Abstract
ABSTRACT Nowadays, extensive data are collected in an automated regime. Combining this, with the increase in accessible computational power, led to large-scale implementations of machine learning (ML). This is also the case of meteor science, where object detection often requires tracking of a moving light source between frames, and the number of false positives can be up to an order of magnitude higher than true meteoric phenomena. While spatiotemporal coincidence of events recorded by close, multiple cameras can eliminate most of the false positives, single-station detections in some camera networks are currently discarded. In this paper, we explore a set of ML models aiming to find an optimal method for re-analysis of this single-station observations, in order to identify and extract real meteors. A set of 15 ML models were trained on features extracted from the meteor movement. Upon testing, we found a top accuracy score of 98,2 per cent, and a recall (i.e. percentage of meteors correctly classified) score of 96 per cent for the best performing models. When combined with the spatiotemporal coincidence of the detection, the recall increases to 99.92 per cent. These 15 ML techniques were selected according to their ability classify tabular data, hence the bundle can be applied to other studies. The same goes for the computed features, which are independent on the camera configuration, thus, the process can be scaled and applied to other networks. These methods are to be implemented to re-analyze the events recorded by the larger, FRIPON network.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have