An automated method to detect and characterise semi-resolved star clusters

We present a novel method for automatically detecting and characterising semi-resolved star clusters: clusters where the observational point-spread function (PSF) is smaller than the cluster’s radius, but larger than the separations between individual stars. We apply our method to a 1.77 deg$^2$ field located in the Large Magellanic Cloud (LMC) using the VISTA survey of the Magellanic Clouds (VMC), which surveyed the LMC in the $YJK_\text{s}$ bands. Our approach first models the position-dependent PSF to detect and remove point sources from deep $K_\text{s}$ images; this leaves behind extended objects such as star clusters and background galaxies. We then analyse the isophotes of these extended objects to characterise their properties, perform integrated photometry, and finally remove any spurious objects this procedure identifies. We demonstrate our approach in practice on a deep VMC $K_\text{s}$ tile that contains the most active star-forming regions in the LMC: 30 Doradus, N158, N159, and N160. We select this tile because it is the most challenging for automated techniques due both to crowding and nebular emission. We detect 682 candidate star clusters, with an estimated contamination rate of 13% from background galaxies and chance blends of physically unrelated stars. We compare our candidates to publicly available James Webb Space Telescope data and find that at least 80% of our detections appear to be star clusters.

💡 Research Summary

The paper introduces a fully automated pipeline designed to detect and characterize semi‑resolved star clusters in the Large Magellanic Cloud (LMC) using deep near‑infrared Kₛ‑band images from the VISTA Survey of the Magellanic Clouds (VMC). Semi‑resolved clusters are defined as objects whose angular radius is at least twice the point‑spread function (PSF) while the typical separation between individual stars is much smaller than the PSF, a regime where traditional star‑by‑star clustering algorithms fail due to crowding.

The authors first construct a position‑dependent PSF model that accounts for variations across the VIRCAM detector mosaic and across multiple observing epochs. Using this model they identify all point‑like sources in the deep Kₛ tile and subtract a scaled PSF from each, producing a residual image that is essentially free of isolated stars.

On the residual image they perform an isophotal analysis: isophotes are extracted at several signal‑to‑noise levels, and for each isophote the centroid, ellipticity, position angle, and radial profile are measured. From these measurements they derive an effective radius, mean surface brightness, asymmetry parameters, and integrated flux for every extended source. A series of automated culling steps—based on shape, surface‑brightness profile, and comparison with a background model—removes spurious detections such as background galaxies and chance alignments of unrelated stars.

The method is applied to a 1.77 deg² VMC tile that includes the most active star‑forming complexes in the LMC (30 Doradus, N158, N159, N160). In this highly crowded and nebular‑rich region the pipeline identifies 682 candidate clusters. Cross‑matching with existing LMC cluster catalogs and visual inspection of publicly available James Webb Space Telescope (JWST) NIRCam images confirm that at least 80 % of the candidates are genuine star clusters, while the estimated contamination from background galaxies and random stellar blends is about 13 %.

Key strengths of the approach are: (1) complete automation that eliminates the need for subjective visual inspection, (2) robust handling of spatially varying PSF, crucial for ground‑based data with typical 1″ seeing, (3) direct measurement of structural parameters from isophotes, enabling subsequent age and mass estimates. Limitations include reduced sensitivity to very low surface‑brightness objects and potential residual artifacts in regions with strong nebular emission where background modeling is challenging.

The authors suggest future improvements such as incorporating multi‑band (Y, J) information for colour‑based filtering, employing deep‑learning techniques to refine residual‑image cleaning, and developing real‑time PSF tracking to further reduce systematic errors. Overall, the study provides a powerful new tool for building homogeneous, large‑scale catalogs of semi‑resolved clusters not only in the LMC but also in other nearby galaxies where similar observational constraints apply.