The Next Generation Fornax Survey (NGFS)

VIII. A support vector machine approach to disentangling globular clusters

¹ Instituto de Estudios Astrofísicos, Facultad de Ingeniería y Ciencias, Universidad Diego Portales Av. Ejército Libertador 441 Santiago, Chile
² Instituto de Astrofísica, Pontificia Universidad Católica de Chile Av. Vicuña Mackenna 4860 Santiago 7820436, Chile
³ Instituto de Astrofísica, Universidad Andres Bello Fernandez Concha 700 Las Condes Santiago, Chile
⁴ Vicerrectoría de Investigación y Postgrado, Universidad de La Serena La Serena 1700000, Chile
⁵ University of Calgary 2500 University Drive NW Calgary Alberta T2N 1N4, Canada
⁶ International Gemini Observatory/NSF NOIRLab Casilla 603 La Serena, Chile

^★ Corresponding author: This email address is being protected from spambots. You need JavaScript enabled to view it.

Received: 17 July 2025
Accepted: 17 December 2025

Abstract

Context. Wide-field, multiband surveys are capable of detecting millions of unresolved sources in nearby galaxy clusters; however, separating globular clusters (GCs) from foreground stars and background galaxies remains challenging. Scalable and automated classification methods are therefore essential to transform forthcoming data from facilities such as the Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST), Euclid, and the Nancy Grace Roman Space Telescope into robust constraints on galaxy assembly.

Aims. We present a supervised machine-learning method to separate GCs, stars, and galaxies using their distribution in color-color space. The primary objective is to recover a clean and reliable GC sample optimized for next-generation survey volumes.

Methods. We analyzed the central 3 deg² of the Next Generation Fornax Survey (NGFS), which images the Fornax cluster in u′g′i′ (BLANCO/DECam) and JK_s (VISTA/VIRCAM). We trained a support vector machine (SVM; svm.SVC implemented in scikit-learn) using spectroscopically confirmed sources. The initial model employed 15 features, including all color combinations from u′g′i′JK_s and basic morphological parameters (e.g., FWHM and ellipticity).

Results. Color combinations linking near-ultraviolet (NUV) and optical to near-infrared (NIR) wavelengths, particularly (u′−g′) versus (g′−K_s), provide the strongest discrimination among object classes. The full 15-feature model achieves an accuracy of 97.3%. A reduced seven-feature model, built from the most informative and least correlated features, attains a 96.6% level of accuracy with a misclassification rate of 10.4%, offering a more efficient and robust solution. Excluding the u′ or NIR bands significantly degrades performance. Tests using LSST-like filters, constructed from NGFS u′g′i′ and Dark Energy Survey r′z′Y data, show that the u′ and Y bands are essential, although models lacking NIR coverage remain suboptimal.

Conclusions. Broad spectral energy distribution coverage combined with simple morphological parameters enables an accurate and scalable classification of unresolved sources. The inclusion of NIR data substantially improves GC identification and the joint exploitation of LSST with Euclid and Roman observations will further enhance machine-learning approaches in large extragalactic surveys.