A Transfer-Learning Pipeline for Specular Meteor Detection

Single-station specular meteor radars (SMR) provide long-term, continuous measurements of mesospheric dynamics, but their heuristic echo filtering methods and fixed detection thresholds limit their performance ceiling. A substantial population of detected meteors is either misclassified or excluded, leading to systematic undercounting and biased sampling of the true meteor population. In this work, we present a transfer-learning-based detection framework designed to improve meteor counts from single-station SMR observations by enhancing the identification of echoes of various morphologies. The model operates on frequency-spectrum radar data and is trained using a combination of simulated underdense meteor echoes and manually curated detections spanning a wide range of SNR, durations, and echo morphologies. Performance is evaluated against conventional filtering methods using both controlled test sets and extended observational periods. The proposed approach increases total meteor detections by a statistically significant margin, with the largest gains occurring near the radar’s sensitivity limit, while maintaining a low false-positive rate. We demonstrate that the additional detections exhibit physically consistent temporal, spatial, and amplitude characteristics, indicating that they represent genuine meteor echoes rather than clutter or noise. These results show that transfer learning is a powerful tool that can recover previously uncounted meteors in legacy and ongoing datasets, improving meteor count statistics without changes to radar hardware or operational parameters.