Dual-coding contrastive learning based on ConvNeXt and ViT models for morphological classification of galaxies in COSMOS-Web

TL;DR

This work tackles scalable galaxy morphology classification in the COSMOS-Web field under limited labeled data by integrating self-supervised contrastive learning with a dual-encoder (ConvNeXt and ViT) framework, CAE denoising, and APCT rotational augmentation. The method first extracts compact features via a dual-encoder contrastive loss, then applies Bagging clustering to label 32,922 galaxies, and finally trains GoogLeNet to classify the remaining 12,366 objects, achieving 73% UML labeling and 27% SML labeling. Validation against parametric (Sérsic $n$, $r_e$) and nonparametric (G, $M_{20}$, $C$, $\Psi$, MID) morphology measures demonstrates strong concordance with galaxy evolution trends. The resulting 45,288-galaxy catalog, with five morphological classes, provides a robust, scalable resource for current and future surveys, including CSST, enabling efficient morphology-driven studies at $0.5<z<6.0$.

Abstract

In our previous works, we proposed a machine learning framework named \texttt{USmorph} for efficiently classifying galaxy morphology. In this study, we propose a self-supervised method called contrastive learning to upgrade the unsupervised machine learning (UML) part of the \texttt{USmorph} framework, aiming to improve the efficiency of feature extraction in this step. The upgraded UML method primarily consists of the following three aspects. (1) We employ a Convolutional Autoencoder to denoise galaxy images and the Adaptive Polar Coordinate Transformation to enhance the model's rotational invariance. (2) A pre-trained dual-encoder convolutional neural network based on ConvNeXt and ViT is used to encode the image data, while contrastive learning is then applied to reduce the dimension of the features. (3) We adopt a Bagging-based clustering model to cluster galaxies with similar features into distinct groups. By carefully dividing the redshift bins, we apply this model to the rest-frame optical images of galaxies in the COSMOS-Web field within the redshift range of $0.5 < z < 6.0$. Compared to the previous algorithm, the improved UML method successfully classifies 73\% galaxies. Using the GoogleNet algorithm, we classify the morphology of the remaining 27\% galaxies. To validate the reliability of our updated algorithm, we compared our classification results with other galaxy morphological parameters and found a good consistency with galaxy evolution. Benefiting from its higher efficiency, this updated algorithm is well-suited for application in future China Space Station Telescope missions.

Figures (12)

• Figure 1: Stellar mass as a function of redshift for the parent sample of this study. The stellar-mass completeness, given by $M_{\rm comp}(z)/M_{\odot} = -1.51 \times 10^6 (1 + z) + 6.81 \times 10^7 (1 + z)^2$, is indicated with the solid yellow line. Background gray dots represent the parent sample selected from the COSMOS2020 catalog (see Section \ref{['sec:2.2']}).
• Figure 2: Comparisons between the original and preprocessed galaxy images for six different morphological types. In each case, images are arranged in a sequence from left to right: original image, denoised image, and image after the polar coordinate transformation.
• Figure 3: Frameworks of the ConvNeXt (left) and ViT (right) models.
• Figure 4: Schematic diagram of UML Clustering Process: Upon completing data preprocessing (i.e., Step1), we employ both the ConvNeXt model and the ViT model to encode the data, thereby extracting compelling features (see Step2). The encoding results from identical data instances are used to construct positive sample pairs, while the remaining encodings from the ViT model serve to form a negative sample set. We subsequently apply the Contrastive Loss (CL) method to minimize loss, which helps in identifying critical features (Step3). For the final determination of labels in Step4, we adopt a voting mechanism based on a Bagging clustering model, supplemented by manual visual adjustments to ensure precision. This approach leverages the strengths of two advanced models, integrating automated learning with human oversight to deliver a precise and reliable feature identification and classification solution.
• Figure 5: Randomly select 100 images from each of the 10 machine clusters for display. We presented all 10 clustering results, which are characterized by small differences in the morphology of galaxies within the group and large differences in the morphology of galaxies outside the group. Therefore, we can visually classify each component into one of the five types (SPH, ETD, LTD, IRR, and UNC). Due to the limited number of categories, the visual efficiency is greatly improved.