Authors: Sascha Carona, Klaas Dijkstra, Christopher Ecknerd, Luc Hendriksa, Guðlaugur Jóhannessonf, Boris Panes, Roberto Ruiz de Austrii, Gabrijela Zaharijasd.
At GeV energies, the sky is dominated by the interstellar emission from the Galaxy. With limited statistics and spatial resolution, accurate separation of point sources is therefore challenging. Here we present the first application of deep learning based algorithms to automatically detect and classify point sources from gamma-ray data. To detect point sources we utilise U-shaped convolutional networks for image segmentation and k-means for source clustering and localisation. We also explore the Centroid-Net algorithm, which is designed to find and count objects. Using two algorithms allows for a cross check of the results, while combination of their results can be used to improve performance. The training data is based on 9.5 years of Fermi-LAT exposure and we use source properties of active galactic nuclei (AGNs) and pulsars (PSRs) from the fourth Fermi-LAT source catalog (4FGL) in addition to several models of background interstellar emission. We compare our algorithms qualitatively with traditional methods and find them to have similar detection thresholds. We also demonstrate the robustness of our source localization algorithms to modifications in the interstellar emission models, which presents a clear advantage over traditional methods. The results of the localization algorithm are fed into a classification neural network that is trained to separate the three general source classes (AGNs, PSRs, and FAKE sources). The network is able to discriminate between these three classes with global accuracy of ∼70%, as long as balanced data sets are used in classification training. We publish our training data sets and analysis scripts and invite the community to the data challenge: How can we best locate and classify gamma-ray point sources?
Preprint available at: https://arxiv.org/pdf/2103.11068.pdf