Deep Learning Assessment of galaxy morphology in S-PLUS Data Release 1

Published in Monthly Notices of the Royal Astronomical Society, 2021

Abstract: The morphological classification of galaxies is a relevant probe for galaxy evolution and unveils its connection with cosmological structure formation. To this scope, it is fundamen- tal to recover galaxy morphologies over large areas of the sky. In this paper, we present a morphological catalogue for galaxies in the Stripe-82 area, observed with S-PLUS, till a mag- nitude limit of r ≤ 17, using the state-of-the-art of Convolutional Neural Networks (CNNs) for computer vision. This analysis will then be extended to the whole S-PLUS survey data, covering ‘ 9300 deg 2 of the celestial sphere in twelve optical bands. We find that the network’s performance increases with 5 broad bands and additional 3 narrow bands compared to our baseline with 3 bands. However, it does lose performance when using the full 12 band image information. Nevertheless, the best result is achieved with 3 bands, when using pre-trained network weights in an ImageNet dataset. These results en- hance the importance of previous knowledge in the neural network weights based on training in non related extensive datasets. Thus, we release a model pre-trained in several bands that could be adapted to other surveys. We develop a catalogue of 3274 galaxies in Stripe-82 that are not presented in Galaxy Zoo 1 (GZ1). We also add classification to 4686 galaxies consid- ered ambiguous in GZ1 dataset. Finally, we present a prospect of a novel way to take advan- tage of 12 bands information for morphological classification using multiband morphometric features. The morphological catalogues are publicly available.

Authors: C. R. Bom, A. Cortesi, G. Lucatelli, L. O. Dias, P. Schubert, G.B. Oliveira Schwarz, N. M. Cardoso, E. V. R. Lima, C. Mendes de Oliveira, L. Sodre Jr., A.V. Smith Castelli, F. Ferrari, G. Damke, R. Overzier, A. Kanaan, T. Ribeiro, W. Schoenell

Paper (PDF): arXiv:2104.00018

Supplementary Material & Data Outputs: Materials and data related with this paper are available at and