Spin Parity of Spiral Galaxies II: A catalogue of 80k spiral galaxies using big data from the Subaru Hyper Suprime-Cam Survey and deep learning

TL;DR

This paper presents an automated deep learning approach to classify a large sample of spiral galaxies using high-resolution Subaru HSC survey data, significantly advancing galaxy morphology studies.

Contribution

It introduces a CNN-based classification method applied to deep, high-resolution images, enabling the identification of spiral structures at higher redshifts than previous surveys.

Findings

Classified nearly 77,000 spiral galaxies with 97.5% accuracy

Identified over 76,000 spiral galaxies at z>0.2, unseen in SDSS data

Demonstrated potential for classifying galaxy features like bars and mergers

Abstract

We report an automated morphological classification of galaxies into S-wise spirals, Z-wise spirals, and non-spirals using big image data taken from Subaru/Hyper Suprime-Cam (HSC) Survey and a convolutional neural network(CNN)-based deep learning technique. The HSC i-band images are about 25 times deeper than those from the Sloan Digital Sky Survey (SDSS) and have a two times higher spatial resolution, allowing us to identify substructures such as spiral arms and bars in galaxies at z>0.1. We train CNN classifiers by using HSC images of 1447 S-spirals, 1382 Z-spirals, and 51,650 non-spirals. As the number of images in each class is unbalanced, we augment the data of spiral galaxies by horizontal flipping, rotation, and rescaling of images to make the numbers of three classes similar. The trained CNN models correctly classify 97.5% of the validation data, which is not used for training. We apply the CNNs to HSC images of a half million galaxies with an i-band magnitude of i<20 over an area of 320 deg^2. 37,917 S-spirals and 38,718 Z-spirals are identified, indicating no significant difference between the numbers of two classes. Among a total of 76,635 spiral galaxies, 48,576 are located at z>0.2, where we are hardly able to identify spiral arms in the SDSS images. Our attempt demonstrates that a combination of the HSC big data and CNNs has a large potential to classify various types of morphology such as bars, mergers and strongly-lensed objects.