Galaxy merger challenge: A comparison study between machine learning-based detection methods

Various galaxy merger detection methods have been applied to diverse datasets. However, it is difficult to understand how they compare. We aim to benchmark the relative performance of machine learning (ML) merger detection methods. We explore six leading ML methods using three main datasets. The first one (the training data) consists of mock observations from the IllustrisTNG simulations and allows us to quantify the performance metrics of the detection methods. The second one consists of mock observations from the Horizon-AGN simulations, introduced to evaluate the performance of classifiers trained on different, but comparable data. The third one consists of real observations from the Hyper Suprime-Cam Subaru Strategic Program (HSC-SSP) survey. For the binary classification task (mergers vs. non-mergers), all methods perform reasonably well in the domain of the training data. At $0.1<z<0.3$, precision and recall range between $\sim$70\% and 80\%, both of which decrease with increasing $z$ as expected (by $\sim$5\% for precision and $\sim$10\% for recall at $0.76<z<1.0$). When transferred to a different domain, the precision of all classifiers is only slightly reduced, but the recall is significantly worse (by $\sim$20-40\% depending on the method). Zoobot offers the best overall performance in terms of precision and F1 score. When applied to real HSC observations, all methods agree well with visual labels of clear mergers but can differ by more than an order of magnitude in predicting the overall fraction of major mergers. For the multi-class classification task to distinguish pre-, post- and non-mergers, none of the methods offer a good performance, which could be partly due to limitations in resolution and depth of the data. With the advent of better quality data (e.g. JWST and Euclid), it is important to improve our ability to detect mergers and distinguish between merger stages.