Machine Learning Techniques to Distinguish Giant Stars from Dwarf Stars Using Only Photometry -- Pushing Redwards

TL;DR

This study addresses the challenge of separating giant stars in Local Group galaxies from foreground Milky Way dwarfs using photometry alone. It leverages the gravity-sensitive NB515 filter, centered at $5150$Å with a FWHM of $77$Å, in combination with broadband $g$ and $i$ data, and builds synthetic photometry from the MaStar and XSL spectral libraries to understand how $NB515-g$ and $g-i$ track stellar parameters. A neural network classifier is trained on field-specific synthetic data (Fornax and two M31 fields) augmented with realistic photometric uncertainties and foreground contamination modeled with Besançon, achieving over 85% accuracy in identifying red giants and their extragalactic membership. The approach extends traditional two-color cuts to metal-rich M-stars, enabling efficient pre-selection for spectroscopic follow-up and advancing studies of star formation histories and chemical evolution in nearby galaxies. The results demonstrate robust gravity-smart photometric classification across a range of metallicities and temperatures, with practical applicability to Local Group surveys.

Abstract

We present our photometric method, which combines Subaru/HSC $NB515$, g, and i band filters to distinguish giant stars in Local Group galaxies from Milky Way dwarf contamination. The $NB515$ filter is a narrow-band filter that covers the MgI+MgH features at $5150$ Å, and is sensitive to stellar surface gravity. Using synthetic photometry derived from large empirical stellar spectral libraries, we model the $NB515$ filter's sensitivity to stellar atmospheric parameters and chemical abundances. Our results demonstrate that the $NB515$ filter effectively separates dwarfs from giants, even for the reddest and coolest M-type stars. To further enhance this separation, we develop machine learning models that improve the classification on the two-color ($g-i$, $NB515-g$) diagram. We apply these models to photometric data from the Fornax dwarf spheroidal galaxy and two fields of M31, successfully identifying red giant branch stars in these galaxies.

Figures (13)

• Figure 1: Total response curves of the $g$, $i$, and $NB515$ filters of HSC, overlaid by the observed spectrum of an M-giant (blue) and of an M-dwarf (purple) from the X-shooter Spectral Library. Both spectra have been down-sampled to a resolving power of 2000 and arbitrarily normalized for clarity. The zoomed-in plot highlights the $NB515$ filter, which is centered around $5150\text{\AA}$ and includes the MgH+Mgb absorption features.
• Figure 2: $Gaia$ color-magnitude diagram of stars in the MaStar Library and two-color diagram of MaStar synthetic photometry, color-coded by $\log(g)$ and [Fe/H] from the mean MaStar Library parameters. A reddening vector is shown in the top-left panel, with length corresponding to the mean extinction of the catalog ($A_v = 0.37$).
• Figure 3: Two-color diagrams of XSL synthetic photometry, color-coded by $\log(g)$ and [Fe/H] from the XSL parameters.
• Figure 4: Color-magnitude and two-color diagrams of the observed data (left) and training set (right) for this field in the inner halo of M31, with giants in the training set color-coded by [Fe/H].
• Figure 5: Two-color diagram and color-magnitude diagram of Fornax, color-coded by the predicted membership label.