The goal of this project was to produce a Machine Learning algorithm to reduce human  annotation effort in identifying star clusters in nearby galaxies. The standard process involves human classification of sources in multi-color high resolution images of galaxies, after an initial selection of the sources performed with automatic tools. These tools identify potential sources, but also include many contaminants (stars, background galaxies, image artifacts). The contaminants can represent between 50% and 70% of the total initial source catalogs, depending on the parameters used for the automatic selection. Hence, there is a need for visual inspection of the automatic catalogs, in order to separate bona-fide clusters from contaminants. This part of securing star cluster catalogs is time consuming, often requiring weeks to months to complete a galaxy, depending on the richness of its cluster population. In addition, different human classifiers can reach different conclusions on the nature of a source. This is because beyond a few Mpc, star clusters are no longer resolved into individual stars, and often appear barely more extended than a point source (e.g., a star). Detailed visualization of the candidate is thus required for obtaining a classification. Because of this problem, Citizen Science approaches have not been successful when applied to galaxies beyond our Local Group.

Machine Learning provides an excellent approach to secure cluster populations while avoiding the bottleneck of human classification, or at least lightening the effort of the human classifiers. In addition to acceleratinng the pace of classifications, by reducing to several minutes to a few hours a task that normally takes weeks-to-months, Machine Learning offers the advantage of repeatability and homogeneity, once the algorithms are trained. Finally, machines do not get tired, while humans do: informal tests have shown that classifications performed by humans late in a day are of worse quality than those performed earlier in the day.

The product of this project, StarcNet, is a custom multi-layer convolutional neural network (CNN) that ingests automatic catalogs of cluster candidates together with multi-band cut-outs of the galaxy images at the position of each candidate and runs them through three pathways operating at different resolutions(see Figure). Each pathway of the CNN consists of seven convolutional layers which are later connected to produce a prediction for the candidate.

Tests run on existing (human classified) star cluster catalogs demonstrate that StarcNet performs as well as human classifiers, with an overall accuracy of 68.6% - human classifiers are around 70%-75% for person-to-person agreement -  when clusters are classified into one of four classes. For a binary classification task (cluster/non-cluster) the accuracy is significantly  higher, around 86% for StarcNet, comparable to the 87% of humans.

Comparisons with other published CNN algorithms show that StarcNet outperforms them, at least in this one task. Finally, tests run on the training sets show that the current existing samples (of about 15,000 classified sources) are barely sufficient for training, and larger sets are more desirable than homogeneous sets.

The results from this project are published in:
Perez, G., Messa, M., Calzetti, D., Maji, S., Jung, D., Adamo, A., &  Sirressi, M., `StarcNet: Machine Learning for Star CLuster Identification', 2021, The Astrophysical Journal, 907, 100.

The StarcNet software and training samples are released through Github and published in the Astrophysics Source Code Library, record ascl:2106.012.  

Last Modified: 11/26/2021
Modified by: Daniela Calzetti