AstroSURE: Learning to Remove Noise from Astronomical Images Without Ground Truth Data

1. Problem Statement

Astronomical imaging faces significant challenges due to low photon counts from faint, distant sources, resulting in images dominated by noise (photon shot noise, readout noise, dark current, and local contaminants like cosmic rays). This noise obscures faint structures and degrades the precision of scientific measurements, particularly object detection.

Traditional denoising approaches face two main hurdles:

1. Lack of Ground Truth: Supervised deep learning requires pairs of noisy and clean images. In astronomy, "clean" reference images do not exist; heavily stacked images are often used as proxies but are computationally expensive, introduce artifacts, and lack generalizability across instruments.
2. Single-Image Constraints: Many scientific scenarios (e.g., time-domain studies or specific calibration frames) only provide single exposures, making standard stacking techniques impossible.

The paper addresses the need for unsupervised or self-supervised denoising methods that can improve faint-source detectability in single or paired noisy astronomical images without requiring clean ground-truth data.

2. Methodology: The AstroSURE Framework

The authors propose AstroSURE, a practical framework evaluating denoising methods trained without clean targets. The workflow involves:

• Data Sources:
  • Synthetic Data: Generated using the Galsim library with a comprehensive noise model (Poisson photon statistics, Gaussian readout noise, dark current, and quantization noise).
  • Real Data: Observations from the Hubble Space Telescope (HST) (WFC3, space-based) and the Canada-France-Hawaii Telescope (CFHT) (MegaCam, ground-based).
• Network Architecture:
  • The primary model is a Modified U-Net, chosen for its balance of performance and computational efficiency compared to larger architectures like Restormer.
  • The architecture includes an encoder-decoder structure with skip connections to preserve fine details.
• Training Strategies (Loss Functions):
  The framework evaluates three distinct unsupervised/self-supervised learning paradigms:
  1. Noise2Noise (N2N): Trains on pairs of independent noisy observations of the same scene. The network learns to predict one noisy image from another, converging to the clean signal in expectation.
  2. Stein's Unbiased Risk Estimator (SURE): A single-image method that estimates the Mean Squared Error (MSE) risk analytically without a clean target. The authors implement SURE for Mixed Poisson-Gaussian noise, utilizing Monte Carlo approximations for the divergence term.
  3. Blind-Spot Methods: Includes Noise2Self (N2S) and Noise2Same (N2Sa), which mask target pixels during prediction to prevent identity mapping.
• Evaluation Metrics:
  Crucially, the paper prioritizes detection-oriented metrics over traditional image quality metrics (like PSNR alone):
  • Correct Detection Rate (CDR): Ratio of true objects correctly identified.
  • False Alarm Rate (FAR): Ratio of noise fluctuations misidentified as objects.
  • Auxiliary Metrics: Unsupervised PSNR (uPSNR), NIQE (Naturalness Image Quality Evaluator), and pixel distribution diagnostics (KL Divergence).

3. Key Contributions

• AstroSURE Pipeline: A systematic evaluation framework for unsupervised denoising in astronomy, specifically tailored to detection tasks rather than photometric precision.
• Comparative Analysis: A rigorous comparison of N2N, SURE, and blind-spot methods on both synthetic and real astronomical data.
• Domain Transfer Insights: An investigation into how models trained on simulated data transfer to real space-based (HST) and ground-based (CFHT) observations.
• Loss Function Optimization: Demonstration that L1 loss generally outperforms L2 loss in preserving faint astronomical features and reducing outliers in this context.

4. Key Results

A. Synthetic Data Performance

• N2N vs. Supervised: Noise2Noise training achieved performance nearly identical to supervised training (Noise2Clean) on synthetic data, validating its efficacy without ground truth.
• SURE Performance: SURE-based denoising was competitive with N2N in single-image settings, offering a viable solution when paired data is unavailable.
• Blind-Spot Limitations: Noise2Self and Noise2Same performed poorly. Noise2Same failed to train effectively due to the sparse nature of astronomical signals (small objects, dark backgrounds), while Noise2Self introduced grid-like artifacts and suppressed signal.
• Detection Gains: Denoised images significantly improved the Correct Detection Rate (CDR) and reduced the False Alarm Rate (FAR). For instance, at a 1% FAR threshold, denoised images achieved ~60% CDR compared to ~30% for raw noisy images.
• L1 vs. L2: Models trained with L1 loss consistently outperformed L2 loss in detection metrics and MAE, better preserving the integrity of faint sources.

B. Real Data Performance (HST vs. CFHT)

• HST (Space-Based): The method showed encouraging gains. After initializing with weights from space-based simulations and adapting via SURE, the CDR increased from 31.12% to 35.72%, and uPSNR improved significantly. This confirms that domain-consistent initialization (space-to-space) is effective.
• CFHT (Ground-Based): Performance gains were limited. The CDR increased only marginally (20.29% to 20.80%), and the FAR slightly increased.
  • Reasoning: The significant "domain gap" between the sharp, diffraction-limited space simulations and the seeing-limited, higher-background ground-based data hindered transfer. Pre-trained features did not generalize well to atmospheric distortions.

5. Significance and Future Work

• Scientific Impact: AstroSURE demonstrates that deep learning can enhance the detectability of faint astronomical sources using only noisy data, potentially reducing the number of exposures required for deep surveys.
• Practical Utility: The framework provides a viable alternative to stacking for single-exposure analysis and offers a method to improve the final products of stacking pipelines.
• Limitations & Future Directions:
  • Domain Adaptation: The study highlights the critical need for initialization weights to match the target instrument's characteristics (e.g., space vs. ground). Future work must address domain-specific retraining.
  • Structured Noise: Current methods focus on unstructured noise. Future iterations aim to handle structured contaminants (cosmic rays, satellite trails) by framing them as image inpainting problems.
  • Architecture: The authors suggest exploring Vision Transformers (ViT) to better capture complex spatial dependencies in future iterations.

In summary, AstroSURE successfully validates that unsupervised deep learning, particularly via Noise2Noise and SURE, can effectively denoise astronomical images to improve source detection, provided that the training domain closely matches the target observational conditions.