An Efficient Galaxy Morphology Classification Framework updated with Two-Stage Feature Extraction with AlexNet Encoding and UMAP

In this study, we present an efficient framework for galaxy morphology classification by combining the AlexNet model for feature extraction with the Uniform Manifold Approximation and Projection (UMAP) algorithm for dimensionality reduction. Using the recently released COSMOS2025 catalogue, we select 100,362 galaxies at 0.5 < z < 6.0 with Imag < 27.5, and classify their morphology based on rest-frame optical images extracted from the COSMOS-Web survey. These galaxies are categorized into five morphological classes: spheroidal galaxies (SPH, 15,015), early-type disc galaxies (ETD, 34,361), late-type disc galaxies (LTD, 19,411), irregular galaxies (IRR, 21,034), and unclassified galaxies (UNC, 10,541). The performance of the supervised classification is assessed by using a validation set, which includes weighted averages of recall, precision, and an F1 score of ∼ 95%. The t-SNE-based visualisation of the feature space shows that different morphological classes form well-separated clusters with minimal overlap, supporting the robustness of the classification framework. We validate the classification results using parametric (Sérsic index n, effective radius re, and compactness Σ1.5) and non-parametric structural measurements. These measurements exhibit systematic trends that are consistent with established morphological sequences and theoretical expectations of galaxy evolution. This method provides a scalable approach for galaxy morphology classification and is well-suited for application to upcoming wide-field surveys, such as CSST.