Stellar Blend Image Classification Using Computationally Efficient Gaussian Processes

TL;DR

This paper introduces a computationally efficient Gaussian process model, MuyGPs, combined with data normalization techniques, to accurately classify stellar blends in astronomical images, especially with limited training data.

Contribution

The study demonstrates that MuyGPs with specific normalization outperforms benchmarks in stellar blend classification, providing confidence measures for improved human-in-the-loop analysis.

Findings

MuyGPs achieves 83.8% accuracy on stellar blend classification.

MuyGPs outperforms benchmark models, especially with limited data.

Confidence bands enable targeted human review.

Abstract

Stellar blends, where two or more stars appear blended in an image, pose a significant visualization challenge in astronomy. Traditionally, distinguishing these blends from single stars has been costly and resource-intensive, involving sophisticated equipment and extensive expert analysis. This is especially problematic for analyzing the vast data volumes from surveys, such as Legacy Survey of Space and Time (LSST), Sloan Digital Sky Survey (SDSS), Dark Energy Spectroscopic Instrument (DESI), Legacy Imaging Survey and the Zwicky Transient Facility (ZTF). To address these challenges, we apply different normalizations and data embeddings on low resolution images of single stars and stellar blends, which are passed as inputs into machine learning methods and to a computationally efficient Gaussian process model (MuyGPs). MuyGPs consistently outperforms the benchmarked models, particularly on limited training data. Moreover, MuyGPs with $r^\text{th}$ root local min-max normalization achieves 83.8% accuracy. Furthermore, MuyGPs' ability to produce confidence bands ensures that predictions with low confidence can be redirected to a specialist for efficient human-assisted labeling.