AstroVision
Deep learning for galaxy morphological classification.
Galaxy10 DECaLS · DINOv2 + morphometrics + SDSS · 83.9% balanced accuracy (DINOv2 fine-tuned) · MC Dropout · Conformal prediction
Quick navigation
| Section | What you'll find |
|---|---|
| Overview | Scientific motivation, key results |
| Data | Galaxy10 DECaLS, morphology classes, SDSS crossmatch |
| Models | CNN, EfficientNet, DINOv2, Late-fusion |
| Results | Benchmarking, uncertainty, physical interpretation |
| Guides | Installation, training, GradCAM, segmentation |
| API | Module reference |
Key result
Central finding
DINOv2 visual representations encode physically meaningful galaxy structure, but only partly: the second principal component of the zero-shot feature space correlates with the Gini coefficient (ρ = 0.53), and a Lasso regression recovers photometric colour g−r from the 768 features with a cross-validated R² of 0.443. More than half of the colour variance is not recovered. Details and caveats: Physical Interpretation.
The late-fusion model reaches 86.3% on its own test split, but this is not a gain from the added modalities (see Evaluation Protocol).