Enhancing astrometric registration of Chinese historical Astronomical Digital Plates with deep learning

1. Problem Statement

China possesses a vast archive of approximately 29,314 digitized historical nighttime astronomical plates (dating from 1900 onwards). While many have been successfully processed, a significant subset of 2,530 single-exposure plates failed to undergo astrometric registration using traditional methods.

Root Causes of Failure:

• Physical Degradation: Long-term storage issues such as scratches, emulsion degradation, uneven backgrounds, and fungal growth.
• Detection Artifacts: Traditional source extraction tools (like SExtractor) struggle to distinguish between genuine stellar sources and artifacts caused by the aforementioned degradation.
• Limitations of Previous Methods: A previous study (Shang et al., 2024) utilized a Support Vector Machine (SVM) classifier based on engineered photometric features. While this recovered 2,184 plates, it failed on the remaining 2,530 plates due to the complexity of the noise and the inability of traditional machine learning to generalize across diverse degradation patterns.

2. Methodology

The authors propose a Deep Learning-based pipeline utilizing a Swin Transformer architecture to improve the classification of stellar vs. non-stellar sources, thereby enabling successful astrometric matching.

A. Workflow Overview

The process follows three main stages:

1. Source Extraction: SExtractor is used to detect raw intensity enhancements on the digitized plates.
2. Stellar Source Classification (The Core Innovation):
   • Instead of using tabular photometric parameters (as in SVM), the method uses image cutouts (64x64 pixels) of detected sources.
   • A Swin Transformer (a hierarchical Vision Transformer with shifted windows) is employed to classify these cutouts as "Stellar" (positive) or "Non-stellar/Artifact" (negative).
   • Training Strategy:
     • Base Model: Trained on a balanced dataset of 1,000 plates (15,696 sources total) derived from previously successfully calibrated plates across five Chinese observatories (SHAO, NAOC, PMO, YNAO, QDO).
     • Fine-tuning Strategy: To address plates with unique degradation or specific telescope characteristics, the authors employ a one-to-one fine-tuning approach. For uncalibrated plates, they construct small, supervised training subsets based on metadata (telescope aperture, exposure time, observatory) to adapt the Base Model to specific conditions.
3. Astrometric Matching: The classified "trustworthy" stellar sources are matched against the Gaia DR3 catalog using Astrometry.net for coarse alignment, followed by iterative refinement to derive the World Coordinate System (WCS).

B. Data Handling

• Training Set: Constructed from 15,696 calibrated plates. Sources were labeled by cross-matching SExtractor catalogs with Gaia DR2 (matching radius < 3x RMS).
• Exclusion Criteria: Plates with severe physical damage (peeling, chemical contamination) or those containing large targets (Moon, M31) were manually excluded, leaving 1,883 plates for the study.
• Quality Grading: The target plates were categorized by degradation: Grade 1 (Intact), Grade 2 (Minor defects <25%), and Grade 3 (Moderate defects <50%).

3. Key Contributions

• Deep Learning Architecture: Successfully adapted the Swin Transformer for the specific domain of historical astronomical plate analysis, demonstrating superior performance over traditional SVMs in distinguishing stars from complex artifacts.
• Hybrid Training Strategy: Developed a robust Base Model + Fine-tuning framework. This allows the system to handle general cases efficiently while adapting to specific, difficult plates (e.g., those from specific telescopes or with unique noise profiles) without requiring massive retraining.
• Metadata-Driven Adaptation: Introduced a hierarchical fine-tuning scheme based on available metadata (aperture, exposure time) to group similar plates, ensuring the model adapts to the specific Point Spread Function (PSF) and noise characteristics of different instruments.

4. Results

The proposed pipeline significantly increased the number of successfully registered plates:

• Base Model Performance:
  • Applied to 1,883 uncalibrated plates.
  • Successfully registered 1,083 plates (57.5% of the total).
  • Achieved classification accuracies ranging from 74% to 85% depending on the observatory and telescope.
• Fine-tuned Model Performance:
  • Applied to the remaining unregistered plates.
  • Successfully registered an additional 270 plates.
• Total Success:
  • 1,353 plates were successfully astrometrically registered out of the 1,883 viable candidates.
  • Overall Success Rate: 71.9%.
• Performance by Plate Quality:
  • Grade 1 (Intact): 82.0% success rate.
  • Grade 2 (Minor defects): 67.9% success rate.
  • Grade 3 (Moderate defects): 68.3% success rate.
  • Note: The model maintained high resilience even as defect coverage doubled from <25% to <50%, indicating the network effectively learns to ignore localized artifacts.

5. Significance and Future Work

• Scientific Value: This work unlocks a massive dataset of historical Chinese astronomical plates that were previously unusable for time-domain astronomy. These data are crucial for long-term studies of stellar variability, proper motions, and transient events.
• Efficiency: The AI-augmented pipeline automates the filtering of spurious sources, a bottleneck in historical plate processing.
• Limitations: Approximately 630 plates (28.1%) still failed. Analysis indicates these failures are due to intrinsic optical defects (defocus, mechanical drift streaks) or severe physical obliteration of source signals (handwriting, fingerprints) that fundamentally alter the photometric/morphological characteristics beyond the model's ability to recover.
• Future Directions: The authors suggest moving toward end-to-end object detection models (e.g., YOLO) to bypass the SExtractor step entirely, further streamlining the pipeline and potentially recovering sources in crowded or highly degraded fields.

Conclusion: The study demonstrates that deep learning, specifically Transformer-based architectures with adaptive fine-tuning, is a powerful tool for recovering scientific value from degraded historical astronomical archives, significantly outperforming traditional machine learning approaches.