Gaussian Process Classification for Galaxy Blend Identification in LSST

J. Buchanan, M. Schneider, R. Armstrong, Amanda Muyskens, Benjamin W. Priest, Ryan DanaThe Astrophysical Journal 2022 · DOI

TL;DR

This paper introduces a Gaussian Process (GP) classifier for identifying blended galaxies in LSST images, outperforming the current peak-finding method and rivaling Convolutional Neural Networks (CNNs). The GP model's key advantage is its ability to provide reliable probability estimates for blends, crucial for robust astronomical analysis.

What Problem Does This Paper Address?

The upcoming Vera C. Rubin Observatory's Legacy Survey of Space and Time (LSST) will observe an unprecedented number of galaxies. However, due to the survey's depth, a significant fraction of these galaxies will overlap along the line of sight, creating "blends." Accurately identifying these blends is critical; misclassifying a blend can lead to incorrect galaxy counts or properties. The current standard method in LSST pipelines relies on counting intensity peaks in smoothed images, but its reliability, especially for LSST's specific characteristics, hasn't been thoroughly studied. This paper aims to address this by evaluating the peak-finding method and proposing a more robust, probabilistic approach.

Methods

The authors developed a novel Gaussian Process (GP) blend classifier. This model treats blend identification as a regression problem, predicting a "blendedness" score for each galaxy "footprint" (a region of interest). The GP model is trained on simulated LSST images and uses a Matérn kernel to define the similarity between different footprints in a PCA-transformed feature space. For comparison, they also evaluated the standard peak-finding method and a Convolutional Neural Network (CNN). The peak-finding method identifies blends by counting local intensity peaks after convolving the image with the telescope's point spread function (PSF). The CNN model is a standard architecture adapted for this task, employing convolutional layers, pooling, and dense layers.

Datasets

The study utilizes simulated LSST i-band images generated from the cosmoDC2 v1.1.4 catalog and the DESC DC2 Truth Match dr6 v1 catalog. These catalogs provide realistic distributions of galaxy properties (luminosity, size, orientation, redshift) and their spatial distribution, based on cosmological simulations. The authors created a benchmark dataset of these simulated images to train and evaluate their models.

Key Insights and Ideas

  • Probabilistic Blending is Key: The most significant insight is the critical need for probabilistic blend identification. The peak-finding method provides a binary classification, lacking uncertainty estimates, whereas the GP model offers reliable probabilities, allowing astronomers to set confidence thresholds for analysis.
  • Footprint Definition Matters: The way "footprints" (regions of interest for classification) are defined significantly impacts the performance of the peak-finding method, highlighting its sensitivity to image processing choices. The GP model, however, shows greater robustness to these variations.
  • GP vs. CNN Trade-offs: While the CNN model achieves high accuracy, the GP model offers comparable performance with significantly lower computational resource requirements (training time and memory), making it more practical for large-scale surveys.
  • Model Complementarity: The GP and CNN models make different types of errors, suggesting that a hybrid approach could potentially yield even better results, though this would complicate probabilistic interpretation.

Key Results

  • The Gaussian Process (GP) classifier achieved competitive accuracy with the Convolutional Neural Network (CNN), often outperforming the standard peak-finding method, especially on "Custom" and "Variant" footprint definitions.
  • For example, on "Custom" footprints, the GP model achieved a balanced accuracy of 84.85%, compared to 77.14% for peak-finding and 87.57% for the CNN.
  • The GP model's probabilistic outputs were found to be generally reliable, meaning the assigned blend probability closely matched the actual fraction of blended footprints in test bins.
  • The peak-finding method's accuracy was highly sensitive to the definition of a "footprint," while the GP model showed much greater robustness.

Limitations and Future Work

  • Model Limitations: The authors acknowledge that while the GP model provides probabilities, it doesn't directly estimate galaxy positions, which is necessary for deblending. The CNN model, while accurate, can produce overconfident predictions, both correct and incorrect.
  • Future Work:
    • Developing models for galaxy localization to complement the blend classification.
  • Exploring multi-class classifiers to distinguish between different numbers of galaxies in a blend (e.g., 2 vs. 3 vs. 4+).
  • Investigating the impact of incorporating other filter bands into the GP model.
  • Refining the GP prior to better account for the a priori likelihood of blends based on galaxy brightness.
  • Exploring hybrid models (e.g., GP + CNN) to leverage their complementary strengths.
  • Testing the models on real LSST data and potentially data from other surveys like HSC.
  • Investigating the impact of more realistic image simulations, including stars and other astrophysical phenomena.

Summary updated

Abstract

Abstract A significant fraction of observed galaxies in the Rubin Observatory Legacy Survey of Space and Time (LSST) will overlap at least one other galaxy along the same line of sight, in a so-called “blend.” The current standard method of assessing blend likelihood in LSST images relies on counting up the number of intensity peaks in the smoothed image of a blend candidate, but the reliability of this procedure has not yet been comprehensively studied. Here we construct a realistic distribution of blended and unblended galaxies through high-fidelity simulations of LSST-like images, and from this we examine the blend classification accuracy of the standard peak-finding method. Furthermore, we develop a novel Gaussian process blend classifier model, and show that this classifier is competitive with both the peak finding method as well as with a convolutional neural network model. Finally, whereas the peak-finding method does not naturally assign probabilities to its classification estimates, the Gaussian process model does, and we show that the Gaussian process classification probabilities are generally reliable.
BibTeX
@article{buchanan2022gaussian,
  author = {J. Buchanan and M. Schneider and R. Armstrong and Amanda Muyskens and Benjamin W. Priest and Ryan Dana},
  title = {{Gaussian Process Classification for Galaxy Blend Identification in LSST}},
  journal = {The Astrophysical Journal},
  year = {2022},
  doi = {10.3847/1538-4357/ac35ca},
}
Figure