Distribution of the Högbom CLEAN Algorithm Using Tiled Images with Feedback

Daniel G. Wright, Karel Adámek, Wesley ArmourarXiv 2024 · DOI · arXiv

Abstract

Data sizes for next generation radio telescopes, such as the Square Kilometre Array (SKA), are far above that of their predecessors. The CLEAN algorithm was originally developed by Högbom [1974], long before such data sizes were thought possible and is still the most popular tool used for deconvolution in interferometric imaging. In order to facilitate these new large data sizes and reduce computation time a distributed approach to the algorithm has been investigated. The serial nature of the CLEAN algorithm, due to its matching pursuit design, makes this challenging. Splitting the image into a number of tiles which can be individually deconvolved has been investigated, but this creates discontinuities in the deconvolved image and makes it difficult to deconvolve faint sources in the presence of a point spread function associated with bright sources in other tiles. A method of feedback between each of the tiles has been developed to deal with these problems. This new approach has been tested on a simulated dataset containing multiple point sources of known intensity. When compared to a standard Högbom deconvolution the tiled feedback version produced a reconstructed image, containing sources up to 2.1 Jy, which agreed to between -0.1 Jy and +0.04 Jy of the standard method across the whole deconvolved image at a speed up to 10.66 times faster.

Introduction

In radio astronomy, telescopes do not capture traditional photographs; instead, they sample signals across sparse arrays, resulting in a "dirty image" cluttered with artifacts. These artifacts are caused by the telescope’s Point Spread Function (PSF), where bright sources are surrounded by complex interference fringes (sidelobes) that obscure fainter objects. The standard tool for addressing this is the Högbom CLEAN algorithm, a matching pursuit technique that iteratively identifies and subtracts the brightest point sources. However, CLEAN is inherently sequential, requiring the identification of the global maximum in each iteration. For next-generation instruments like the Square Kilometre Array (SKA), this serial bottleneck is a major hurdle for processing unprecedented data volumes.

Methodology: Distributed Tiling with Feedback

To address the serial bottleneck, researchers have developed a distributed approach that partitions the dirty image into a grid of tiles. While simple tiling typically fails because sidelobes from one tile "leak" into others—creating ghost artifacts—this method introduces a critical synchronization step known as cross-coupling feedback.

  1. Tiling: The image is split into a grid (e.g., 4 to 256 tiles), with each tile assigned to a separate worker node.
  2. Parallel Deconvolution: Workers independently run the CLEAN algorithm on their assigned tiles until a specific flux threshold is met.
  3. Cross-Coupling Feedback: This is the core innovation. Workers convolve their identified CLEAN components with the full PSF to create a cross-coupling image, showing how subtractions in one tile affect the rest of the field.
  4. Synchronization: These cross-coupling effects are shared and subtracted from the residual images of all other tiles, ensuring global consistency across the entire image.
The synchronization workflow for the distributed CLEAN algorithm, showing parallel processing interleaved with feedback steps.

Experimental Evaluation

The method was evaluated using a simulated point source dataset featuring intensities up to 2.1 Jy. The researchers utilized the unique and challenging PSF of the Very Large Array (VLA) in its "D" configuration. Performance benchmarks were conducted using 8 Dask workers on an AMD Ryzen 9 processor to compare the distributed method against the standard serial Högbom CLEAN.

Results and Performance

The distributed approach achieved a significant speedup of up to 10.66 times compared to the standard serial method. Critically, this speedup did not come at the cost of image fidelity. The difference between the distributed and standard methods remained between -0.1 Jy and +0.04 Jy across the entire reconstructed image. The study also noted that scalability generally improves with the number of tiles, provided the tiles do not become smaller than the sources or significant PSF features.

Comparison of the parallel skymodel with feedback against the standard method, including a difference image and histogram showing high fidelity.

Conclusion and Future Work

This research provides a vital proof-of-concept for the SKA era, demonstrating that one of astronomy's most trusted algorithms can be adapted for distributed computing. While the current study focused on point sources, future work will address the deconvolution of complex extended sources and the scaling of the algorithm across multi-node clusters.

Summary updated

Abstract

Data sizes for next generation radio telescopes, such as the Square Kilometre Array (SKA), are far above that of their predecessors. The CLEAN algorithm was originally developed by Högbom [1974], long before such data sizes were thought possible and is still the most popular tool used for deconvolution in interferometric imaging. In order to facilitate these new large data sizes and reduce computation time a distributed approach to the algorithm has been investigated. The serial nature of the CLEAN algorithm, due to its matching pursuit design, makes this challenging. Splitting the image into a number of tiles which can be individually deconvolved has been investigated, but this creates discontinuities in the deconvolved image and makes it difficult to deconvolve faint sources in the presence of a point spread function associated with bright sources in other tiles. A method of feedback between each of the tiles has been developed to deal with these problems. This new approach has been tested on a simulated dataset containing multiple point sources of known intensity. When compared to a standard Högbom deconvolution the tiled feedback version produced a reconstructed image, containing sources up to 2.1 Jy, which agreed to between -0.1 Jy and +0.04 Jy of the standard method across the whole deconvolved image at a speed up to 10.66 times faster.
BibTeX
@misc{wright2024distribution,
  author = {Daniel G. Wright and Karel Adámek and Wesley Armour},
  title = {{Distribution of the Högbom CLEAN Algorithm Using Tiled Images with Feedback}},
  year = {2024},
  eprint = {2501.01461},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.01461},
}
Figure