Morphological classification of galaxies with deep learning: comparing 3-way and 4-way CNNs

TL;DR

This paper develops and compares deep learning CNN models for galaxy morphological classification, achieving high accuracy in 3-way and 4-way schemes, and provides insights into classification challenges and physical correlations.

Contribution

Introduces a new CNN architecture that outperforms existing models in galaxy morphology classification and analyzes the physical significance of misclassifications.

Findings

New CNN architecture achieves 83% and 81% accuracy in 3-way and 4-way classification.

Ellipticals and spirals are distinguished with over 98% accuracy.

Misclassifications show physical trends, e.g., massive lenticulars often misclassified as ellipticals.

Abstract

Classifying the morphologies of galaxies is an important step in understanding their physical properties and evolutionary histories. The advent of large-scale surveys has hastened the need to develop techniques for automated morphological classification. We train and test several convolutional neural network architectures to classify the morphologies of galaxies in both a 3-class (elliptical, lenticular, spiral) and 4-class (+irregular/miscellaneous) schema with a dataset of 14034 visually-classified SDSS images. We develop a new CNN architecture that outperforms existing models in both 3 and 4-way classification, with overall classification accuracies of 83% and 81% respectively. We also compare the accuracies of 2-way / binary classifications between all four classes, showing that ellipticals and spirals are most easily distinguished (>98% accuracy), while spirals and irregulars are hardest to differentiate (78% accuracy). Through an analysis of all classified samples, we find tentative evidence that misclassifications are physically meaningful, with lenticulars misclassified as ellipticals tending to be more massive, among other trends. We further combine our binary CNN classifiers to perform a hierarchical classification of samples, obtaining comparable accuracies (81%) to the direct 3-class CNN, but considerably worse accuracies in the 4-way case (65%). As an additional verification, we apply our networks to a small sample of Galaxy Zoo images, obtaining accuracies of 92%, 82% and 77% for the binary, 3-way and 4-way classifications respectively.