy result was also achieved by the imbalanced YOLOv4 model, with a score of 90%. Overall, all the four models were successful at labeling meteors with a confidence more than 95%. The proposed study represents a significant contribution to the field of meteor-related image analysis using low-cost cameras and machine learning. It also holds promising implications for further research and development in this area.