Dušan Pavlović

The applications of machine/deep learning in astronomy are numerous, and in this tutorial session, we will tackle only one of them: the detection of meteors in the data from radio observations. Meteors are phenomena of light and ionization processes through which a small body goes and produces while interacting with the atmosphere. They can be observed via a wide range of techniques, and one of them is the so-called forward-scattering technique used in radio meteor astronomy. It works by reflecting radio waves emitted from the earth's surface onto the ionized trails of meteors in the atmosphere, and detecting the reflected signal (radio echo). As meteor echoes can be highly irregular in shape, and as there are also planes and aircrafts that reflect the radio signal during the observation period, too, the human eye is still the best instrument for the problem of meteor detection in the radio meteor data, but more and more people are trying to approach this problem from the machine learning and neural network perspective. The main idea behind this talk is to try to achieve a development of a technique for meteor detection in the radio spectrogram data, through the image segmentation approach using U-net CNN architecture in PyTorch, which will achieve human-like performance for the detection of meteor echoes on the (labeled) spectrograms provided for this purpose by the BRAMS (Belgian RAdio Meteor Stations) network. Also, several problems as bounding boxes aggregation, evaluation metrics and image transforms used for the purpose of this model will be tackled in the talk.