Automatic Morphological Classification of Galaxies: Convolutional Autoencoder and Bagging-based Multiclustering Model

Chichun Zhou, Yizhou Gu, Zesen Lin, Guanwen Fang

Abstract:

In order to obtain morphological information of unlabeled galaxies, we present an unsupervised machine-learning
(UML) method for morphological classification of galaxies, which can be summarized as two aspects: (1) the
methodology of convolutional autoencoder (CAE) is used to reduce the dimensions and extract features from the
imaging data; (2) the bagging-based multiclustering model is proposed to obtain the classifications with high
confidence at the cost of rejecting the disputed sources that are inconsistently voted. We apply this method on the
sample of galaxies with H < 24.5 in CANDELS. Galaxies are clustered into 100 groups, each contains galaxies
with analogous characteristics. To explore the robustness of the morphological classifications, we merge 100
groups into five categories by visual verification, including spheroid, early-type disk, late-type disk, irregular, and
unclassifiable. After eliminating the unclassifiable category and the sources with inconsistent voting, the purity of
the remaining four subclasses are significantly improved. Massive galaxies (M* > 1010Me) are selected to
investigate the connection with other physical properties. The classification scheme separates galaxies well in the
U − V and V − J color space and Gini–M20 space. The gradual tendency of Sérsic indexes and effective radii is
shown from the spheroid subclass to the irregular subclass. It suggests that the combination of CAE and
multiclustering strategy is an effective method to cluster galaxies with similar features and can yield high-quality
morphological classifications. Our study demonstrates the feasibility of UML in morphological analysis that would
develop and serve the future observations made with China Space Station telescope.