Resumo:
Morphological classification of galaxies in large photometric surveys is central to galaxy evolution
studies, but it remains challenging when deep learning models are transferred across
different data domains. This dissertation addresses the problem by developing and evaluating
deep learning models for binary galaxy classification (regular vs. peculiar), with emphasis on
quantifying the domain shift between the Sloan Digital Sky Survey (SDSS) and the Southern
Photometric Local Universe Survey (S-PLUS).
Using transfer learning, a benchmark of 11 pre-trained architectures was conducted, including
EfficientNet (B0-B4), ResNet/ResNeXt, and Vision Transformer (ViT-B/16 and ViT-B/32), plus
a simple CNN trained from scratch. The models were trained and validated on a curated dataset
of 11,662 galaxies with images from SDSS and S-PLUS, labeled from the RC3 catalog.
In the SDSS domain, the best model was ViT-B/32, with an F1-score of 0.923 and an AUC
of 0.963, followed by EfficientNet-B4 (F1-score 0.910) and ViT-B/16 (F1-score 0.899). The
superiority of attention-based models suggests that long-range spatial modeling is advantageous
for this task. In contrast, the simple CNN showed the worst performance (F1-score 0.655),
reinforcing the importance of transfer learning.
Evaluation of the best model on S-PLUS quantifies the performance degradation, with a drop
of 14.8 percentage points in F1-score (from 0.923 to 0.775), confirming a significant domain
shift effect. The Peculiar class is the most affected, with a 29.8% drop in recall, while average
predictive uncertainty increases by 65.4% and calibration worsens by more than 50%.
This work contributes a comparative architecture benchmark, a curated dataset, and a quantitative
analysis of cross-survey generalization, showing that domain adaptation is an indispensable
methodological step for the practical use of these models in different observational contexts.
