VisRec: A Semi-Supervised Approach to Visibility Data Reconstruction in Radio Astronomy

Ruoqi Wang, Haitao Wang, Qiong Luo, Feng Wang, Hejun WuAAAI 2025 · DOI

TL;DR

This paper introduces VisRec, a novel semi-supervised learning framework for reconstructing sparse and noisy visibility data from radio telescopes into high-fidelity images. By effectively combining supervised learning with data augmentation and unsupervised consistency training, VisRec significantly reduces the reliance on expensive labeled data and demonstrates superior robustness and generalizability compared to existing methods.

What Problem Does This Paper Address?

Radio astronomy relies on interferometers to observe celestial objects. These instruments produce "visibility data," which is essentially a sparse and noisy representation of the sky in the frequency domain. To create usable images, this visibility data must be reconstructed. Traditional methods like CLEAN are computationally intensive and can struggle with complex structures. More recent deep learning approaches have shown promise but require vast amounts of labeled data (ground truth images and their corresponding visibility data) for training. Obtaining this labeled data is a major bottleneck in radio astronomy due to the difficulty and expertise required for manual labeling and the inherent limitations of telescope observations in capturing complete data. This paper addresses the challenge of label scarcity by proposing a semi-supervised method that can leverage both limited labeled data and abundant unlabeled data to achieve high-quality reconstruction.

Methods

VisRec is a "model-agnostic" framework, meaning it can be applied to various underlying reconstruction models (like neural fields or CNNs). Its core innovation lies in its semi-supervised learning approach, which combines two key modules:

  • Supervised Module: This module uses the available labeled data. To enhance its performance, VisRec introduces a suite of data augmentation functions. These augmentations are categorized into:
  • Label-invariant augmentations: These modify the visibility data without altering the corresponding ground truth image (e.g., adding noise, shifting positions).
  • Label-variant augmentations: These apply transformations to both the visibility data and the ground truth image simultaneously to maintain consistency (e.g., geometric transformations in the uv-plane and corresponding image rotations/flips).
  • Unsupervised Module: This module leverages the abundant unlabeled visibility data. It works by:
  • Corruption Augmentation: Unlabeled visibility data is artificially corrupted using a predefined model (Tcorr) that simulates realistic observational impairments like noise, missing data, and antenna offsets.
  • Consistency Training: The model reconstructs both the original and the corrupted unlabeled visibility data. The consistency loss is then computed between these two reconstructions. This encourages the model to learn to reverse the effects of the corruption, making it more robust.

The overall training objective is to minimize a weighted sum of the supervised loss (from labeled data) and the consistency loss (from unlabeled data).

Datasets

The authors used real astronomical images from the Galaxy10 DECals dataset as a basis for generating synthetic visibility data. They employed the eht-imaging toolkit to simulate observations mimicking an 8-telescope Event Horizon Telescope (EHT) setup and also used Very Long Baseline Array (VLBA) uv-coverage.

For evaluation, they created datasets with varying proportions of labeled and unlabeled data. A small subset of the training data was randomly selected as labeled, while the rest was used as unlabeled.

Key Insights and Ideas

  • Leveraging Unlabeled Data is Crucial: The paper strongly demonstrates that effectively utilizing unlabeled data through consistency training significantly boosts performance, especially in label-scarce scenarios.
  • Data Augmentation is Key for Robustness: Both label-invariant and label-variant augmentations play a vital role in improving the model's ability to handle noisy and imperfect data.
  • Model Agnosticism is a Strength: VisRec's ability to integrate with different underlying reconstruction models makes it a versatile framework applicable to various architectures.
  • Consistency Training Enhances Robustness: By training the model to be invariant to simulated corruptions, VisRec inherently improves its resilience to real-world observational imperfections.

Key Results

  • Superior Reconstruction Quality: VisRec consistently outperformed baseline methods (including supervised learning with augmentation) in terms of LFD, PSNR, and SSIM metrics on both EHT and VLBA datasets.
  • Label Efficiency: With as few as 64 labeled examples, VisRec significantly outperformed traditional methods and performed comparably to supervised methods trained with over 16,000 labeled examples when testing data was ideal.
  • Robustness to Noise and Missing Data: VisRec showed significantly less performance degradation than supervised methods when faced with increasing noise levels or missing visibility points. For instance, when testing data had a 50% sample loss rate, VisRec's performance drop was substantially smaller than that of supervised methods.
  • Generalizability: VisRec demonstrated better generalizability to unseen telescope configurations compared to traditional supervised methods.

Limitations and Future Work

  • Limitations: While VisRec shows strong performance, the paper notes that in some specific metrics, variants like "Supervised + Aug" and "VisRec w/o Sup-Aug" might show marginal improvements. The impact of the consistency loss weight ($\lambda$) was found to be small across a range of values, suggesting it might not require extensive tuning, but this could also imply a limited dynamic range of its influence.
  • Future Work: The authors suggest extending VisRec to other domains beyond radio astronomy, such as magnetic resonance imaging, seismic imaging, and point cloud data analysis. They acknowledge that direct application might not be feasible, and further research will be needed to adapt the training approach and augmentations for these new fields.

Summary updated

BibTeX
@inproceedings{wang2025visrec,
  author = {Ruoqi Wang and Haitao Wang and Qiong Luo and Feng Wang and Hejun Wu},
  title = {{VisRec: A Semi-Supervised Approach to Visibility Data Reconstruction in Radio Astronomy}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i1.32069},
}
Figure