🛰️ Practical application tutorial session II: machine learning with survey images
https://github.com/csabiu/KAML-2025/blob/main/KAML_images.ipynb
🛰️ Autoencoder-Based Galaxy Image Search: A New Approach to Morphological Classification and Galaxy Research Methods
Suk Kim (CNU)
potential application of and Autoencoder model
- Search for similar images
Tool for Searching a similar Galaxy image based on a Convolutional Autoencoder using Simailarity - Morphological classification
Semi-supervised learning, Clustering - Denoising
Directions for Future Improvement
Key Challenges Identified:
• Data Imbalance: The current dataset is unbalanced across morphological types, especially for complex structures, making It harder to find similar images for underrepresented types.
• Latent Feature Bias: The latent features are sometimes biased towards global characteristics like color and overall shape, rather than detalled structures such as spiral arms or bars, leading to misclassifications. The model can also struggle to reconstruct delicate, faint structures.
Planned Solutions:
• Data Augmentation: Use Variational Autoencoders (VAEs) to generate diverse images, enriching the dataset for sparse morphological types and enhancing diversity.
• Improved Training Data: Incorporate residual images from galaxy-profile fitting Into the training data. This is intended to highlight and emphasize the detailed structural features crucial for accurate classification, helping to mitigate the bias towards color and shape.
• Feature Analysis: Further examine the correlation between latent features and photometric parameters to better understand and refine their relative significance in similarity calculations.
Key strategies to Activate Research
To effectively integrate machine learning into astronomy, the following initiatives are essential:
(1) Systematic Database Development
Purpose-Built Databases:
- Consistent Data Format
- Large-Scale Refined Data
- Labeling and Metadata for usability
(2) Data Accessibility and Collaborative Research - Promoting data sharing to expand collaborative research
- Leveraging open data platforms to facilitate widespread access
Collaborative Research Teams: Encouraging partnerships between machine learning experts and astronomers
