How Do Convolutional Neural Networks (CNNs) Compare to Transformer-Based Models in Classifying Galaxy Morphologies in Large-Scale Sky Surveys?

The research evaluates the performance of Convolutional Neural Networks (CNNs) and Vision Transformer (ViT) models when used for galaxy morphology classification on Sloan Digital Sky Survey (SDSS) and Galaxy Zoo 2 (GZ2) survey data. The classification of galaxy morphology serves as a fundamental tool for studying cosmic structure and evolution but human analysis becomes impossible when dealing with millions of images. The research team creates a labeled dataset from GZ2 data before applying image preprocessing to galaxies and trains CNN and ViT models under identical parameters to evaluate their performance and convergence and their capacity to handle unbalanced data. The CNN model achieved superior results by reaching 43.33% accuracy compared to the ViT model which reached 39.61% on the same dataset. The results show CNNs perform better on small to medium-sized datasets because they excel at detecting local spatial patterns yet ViTs need extensive pretraining and large datasets to achieve their best global pattern recognition abilities. The research demonstrates that future astronomical machine learning needs to adopt transfer learning methods and develop hybrid CNN-Transformer models and implement data augmentation techniques to enhance galaxy classification accuracy.