Radio Morphology Eval

---
name: radio-morphology-eval
description: Probes the transfer learning capability of self-supervised vision models on radio astronomy morphology classification tasks across heterogeneous imaging pipelines, telescopes, and label granularities. Use when the user wants to benchmark on MiraBest, LoTSS DR2, Radio Galaxy Zoo DR1, or asks about evaluating this task. Reports accuracy.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2603.29660
  bibtex_key: [REDACTED]
  confidence: high
---

# radio-morphology-eval

> STRADAViT: Towards a Foundational Model for Radio Astronomy through Self-Supervised Transfer — DeMarco et al. (2026) (arXiv:2603.29660, 2026)

## What this evaluates

Probes the transfer learning capability of self-supervised vision models on radio astronomy morphology classification tasks across heterogeneous imaging pipelines, telescopes, and label granularities.

## Datasets

- **MiraBest** — total 1563; splits: test (1563)
- **LoTSS DR2** — total 8805; splits: test (8805)
- **Radio Galaxy Zoo DR1** — total 98391; splits: test (98391)

## Metrics

- `accuracy` **(primary)** — range: [0, 1]
  - Standard classification accuracy: the fraction of correctly predicted morphology labels out of the total number of evaluated samples.

## Input / output format

**Input**: Single-channel radio astronomy cutout images, standardized via per-image ZScale contrast stretch, replicated to three channels, and resized to 224×224 (ViT-MAE) or 518×518 (DINOv2).

**Output**: Discrete morphology class label (e.g., FR I/FR II for MiraBest; FR I, FR II, Hybrid, Spiral, Relaxed double for LoTSS; 1c1p, 1c2p, 1c3p, 2c2p, 2c3p, 3c3p for RGZ).

## Scoring recipe

```python
def compute_accuracy(predictions, gold_labels):
    correct = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
    return correct / len(gold_labels)
```

## Common pitfalls

- Excluding uncertain labels or hybrid sources significantly alters dataset size and class distribution.
- Model input resolution is backbone-dependent (224 vs 518), which must be accounted for when comparing performance across architectures.
- FITS cutouts may contain NaN/Inf or blank values that require explicit filtering before evaluation.

## Evidence (verbatim from paper)

> We evaluate transfer learning on three public, morphology-focused radio astronomy imaging datasets that span different telescopes, labeling paradigms, and class granularities: MiraBest, the LoTSS DR2 visual-classification sample of Horton et al. (2025), and Radio Galaxy Zoo DR1 (Wong et al., 2024).

## Citation

```bibtex
@misc{demarco2026stradavit,
  title={STRADAViT: Towards a Foundational Model for Radio Astronomy through Self-Supervised Transfer},
  author={DeMarco et al. (2026)},
  year={2026},
  note={arXiv:2603.29660}
}
```

- arXiv: 2603.29660