Deep Learning Assessment of galaxy morphology in S-PLUS DataRelease 1

TL;DR

This paper develops and evaluates deep learning methods for galaxy morphology classification using multi-band data from S-PLUS, demonstrating the importance of prior training and providing a new galaxy morphology catalogue.

Contribution

It introduces a novel multi-band morphometric fitting technique, compares CNN performance across different band combinations, and releases a new galaxy morphology catalogue for the S-PLUS survey.

Findings

CNN performance improves with 5 broad and 3 narrow bands over 3 bands

Pre-trained networks on ImageNet perform well with only 3 bands

The catalogue includes 3274 new galaxies not in GZ1 and classifications for ambiguous GZ1 galaxies

Abstract

The morphological diversity of galaxies is a relevant probe of galaxy evolution and cosmological structure formation, but the classification of galaxies in large sky surveys is becoming a significant challenge. We use data from the Stripe-82 area observed by the Southern Photometric Local Universe Survey (S-PLUS) in twelve optical bands, and present a catalogue of the morphologies of galaxies brighter than $r=17$ mag determined both using a novel multi-band morphometric fitting technique and Convolutional Neural Networks (CNNs) for computer vision. Using the CNNs we find that, compared to our baseline results with 3 bands, the performance increases when using 5 broad and 3 narrow bands, but is poorer when using the full $12$ band S-PLUS image set. However, the best result is still achieved with just 3 optical bands when using pre-trained network weights from an ImageNet data set. These results demonstrate the importance of using prior knowledge about neural network weights based on training in unrelated, extensive data sets, when available. Our catalogue contains 3274 galaxies in Stripe-82 that are not present in Galaxy Zoo 1 (GZ1), and we also provide our classifications for 4686 galaxies that were considered ambiguous in GZ1. Finally, we present a prospect of a novel way to take advantage of $12$ band information for morphological classification using morphometric features, and we release a model that has been pre-trained on several bands that could be adapted for classifications using data from other surveys. The morphological catalogues are publicly available.