Celestial Image Classification using Ensemble Learning and Vision Transformers

Title

Celestial Image Classification using Ensemble Learning and Vision Transformers

Creator

Varghese, Nisha; Justin, Jesvin K.; Kezya Steffyn, S.

Description

Astronomical image classification plays a crucial role in understanding the universe, but deep learning models often stumble when faced with scarce labeled data. In our work, we address this gap in two key ways: first, by building a richly varied dataset from just 600 Hubble Space Telescope images and, through targeted augmentation, expanding it to 4,500 distinct training examples; and second, by introducing a hybrid learning strategy that marries transformer-driven feature extraction with gradient-boosted decision trees. We used benchmark standard convolutional architectures (ResNet-50, DenseNet-121) alongside the Data-Efficient Image Transformer (DeiT) and two novel hybrids-DeiT-RF and DeiT-XGBoost (DXg). In DXg, DeiT captures complex spatial patterns, an adaptive dimensionality reduction layer hones in on the most informative features, and XGBoost delivers the final classification. This fusion not only boosts accuracy across nebulae, galaxies, and star clusters but also enhances interpretability by revealing which transformer-derived features most influence the model's decisions. 2025 IEEE.

Source

2025 IEEE International Conference on Electronics, Computing and Communication Technologies, CONECCT 2025;

Date

01-01-2025

Publisher

Institute of Electrical and Electronics Engineers Inc.

Subject

Astronomical Image Classification; Convolutional Neural Networks (CNNs); Data Augmentation; Deep Learning; Dimensionality Reduction; Ensemble Learning; Hybrid Models; Transformers; XGBoost

Coverage

Varghese N., Christ University, Department of Computer Science, Karnataka, Bangalore, 560029, India; Justin J.K., Christ University, Department of Computer Science, Karnataka, Bangalore, 560029, India; Kezya Steffyn S., Christ University, Department of Computer Science, Karnataka, Bangalore, 560029, India

Rights

Restricted Access; Hardcopy may be available in the library

Relation

ISBN: 979-833150236-2;

Format

online

Language

English

Type

Conference paper

Identifier

https://doi.org/10.1109/CONECCT65861.2025.11306644

https://www.scopus.com/pages/publications/105032701371?origin=resultslist