Unsupervised Galaxy Morphological Visual Representation with Deep Contrastive Learning

TL;DR

This paper introduces a contrastive learning approach using vision transformers and convolutional networks to learn galaxy morphological representations from unlabeled data, achieving high accuracy across multiple datasets.

Contribution

It presents a novel unsupervised contrastive learning method combining vision transformers and CNNs for galaxy morphology classification, reducing reliance on labeled data.

Findings

Achieved over 94% accuracy on Galaxy Zoo 2 and SDSS-DR17 datasets.

Demonstrated strong transfer and generalization ability across datasets.

Code and pretrained models are publicly available for easy adaptation.

Abstract

Galaxy morphology reflects structural properties which contribute to understand the formation and evolution of galaxies. Deep convolutional networks have proven to be very successful in learning hidden features that allow for unprecedented performance on galaxy morphological classification. Such networks mostly follow the supervised learning paradigm which requires sufficient labelled data for training. However, it is an expensive and complicated process of labeling for million galaxies, particularly for the forthcoming survey projects. In this paper, we present an approach based on contrastive learning with aim for learning galaxy morphological visual representation using only unlabeled data. Considering the properties of low semantic information and contour dominated of galaxy image, the feature extraction layer of the proposed method incorporates vision transformers and convolutional network to provide rich semantic representation via the fusion of the multi-hierarchy features. We train and test our method on 3 classifications of datasets from Galaxy Zoo 2 and SDSS-DR17, and 4 classifications from Galaxy Zoo DECaLS. The testing accuracy achieves 94.7%, 96.5% and 89.9% respectively. The experiment of cross validation demonstrates our model possesses transfer and generalization ability when applied to the new datasets. The code that reveals our proposed method and pretrained models are publicly available and can be easily adapted to new surveys.