The Next Generation Fornax Survey (NGFS).VIII. A Support Vector Machine Approach for Disentangling Globular Clusters from other Sources

TL;DR

This paper presents a support vector machine-based method for accurately distinguishing globular clusters from stars and galaxies in large survey data, leveraging multi-band photometry and morphological features to enable scalable, automated classification for future astronomical surveys.

Contribution

It introduces a supervised SVM classification approach utilizing broad spectral and morphological features, optimized for next-generation survey volumes, achieving over 96% accuracy in separating unresolved sources.

Findings

Achieved 97.3% accuracy with full feature set.

Omitting u' or near-IR bands reduces performance.

NIR bands significantly improve globular cluster classification.

Abstract

Wide-field, multi-band surveys now detect millions of unresolved sources in nearby galaxy clusters, yet separating globular clusters (GCs) from foreground stars and background galaxies remains challenging. Scalable, automated classification is therefore essential to convert the forthcoming data from facilities such as the Vera C. Rubin/LSST, the Roman and Euclid into robust constraints on galaxy assembly. We introduce a supervised classification method to separate GCs, stars, and galaxies based on their locations in color-color diagrams. The main objective is to recover a clean GC sample for future scientific analysis. The method exploits broad spectral energy distribution coverage, deep photometry, and is optimized for next-generation survey volumes. We use the central 3deg2 of the Next Generation Fornax Survey (NGFS), which images the Fornax cluster in u'g'i'JKs. We build a Support Vector Machine (SVM; svm.SVC, scikit-learn) using 15 features: all color combinations and basic morphological parameters. Spectroscopically confirmed sources define the training classes. Color pairs connecting near-UV/optical/near-IR. The full 15 feature model achieves 97.3% accuracy and a pruned 7 feature model built from the most informative, least correlated features achieves 96.6% accuracy. Misclassifications amount 8.4% and 10.4%, respectively. Omitting the u' or/and near-IR bands degrades performance. Emulating LSST filters with NGFS u'g'i' and DES r'z'Y shows that u' and Y bands are crucial, but models lacking NIR remain suboptimal. Combining broad SED coverage with simple morphological parameters enables precise, scalable separation of unresolved sources. Including NIR bands significantly improves GC classification, and joining LSST with forthcoming Euclid and Roman data will further enhance machine-learning frameworks.