Deep Transfer Learning for Star Cluster Classification

Wei Wei, E. A. Huerta, Brad Whitmore, Janice Lee, Stephen HannonBulletin 2020

TL;DR

This paper demonstrates that deep learning, specifically transfer learning, can effectively classify star clusters in astronomical images. By leveraging pre-trained models, the authors achieve classification accuracies comparable to human experts, paving the way for automated, large-scale analysis of star cluster populations in upcoming astronomical surveys.

What Problem Does This Paper Address?

Classifying astronomical objects, like star clusters, has traditionally relied on human visual inspection. While effective for smaller datasets, this approach is not scalable for the massive amounts of data generated by modern astronomical surveys. Existing machine learning methods for image classification have shown great promise, but applying them to specialized astronomical tasks like star cluster morphology requires adapting them to smaller, domain-specific datasets. This paper tackles the challenge of accurately classifying star clusters into four morphological categories using deep learning, aiming to bridge the gap between human expertise and the demands of big data astronomy.

Methods

The core of this work lies in deep transfer learning. The authors take pre-trained neural network architectures, specifically ResNet18 and VGG19-BN, which have been trained on the massive ImageNet dataset for general object recognition. They then "fine-tune" these models using a smaller, curated dataset of star cluster images. This process leverages the general feature extraction capabilities learned from ImageNet and adapts them to the specific task of star cluster classification.

To handle the fact that star cluster images have fewer input channels (5 filters) than the typical 3 channels of ImageNet data, they concatenate two copies of the neural network architecture, effectively creating a model with 6 input channels. They also explore the impact of different image sizes (25x25, 50x50, 100x100 pixels) for training and the importance of individual imaging filters.

Datasets

The primary datasets used are:

  • LEGUS (Legacy ExtraGalactic UV Survey) dataset: This dataset contains human-classified star cluster candidates from various nearby galaxies. The authors utilize two versions:
  • Classifications primarily performed by a single expert (BCW).
  • Consensus classifications from three LEGUS team members.
  • PHANGS-HST (Physics at High Angular Resolution in Nearby Galaxies) Survey data: This dataset, specifically observations of the galaxy NGC 1559, is used as an independent test set. NGC 1559 was not included in the training data, allowing the authors to assess the models' generalization capabilities.

Key Insights and Ideas

  • Transfer learning is powerful for specialized astronomical tasks: The success of models pre-trained on general image recognition (ImageNet) for a specific astronomical problem (star cluster classification) highlights the broad applicability of transfer learning.
  • Human classification consistency is a key benchmark: The authors meticulously compare their model's performance against human classifications, establishing a baseline for "accuracy" based on inter-expert agreement. This is crucial for evaluating the success of any automated method.
  • Model robustness to training data variations: Surprisingly, the models' performance was not significantly affected by whether the training data was classified by a single expert or a consensus of three, nor by the specific galaxies included in the training set. This suggests that deep learning models can generalize well even with variations in the training data's origin and composition.
  • The V-band filter is critical for classification: Removing the V-band filter from the input data led to the most significant drop in classification accuracy, indicating its importance for distinguishing star cluster morphologies, mirroring human visual inspection strategies.

Key Results

  • Competitive Classification Accuracy: The models achieved classification accuracies for NGC 1559 (a galaxy not in the training set) of approximately 70% for Class 1 (symmetric compact clusters), 40-50% for Class 2 (asymmetric compact clusters), 40-50% for Class 3 (compact associations), and 50-70% for Class 4 (non-clusters). These accuracies are comparable to previously published human and automated methods.
  • Model Robustness: The choice of neural network architecture (ResNet18 vs. VGG19-BN) and the origin of the training data (single expert vs. consensus) had minimal impact on overall classification accuracy.
  • Image Size Insensitivity: Varying the size of the input images (from 25x25 to 100x100 pixels) did not significantly alter the classification performance, suggesting that the models are adept at extracting relevant features from smaller image patches.

Limitations and Future Work

  • Dataset Size: While transfer learning mitigates the need for massive datasets, the authors acknowledge that the current training sets are still relatively small. They suggest that training with larger, combined datasets would further improve performance.
  • Standardized Datasets: A key limitation identified is the lack of a universally agreed-upon, standardized dataset of human-labeled star cluster classifications. The authors emphasize the need to create such a dataset, agreed upon by a broad range of experts, to further enhance the models.
  • Future Work: The authors propose future work to include training with larger datasets, developing standardized classification protocols, and potentially conducting "classification challenges" to foster expert consensus. They aim to use these improved models to automate the classification of star clusters in the ongoing PHANGS-HST survey and future astronomical endeavors.

Summary updated

BibTeX
@article{wei2020deep,
  author = {Wei Wei and E. A. Huerta and Brad Whitmore and Janice Lee and Stephen Hannon},
  title = {{Deep Transfer Learning for Star Cluster Classification}},
  journal = {Bulletin},
  year = {2020},
}
Figure