Improving Galaxy Classification Accuracy Using Multi-Task Learning and Transfer Learning

School of Biology and Environmental Science, University College Dublin Belfield, Dublin 4, D04 V1W8, Ireland

^* Corresponding author: lei.yi@ucdconnect.ie

Abstract

In large-scale astronomical surveys, accurate galaxy classification is essential for advancing our understanding of the universe. This study systematically compares and evaluates the performance of three deep learning models for galaxy image classification. This paper uses the Galaxy10 DECals dataset and trained three models: a convolutional neural network (CNN), an autoencoder–classifier joint model and a transfer learning based ResNet50.For three models, this study employed multiple performance-evaluation methods, such as Kullback–Leibler divergence and accuracy, to assess their performance. The results show that adding multi- task learning and transfer learning gradually improves the accuracy of galaxy classification. In particular, the fine-tuned ResNet-50 achieved the highest classification accuracy by leveraging ImageNet pre-trained features via transfer learning. This result underscores the efficacy of transfer learning in bridging the domain gap between natural images and astronomical data. It also provides a robust baseline and practical methodology for employing advanced deep learning models in astrophysical image analysis.