Enhanced Galaxy Morphology Classification with Synthetic Data Augmentation and Vision Transformers
Year: 2026
Type: conference-paper
Abstract: Modern astronomical surveys like LSST (The legacy survey of space and time) and DES (dark energy survey) capture thousands of galaxy images every night which is creating a growing need for automated systems that can reliably identify galaxy structures at scale. Galaxy morphology is a key indicator of a galaxy's formation history, dynamical state and evolutionary processes thus making morphological classification essential for studying how galaxies form and change over time. However, manual labeling is time consuming and subjective. Furthermore, automated methods often struggle when it comes to rare or visually complex galaxy types that are underrepresented in training data. Using the Galaxy10 DECaLS dataset, which contains ten visually distinct galaxy classes, we generate additional training samples for minority classes using a deep convolutional generative adversarial network (DCGAN) to reduce the impact of class imbalance. These augmented datasets are then used to train a microsoft/swin-tiny-patch4-window7-224 image classification model on galaxy images. The resulting system achieves an overall classification accuracy of 89% across the ten morphological categories. Along with the increased performance by the proposed methodology, this study provides a comparative framework showcasing the other important advantages synthetic data augmentation yields and the improvements that can be achieved when paired with a state of the art vision transformer model.
Institution: PES University
Language: English
Subfield: Astronomy and Astrophysics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7213535005