Skip to main content

Scientific Motivation

Galaxy morphology (smooth, barred spiral, merging, edge-on…) correlates with star-formation history and assembly history. Classifying it at survey scale needs models that are accurate, calibrated and, ideally, interpretable in physical terms. AstroVision uses Galaxy10 DECaLS as a controlled testbed for three questions.

1. Do foundation-model features transfer to galaxies?​

DINOv2 was trained on natural images. A linear probe on its frozen features reaches only 55.5% balanced accuracy, below a small CNN trained from scratch (59.5%). Full fine-tuning lifts it to 83.9%, above fine-tuned EfficientNet-B0 (82.7%). The gap between probe and fine-tuning is the measurable domain gap (probe, fine-tuning).

2. Do the features carry physical information?​

Two quantities were checked against the zero-shot DINOv2 features, using the SDSS DR16 cross-match and CAS/Gini/M20 morphometry:

  • the second principal component (13.2% of feature variance) correlates with the Gini coefficient (ρ = 0.53) and with the g−r colour (ρ = 0.30);
  • a Lasso regression of g−r on the 768 features reaches R² = 0.443 ± 0.103 in cross-validation.

These are moderate correlations, not proof that the model has learned physics; the Hubble sequence links concentration and colour, so a shared axis is expected from morphology alone. Details and caveats: Physical Interpretation.

Provenance. Model scores in this section are copied from the output cells of the AstroVision notebooks (and, for DINOv2 per-class recall, read from an exported figure). The documentation data pack contains no metrics file for them, so they cannot be re-derived from the pack; the cell of origin is listed for each value in _facts/astrovision/FACTS_INDEX.md §3 and shown when hovering over a number.

3. Can predictions be trusted and explained?​

Misclassifications concentrate in physically ambiguous classes (disturbed systems, loose spirals). AstroVision therefore pairs the classifier with uncertainty estimates (MC Dropout, calibration, conformal sets) and with attention- and gradient-based explanations (Grad-CAM and attention).

Open question: images × spectra​

The longer-term goal is to relate visual morphology to spectroscopic quantities such as metallicity — the link to AstroSpectro. A cross-match of Galaxy10 with LAMOST DR7 was attempted in the synthesis notebook and returned no data (the VizieR table was not available on the CDS XMatch server), so this remains a research direction, not a result.