Abstract
We present a labeled set of 14,600 cutouts from 312 digitized DASCH plates and a classifier that sorts detections into stars, five kinds of defect, and handwritten annotations. The classifier agrees with human reviewers on 93 percent of held-out cutouts. We release the labels, the model, and a catalog of about 41,000 classified detections from 1,200 plates.
1. Background
Between 1885 and 1992, Harvard College Observatory exposed and archived several hundred thousand astronomical photographic plates. The DASCH project scanned the collection, finishing in 2024, and its final data release provides the scans together with calibrated catalogs (Williams 2025).
Automated source extraction on plate scans finds many objects that are not stars. Emulsion contains grains and bubbles, plates collect dust and scratches, and many plates carry ink marks made by the astronomers who used them. DASCH’s own pipeline handles much of this. The aim of this project is to see how far a simple image classifier can help with the detections that remain hard to sort.
2. Sample
We selected 312 plates from the DASCH data release, spread across the decades of the collection, across telescope series and across the sky. From each plate we took cutouts of 64 × 64 pixels around a random sample of detections, giving 14,600 cutouts.
3. Labeling
Each cutout was labeled independently by two members, using a simple web tool, into one of seven classes:
| Class | Description | Share of labels |
|---|---|---|
| Star | A real astronomical image, including blended stars | 61.2% |
| Grain | Small emulsion grain or clump | 11.8% |
| Dust | Dust or debris on the plate or scanner | 8.9% |
| Scratch | Linear damage to the emulsion | 4.1% |
| Bubble | Bubble or void in the emulsion | 2.6% |
| Edge | Plate edge, label or scanner artifact | 3.7% |
| Annotation | Ink or pencil mark made by hand | 7.7% |
The two labelers agreed on 89 percent of cutouts. Disagreements were settled by a third member. Most disagreements were between “star” and “grain” for very faint objects.
4. Classifier
We fine-tuned a ResNet-18 (He et al. 2016) pre-trained on everyday images, with the plate cutouts as single-channel input. Plates, not cutouts, were split between training (70 percent), validation (15 percent) and test (15 percent) sets, so no plate contributes to more than one set.
5. Results
On the test set the classifier agreed with the human label on 93.0 percent of cutouts.
| Class | Precision | Recall |
|---|---|---|
| Star | 0.96 | 0.97 |
| Grain | 0.81 | 0.78 |
| Dust | 0.90 | 0.88 |
| Scratch | 0.92 | 0.89 |
| Bubble | 0.87 | 0.84 |
| Edge | 0.95 | 0.94 |
| Annotation | 0.94 | 0.93 |
As with the human labelers, the main confusion is between faint stars and grains. Performance is lower on plates from the earliest decades, which are under-represented in the sample.
6. Catalog
We ran the classifier over about 41,000 detections from 1,200 plates. The catalog lists each detection’s plate, position on the plate, sky coordinates from the DASCH astrometric solution, predicted class and class probabilities. It does not replace the DASCH catalogs and should be used together with them.
7. Availability
Labels, model weights and catalog: CC BY 4.0. Code: MIT license. Users of the catalog should also acknowledge DASCH as described below.
Acknowledgments
This work has made use of data provided by Digital Access to a Sky Century @ Harvard (DASCH). The DASCH project at Harvard is grateful for partial support from NSF grants AST-0407380, AST-0909073 and AST-1313370. We thank the many ASRU members who took part in labeling sessions.
References
- Grindlay, J., Tang, S., Los, E. & Servillat, M. 2012, Proceedings of the International Astronomical Union, IAU Symposium 285, 29
- He, K., Zhang, X., Ren, S. & Sun, J. 2016, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 770
- Williams, P. K. G. 2025, arXiv:2501.12977