Abstract
With billions of galaxies in the observable universe, manual morphological classification has become an insurmountable task for astronomers. This paper explores the application of the VGG16 convolutional neural network (CNN) to automate this process using the Galaxy10 DECals dataset. By leveraging deep learning, we achieved an overall classification accuracy of 81%. Our findings suggest that while standard CNN architectures are highly effective for structured morphologies like elliptical and spiral galaxies, they face significant challenges with "disturbed" or irregular systems.
Introduction: The Challenge of Big Data Astronomy
Galaxy morphology—the study of a galaxy's shape and structure—provides critical insights into its evolutionary history and physical properties. Since the development of the Hubble Sequence, astronomers have relied on these visual cues to understand the universe. However, modern surveys now produce data at a scale that far exceeds human capacity for manual inspection or even large-scale crowdsourcing efforts like Galaxy Zoo. There is an urgent need for automated, scalable, and accurate classification systems.
Methods: Leveraging VGG16
This study utilizes the VGG16 architecture, a 16-layer convolutional neural network known for its simplicity and effectiveness in image recognition. The model consists of 13 convolutional layers using $3 \times 3$ filters and 3 dense layers.
To prepare the data, raw $256 \times 256$ pixel images were cropped to $224 \times 224$ pixels. This preprocessing step focuses the model on the centered galactic structures while reducing computational noise from the image borders. Feature map analysis was conducted to ensure the model was identifying relevant astronomical features, such as galactic cores and spiral arms, rather than background artifacts.
Dataset: Galaxy10 DECals
The model was trained and evaluated on the Galaxy10 DECals dataset, which contains 17,736 color images across 10 distinct morphological classes. These classes range from "Round Smooth" and "Cigar Shaped" to more complex "Merging" and "Disturbed" systems.
The dataset exhibits a varied distribution across these categories, which impacts model performance. For instance, classes with high internal symmetry are more easily identified than those with irregular structures.
Key Insights and Results
The VGG16 model reached an overall accuracy of 81% after 16 training epochs. Performance varied significantly by class:
- High Precision: The "Round Smooth" category achieved a precision of 0.94, indicating the model is highly reliable when identifying simple elliptical shapes.
- High Recall: The "Edge-on without Bulge" category reached a recall of 0.97, showing the model's strength in detecting specific disk-like orientations.
- The "Disturbed" Difficulty: The model struggled most with "Disturbed" galaxies, where incorrect classifications outnumbered correct ones. This suggests that the lack of consistent structural patterns in irregular galaxies makes them difficult for standard CNNs to generalize.
Conclusion and Future Work
The 81% accuracy achieved by VGG16 demonstrates that deep learning is a viable path for large-scale astronomical classification. However, the limitations observed in irregular "Disturbed" classes suggest that future work should explore more advanced architectures, such as Vision Transformers (ViT), or incorporate data augmentation techniques to better handle class imbalances and morphological complexity.
Summary updated