| Abstract [eng] |
Galaxy morphology classification has traditionally been performed manually by experts or groups of volunteers. However, the rapid growth of astronomical data driven by advances in observational technology has made manual classification increasingly impractical. Although automated approaches have been introduced, they are often limited in classification scope due to data complexity, class imbalance, and variability across astronomical surveys. In this context, this research focuses on the application of convolutional neural networks (CNNs) as an effective and robust solution for automated galaxy morphology classification. The main objective of this work is to design, implement, and evaluate a CNN–based model for classifying galaxies according to their morphological features. The study utilizes data from the Galaxy Zoo project and multiple cosmological surveys, focusing on elliptical (smooth-round, cigar-shaped), spiral (barred and unbarred), and edge-on galaxies derived from the corresponding Galaxy Zoo decision trees. In addition, a convolutional vision transformer model (GalaxyCvT) is implemented solely for comparative purposes, representing a newer architectural approach. A comprehensive experimental evaluation is conducted under conditions of significant class imbalance, using the macro-averaged F1 score as the primary metric. Model performance is assessed across multiple datasets with differing image acquisition methods, spectral ranges, and observational equipment, and additional experiments analyze sensitivity to the amount of available training data. Model interpretability is examined using Grad-CAM visualizations to evaluate spatial stability and feature focus under image transformations and background noise, with results compared against attention-based visualizations from the CvT model. Furthermore, computational efficiency is evaluated in terms of training and inference performance. The proposed GalaxyCNN model achieves a high overall classification performance F1_macro = 0.897, outperforming the comparative GalaxyCvT model (F1_macro = 0.884), with particularly strong results for challenging classes (i.e. barred spiral galaxies) and improved robustness when trained on smaller datasets. Most misclassifications occur between morphologically similar classes, indicating that misclassifications are primarily caused by intrinsic ambiguity in galaxy morphology rather than architectural limitations. Visualization results confirm that the CNN model bases its decisions on stable, local morphological features that are robust to background noise, while the CvT model exhibits greater sensitivity to input variations. In conclusion, this thesis confirms that CNNs remain a highly effective and reliable approach for automated galaxy morphology classification, particularly in scenarios with limited and heterogeneous data. While transformer-based models represent a promising direction, their advantages are not fully realized under current data constraints; therefore, future work should focus on enhancing CNN architectures, expanding labeled datasets, integrating additional spectral information, and improving robustness and interpretability for large-scale astronomical surveys. |
