Deep Learning on Jovian Decametric Emissions

Marques et al. (2017) built a database covering 26 years of daily observation of Jupiter from the Nançay Decameter Array (NDA). We use this database to train semantic segmentation algorithms to automatically identify and localise Jupiter emissions (induced or not by Io and of A,B,C or D types) in the observations. We explore different segmentation models (Unet, FPN, SwinUnet) and compare their performance on our task. We present the results through a web interface, which is now operational on any daily NDA observation of Jupiter. For each observation in the catalog, we represent the classification from Marques et al. (2017) and our algorithm output. We also provide indicators to assess the performance of the algorithm, and visualizations of the data from ExPRES simulation and from the Jupiter Probability Tool. This approach is directly applicable to any other ground-based radio observation of Jupiter, such as those of NenuFAR.