Deep Learning Recovers 1,353 Historical Astronomical Plates: China's Century-Old Night Sky Photographs Digitally Restored
China has been systematically collecting nighttime astronomical photographs since 1900. Now, a Transformer-based deep learning model has rescued over 1,300 plates that were previously unprocessable due to environmental deterioration.
The Archive
China's historical astronomical plate collection spans over a century of observations, digitized using optical scanners. These plates contain irreplaceable scientific data about:
- Variable stars
- Asteroid positions
- Cometary observations
- Long-term astronomical changes
The Problem
Suboptimal early storage conditions and environmental deterioration caused thousands of digitized plates to fail standard processing:
- Source extraction (SExtractor) couldn't identify stars reliably
- Astrometric matching with Gaia catalog failed
- Astrometry.net couldn't calibrate positions
The AI Solution
A Transformer-based classification model that:
- Takes cutouts of detected sources as input
- Uses multi-scale feature fusion to identify trustworthy stellar sources
- Trains on successful plates to learn what reliable sources look like
- Applied to 1,883 failed plates — successfully registered 1,353
Results
| Metric | Value |
|---|---|
| Failed plates attempted | 1,883 |
| Successfully recovered | 1,353 (71.9%) |
| Method | Transformer + multi-scale feature fusion |
| Reference catalogs | Astrometry.net + Gaia |
Why It Matters
- Scientific heritage — Recovering century-old observations that were thought lost
- Long-term astronomy — Enables studies of astronomical changes over 100+ years
- Digital preservation — Pipeline applicable to other historical archives worldwide
- AI for science — Demonstrates deep learning's value beyond typical image tasks
The Chinese astronomical plate archive joins similar digitization efforts worldwide, using AI to unlock scientific data trapped in deteriorating physical media.