A self-organizing map-based solution to the automatic detection of meteor echoes in radio spectrograms

Abstract

This paper aims at introducing a novel approach to the automatic detection of meteor echoes in spectrograms of radio recordings. The proposed meteor detection solution uses an artificial neural network to analyze data extracted from spectrograms and identify the meteor echoes present within. The success of the neural network based solution greatly depends on the network's architecture and training process. Several tests were performed to find the optimal neural network, while its training process was done using a manually built data set to insure the presentation of sufficiently large set of data examples to the neural network. The final trained network is evaluated using a new set of spectrogram data and two distinct feature areas identification methods. The result obtained by this neural network were dount to provide a statistical significant count of the meteor echoes in the BRAMS spectrograms, with correct data classification rates of over 88%.