Classification of Spiral Galaxies by Spiral Arm Number using Convolutional Neural Network

TL;DR

This study developed CNN classifiers based on EfficientNet architectures to categorize spiral galaxies by arm number, achieving high accuracy and providing insights into galaxy structure and evolution.

Contribution

The paper introduces a CNN-based classification method using EfficientNet variants for spiral arm number categorization, with improved accuracy through class merging and interpretability analysis.

Findings

High accuracy (>0.8) in classifying spiral arm numbers except for 4 arms.

Merging classes with more than 4 arms improved 4-arm classification accuracy.

GradCAM++ and SmoothGrad effectively highlighted galaxy structures and spiral arms.

Abstract

The structural information of spiral galaxies such as the spiral arm number, offer valuable insights into their formation processes and physical roles in galaxy evolution. We developed classifiers based on CNNs using variants of the EfficientNet architecture with different transfer learning techniques and pre-trained weights to categorise spiral galaxies by their number of spiral arms. A dataset from GZ2, comprising 11718 images filtered based on appropriate criteria is used for training and evaluation. Both the EfficientNetV2M model fine-tuned on ImageNet and the EfficientNetB0 model with Zoobot pre-trained weights achieved high accuracy on the down-sampled dataset, with most performance metrics exceeding 0.8 across all classes, except for galaxies with 4 arms due to the limited number of samples in this category. Merging higher-arm-number classes (more than 4 arms) improved the V2M model's accuracy significantly for 4-arm galaxies, as this approach allowed the model to focus on more distinct features with a more balanced class distribution. GradCAM++ and SmoothGrad highlight the networks' effectiveness in classifying galaxies, through the distinction of the galaxy structures and the extraction of the spiral arms, with the V2M model showing better capabilities in both tasks. Lower-arm galaxies tend to be misclassified as "can't tell" when their spiral arms are not clearly visible, while higher-arm galaxies tend to be misclassified as having fewer arms when their features are only partially detected. The study also found that galaxies with 3 arms tend to have lower stellar masses, and this tendency is reduced in the model predictions. The models' mispredictions between 2-arm and 1/3-arm are likely resulting from external interference and dynamic nature of spiral arms. The V2M model prediction also shows a slight tendency towards higher stellar mass in high-arm galaxies.