Galaxy mergers classification using CNNs trained on Sérsic models, residuals and raw images

TL;DR

This work finds that galaxy merger classification is possible using either faint features or the position and Sersic profile information present in residual and model images, respectively, and demonstrates that not only faint features but also source position information play complementary roles in merger classification.

Abstract

Galaxy mergers are crucial for understanding galaxy evolution, and with large upcoming datasets, automated methods such as Convolutional Neural Networks (CNNs) are essential for efficient detection. It is understood that CNNs classify mergers by identifying deviations from the regular, expected shapes of galaxies, particularly faint features that are indicative of a merger event. In this work, we present a novel investigation of the relative importance of different morphological components, namely faint residual features and position and spatial structure, in CNN-based binary classification of galaxies into merger and non-merger classes. Using mock images from the IllustrisTNG simulations processed to mimic Hyper Suprime-Cam (HSC) observations, we fit Sérsic profiles to each galaxy and generate three datasets: original images, model images containing only smooth Sérsic profiles, and residual images highlighting faint features after model subtraction. We train three identical CNNs on these datasets: CNN1 on original images, CNN2 on model images, and CNN3 on residual images. CNN1, trained on full images, achieves the highest accuracy of 74 percent. CNN2, using only shape information including source position, achieves 70 percent, while CNN3, using only faint residual features, achieves 68 percent. We find that galaxy merger classification is possible using either faint features or the position and Sérsic profile information present in residual and model images, respectively. Our results demonstrate that not only faint features but also source position information play complementary roles in merger classification. This has important implications for the design and interpretation of machine learning methods for galaxy morphology, particularly in regimes where specific image components may be enhanced or suppressed.

Figures (13)

• Figure 1: Distribution of galaxies across redshifts. Y axis shows number of galaxies in each redshift bin and X axis shows the redshift bin.
• Figure 3: Preprocessing steps. a) Original image, b) segmentation map, c) deblending to isolate overlapping galaxies, d) Sérsic profile fits, and e) subtraction to produce residual image.
• Figure 4: Examples of mergers (a) and non-mergers (b) used for training. Left: original images. Middle: fitted Sérsic profiles. Right: residual images. A power-law normalization with exponent 0.2 was used to display the images.
• Figure 5: Data pipeline summary. Three datasets were created - raw images, Sérsic profile images, and residual images. Images were cropped to 128 × 128 pixels, linearly normalized, and divided in a 75:15:10 ratio into the training set, the validation set, and the test set.
• Figure 6: Performance metrics as a function of threshold of CNN1 (top panel), CNN2 (middle panel), and CNN3 (lower panel). Accuracy is shown with blue solid line. Precision with red dotted line. Recall with green dashed line, and F1 with black dashed dotted line.