[Paper Review] Evolutionary Deep Learning to Identify Galaxies in the Zone of Avoidance

David Jones, Anja Schroeder+1 more|arXiv (Cornell University)|Mar 6, 2019

Spectroscopy and Chemometric AnalysesChemistry38 references3 citations

TL;DR

This study proposes an evolutionary deep learning approach using a genetic algorithm to automatically evolve optimal Convolutional Neural Network (CNN) topologies for identifying galaxies in the Zone of Avoidance (ZoA), where high extinction and star density obscure extragalactic objects. The method achieved 91.75% accuracy on 32×32 near-infrared images using all J, H, and K passbands, successfully identifying 3,907 potential galaxies in unseen data, demonstrating the efficacy of automated topology evolution over manual tuning in challenging astronomical environments.

ABSTRACT

The Zone of Avoidance makes it difficult for astronomers to catalogue galaxies at low latitudes to our galactic plane due to high star densities and extinction. However, having a complete sky map of galaxies is important in a number of fields of research in astronomy. There are many unclassified sources of light in the Zone of Avoidance and it is therefore important that there exists an accurate automated system to identify and classify galaxies in this region. This study aims to evaluate the efficiency and accuracy of using an evolutionary algorithm to evolve the topology and configuration of Convolutional Neural Network (CNNs) to automatically identify galaxies in the Zone of Avoidance. A supervised learning method is used with data containing near-infrared images. Input image resolution and number of near-infrared passbands needed by the evolutionary algorithm is also analyzed while the accuracy of the best evolved CNN is compared to other CNN variants.

Motivation & Objective

• Address the challenge of galaxy identification in the Zone of Avoidance (ZoA), where high stellar density and extinction obscure extragalactic sources.
• Develop an automated system to classify galaxies in the ZoA, reducing reliance on time-consuming manual inspection.
• Evaluate whether evolutionary algorithms can optimize CNN topology and hyperparameters for improved performance in noisy, complex ZoA environments.
• Assess the impact of input resolution and multi-band near-infrared (J, H, K) data on classification accuracy.
• Provide astronomers with a scalable, automated tool to enhance galaxy cataloging and support cosmological research in the local universe.

Proposed method

• Employ an evolutionary algorithm (EA) to evolve the architecture and hyperparameters of Convolutional Neural Networks (CNNs), including filter counts, kernel sizes, and layer configurations.
• Use supervised learning with labeled near-infrared FITS images from the ZoA, where galaxies and non-galaxy sources are pre-labeled.
• Train and evaluate CNNs across multiple generations of the EA, using accuracy on a validation set as the fitness function.
• Apply the best-evolved CNN to unseen FITS data using the Galyxi Fynder inference module to detect potential galaxies.
• Compare performance across different input resolutions (e.g., 32×32) and number of passbands (J, H, K) to assess data input impact.
• Limit activation function evolution to a single function across all fully-connected layers, as a design constraint due to computational and data limitations.

Experimental results

Research questions

• RQ1 Can an evolutionary algorithm effectively optimize CNN topology for galaxy identification in the Zone of Avoidance, where data is noisy and complex?
• RQ2 Does using multi-band near-infrared (J, H, K) data improve classification accuracy compared to single-band inputs?
• RQ3 How sensitive is the performance of the evolved CNN to input image resolution, particularly at small sizes like 32×32?
• RQ4 To what extent does the evolutionary approach outperform manually tuned CNN configurations in terms of accuracy and generalization in the ZoA?
• RQ5 What is the potential of the evolved CNN to generalize to unseen data and identify new candidate galaxies in real astronomical surveys?

Key findings

• The evolved CNN achieved 91.75% accuracy in identifying galaxies in the Zone of Avoidance using only 32×32 near-infrared images, demonstrating strong generalization despite low resolution.
• The use of all three near-infrared passbands (J, H, K) marginally improved performance, supporting the hypothesis that multi-band data enhances detection capability.
• Contrary to expectations, very low input resolution (32×32) was sufficient for high accuracy, suggesting that learning on small images is efficient and effective for this task.
• The evolutionary algorithm converged in under 15 generations with relatively short processing time, indicating computational efficiency and practicality for astronomical workflows.
• The system successfully identified 3,907 potential galaxies in an unseen FITS image containing 20,690 cataloged light sources, highlighting its utility for large-scale sky surveys.
• The method outperformed prior attempts in the ZoA, including those by Lahav et al. (80–96%) and Drinkwater et al., and approached or exceeded results from non-ZoA studies (e.g., 97% accuracy by Khalifa et al.), despite the extreme observational challenges of the ZoA.