Robustness Analysis of USmorph: I. Generalization Efficiency of Unsupervised Strategies and Supervised Learning in Galaxy Morphological Classification

We conduct a systematic robustness analysis of the hybrid machine learning framework \texttt{USmorph}, which integrates unsupervised and supervised learning for galaxy morphological classification. Although \texttt{USmorph} has already been applied to nearly 100,000 $I$-band galaxy images in the COSMOS field ($0.2 < z < 1.2$, $I_{\mathrm{mag}} < 25$), the stability of its core modules has not been quantitatively assessed. Our tests show that the convolutional autoencoder (CAE) achieves the best performance in preserving structural information when adopting an intermediate network depth, $5\times5$ convolutional kernels, and a 40-dimensional latent representation. The adaptive polar coordinate transform (APCT) effectively enhances rotational invariance and improves the robustness of downstream tasks. In the unsupervised stage, a bagging clustering number of $K=50$ provides the optimal trade-off between classification granularity and labeling efficiency. For supervised learning, we employ GoogLeNet, which exhibits stable performance without overfitting. We validate the reliability of the final classifications through two independent tests: (1) the t-distributed stochastic neighbor embedding (t-SNE) visualization reveals clear clustering boundaries in the low-dimensional space; and (2) the morphological classifications are consistent with theoretical expectations of galaxy evolution, with both true and false positives showing unbiased distributions in the parameter space. These results demonstrate the strong robustness of the \texttt{USmorph} algorithm, providing guidance for its future application to the China Space Station Telescope (CSST) mission.

💡 Research Summary

This paper presents a comprehensive robustness analysis of the hybrid machine‑learning framework USmorph, which combines unsupervised feature extraction with supervised classification for galaxy morphology. Using a well‑defined sample of 99,806 I‑band galaxies from the COSMOS2020 “Farmer” catalog (0.2 < z < 1.2, I mag < 25), the authors evaluate each core module under a variety of hyper‑parameter settings.

In the unsupervised stage, a convolutional autoencoder (CAE) is employed to denoise and compress the 100 × 100 pixel images. Experiments show that a moderate network depth with two 5 × 5 convolutional layers, each followed by max‑pooling, and a latent space of 40 dimensions yields the lowest reconstruction loss while preserving fine structural details. Kernel sizes of 3 × 3, 5 × 5, and 7 × 7 are compared; the 5 × 5 configuration offers the best trade‑off between noise removal and retention of faint features.

An adaptive polar coordinate transform (APCT) is then applied to the CAE embeddings, effectively normalizing rotational variance so that rotated versions of the same galaxy map to identical feature vectors. This step markedly sharpens cluster boundaries in subsequent analysis.

For clustering, a bagging‑based multi‑clustering approach is tested with the number of clusters K ranging from 20 to 100. The authors find that K = 50 optimally balances labeling efficiency and morphological granularity, minimizing intra‑cluster variance while providing sufficient resolution for downstream tasks.

The supervised stage uses GoogLeNet as the final classifier. Across a range of learning rates, batch sizes, and epochs, the network exhibits stable performance without over‑fitting, achieving an overall validation accuracy of ~94 %.

Robustness is validated through two independent methods. First, t‑distributed stochastic neighbor embedding (t‑SNE) visualizations reveal well‑separated low‑dimensional clusters corresponding to distinct morphological classes. Second, the predicted labels are compared against physical parameters such as Sérsic index, effective radius, Gini coefficient, and M20. Both true and false positives display unbiased distributions in this parameter space, confirming that the classifications are consistent with established galaxy‑evolution trends.

The study concludes that USmorph’s components—CAE architecture, APCT, bagging clustering with K = 50, and GoogLeNet classification—are individually and collectively robust. These findings support the deployment of USmorph for upcoming large‑scale surveys, notably the China Space Station Telescope (CSST), where automated, reliable morphology classification will be essential.