Abstract

Poster - Splinter General   (Kuppelsaal / virtual plenum)

Detection of Low-Surface-Brightness Galaxies Using a Compact Convolutional Neural Network

Henri Cecatka, Günther Heemann
Astronomical Institute at the Ruhr-University Bochum

The detection of low-surface-brightness galaxies (LSBGs) in astronomical surveys is challenged by their faint signals and poor signal-to-noise ratios. While new deep-sky surveys like the Legacy Survey of Space and Time and Euclid promise increased discoveries of LSBGs, the vast volume of available data poses a substantial problem for traditional source detection methods. We demonstrate that a compact convolutional neural network can achieve high‐precision LSBG detection. We train our model on Hyper Suprime-Cam Subaru Strategic Program (HSC) images using the LSBG catalogue by Greco et al. (2018) and the SuperBoRG catalogue by Morishita (2021) to generate accurate labels of confirmed faint sources. This dataset yields robust feature representation despite the small number of data points and the network’s small footprint. On the test dataset, the network achieves a precision of 97.46 % and a recall of 89.84 %, while both the false positive rate of 0.21 % and the false negative rate of 10.16 % are reasonably small. We further apply the trained network to HSC tracts, using a sliding-window algorithm that produces probability maps, to create a catalogue of LSBG candidates.