Source-linked AI summary
Compact continuum source-finding for next generation radio surveys
Paul J Hancock, Tara Murphy, Bryan M Gaensler, Andrew Hopkins, James R Curran
TL;DR
Next-generation radio surveys require automated source finding that remains complete and reliable despite rare failures becoming numerous in large, high-cadence datasets. The paper analyzes four established packages and ASKAP’s program, identifies algorithmic failure modes, and demonstrates Aegean, which produces highly complete and reliable catalogues while accurately characterising multi-component islands. Its conclusions are scoped to Stokes I continuum source finding and assume structured backgrounds have been pre-processed.
Problem
Rare failures in otherwise highly complete and reliable source finders become consequential for next-generation surveys producing many thousands of sources daily, especially blind transient surveys.
Method
The paper evaluates widely used source-finding packages and Selavy, analyzes their failure modes, and develops Aegean using curvature-based parameter estimation and constraints for multiple-Gaussian fitting.
Results
Aegean achieves 5σ completeness better than the estimated ideal source finder without sacrificing reliability and is the most reliable tested algorithm.
Takeaways & Limitations
Adopting Aegean could support more complete, reliable, and accurate radio-survey catalogues without significant manual intervention.
Takeaways & Limitations
The study focuses on Stokes I continuum emission and assumes images have been pre-processed to remove structured background emission.
Abstract
from arXiv · showhide
We present a detailed analysis of four of the most widely used radio source finding packages in radio astronomy, and a program being developed for the Australian Square Kilometer Array Pathfinder (ASKAP) telescope. The four packages; SExtractor, SFind, IMSAD and Selavy are shown to produce source catalogues with high completeness and reliability. In this paper we analyse the small fraction (~1%) of cases in which these packages do not perform well. This small fraction of sources will be of concern for the next generation of radio surveys which will produce many thousands of sources on a daily basis, in particular for blind radio transients surveys. From our analysis we identify the ways in which the underlying source finding algorithms fail. We demonstrate a new source finding algorithm Aegean, based on the application of a Laplacian kernel, which can avoid these problems and can produce complete and reliable source catalogues for the next generation of radio surveys.
1 INTRODUCTION
Radio source finding aims to build complete, reliable catalogues, but next-generation surveys make even rare failures consequential. The paper analyzes these failures and uses them to develop an improved source-finding algorithm while defining its scope.
- Large-scale radio surveys seek an unbiased census of the sky through catalogues that are both complete and reliable.Completeness means finding all sources; reliability means that detected and extracted sources are real.
- Standard source finders miss only a small fraction of sources, with problem cases often handled through preprocessing, post-processing, or manual correction.
- Next-generation surveys will produce much more data at high cadence, leaving limited scope for manual intervention and correction.
- At 99% completeness and reliability, a VAST-like survey could generate approximately 10,000 false sources and miss approximately 10,000 real sources per day.
- The paper finds that source-finding failures arise from both random noise-related effects and deterministic effects related to the underlying algorithms.
- The analysis identifies improvements and uses them to build an algorithm intended to produce more complete and reliable catalogues than existing approaches.
- The study focuses on Stokes I continuum emission and does not directly address spectral-line, polarization, extended-source, or diffuse-emission complications.
2 SOURCE-FINDING IN RADIO ASTRONOMY
Radio source finding separates sources from background, identifies connected pixel islands, characterises their sources, and records the fitted parameters in catalogues. The main technical difficulty is accurately fitting multiple Gaussian components under structured-background and initialization constraints.
- Most radio source-finding pipelines estimate and subtract the background, identify source pixels, characterise sources, and catalogue their parameters.
- Background estimation and subtraction: Thresholding separates source pixels from background, with the threshold balancing real-source detection against false detections; blind surveys typically use 5σ.
- Background estimation and subtraction: Structured backgrounds require image filtering before source finding, and this paper assumes the images have already been pre-processed to remove such structure.
- Source identification: Source identification groups above-threshold pixels into contiguous islands, each representing one or more sources, but island lists alone are not generally useful catalogues.
- Source characterisation: Source characterisation measures properties such as flux and angular size, using single or multiple Gaussian components according to source morphology and survey resolution.
- Source characterisation: Multiple-Gaussian fitting is sensitive to initial parameters and local minima, so effective fitting requires intelligent starting estimates and sensible parameter constraints.
- Source characterisation: De-blending and iterative fitting are two approaches for obtaining initial parameters for multiple-Gaussian characterisation.
- Cataloguing: Catalogues should retain fitted parameters, uncertainties, and warnings about inadequate characterisation so poorly measured sources remain available for remeasurement rather than being excluded.
3 SOURCE-FINDING PACKAGES AND THEIR ALGORITHMS
The paper reviews common radio source-finding algorithms and packages, then describes FloodFill and Aegean as approaches for identifying and characterising compact continuum sources. It also documents implementation choices and known limitations in several packages.
- Common approaches: Most major radio source-finding packages use common stages for background estimation, source identification, characterisation, and cataloguing.Wavelet-based packages were excluded because none of the most widely used packages considered relies on wavelet analysis.
- SExtractor: SExtractor detects peaks above 5σ, characterises islands above 3σ containing at least five pixels, and uses a large background-estimation box.For the simulated images, its source-characterisation results were insensitive to the DEBLEND_MINCONT parameters.
- imsad: imsad uses a user-specified flux or SNR threshold, estimates background noise from pixel histograms, and performs one Gaussian fit per island.The analysis used a threshold of 5σ = 125µJy/beam.
- Selavy: Selavy is an ASKAPsoft package for spectral cubes and continuum images that can parameterise islands with multiple Gaussian components.In compact continuum source finding, its stated difference from Duchamp is the ability to fit multiple Gaussian components to an island.
- sfind: sfind estimates a varying background by measuring the mean and rms in user-defined sub-regions, but source contamination can distort boundary-source shapes and flux distributions.Using sub-regions larger than the image can largely remove these contamination effects for images with zero mean and constant rms.
- FloodFill: FloodFill grows contiguous islands from seed pixels above σs using adjacent pixels above the lower flood threshold σf.Its output is a disjoint list of islands, and its operation is invariant to the order in which seed pixels are selected.
- Aegean: Aegean uses FloodFill to identify islands, then uses a curvature map and constrained Gaussian fitting to determine and fit multiple components.BLOBCAT instead characterises islands without assuming a particular source structure, whereas Aegean assumes compact sources.
- Catalogue construction: Source-catalogue construction requires more than running a source finder because observing strategy, hardware, and data reduction vary across surveys.The supplied passage describes standard packages as starting points for catalogue creation rather than complete catalogue-production workflows.
4 TEST DATA
The evaluation uses simulated images generated from a controlled master source catalogue, allowing the input sources and image properties to be known exactly. The simulations specify source fluxes, positions, morphologies, beam, noise, and image sampling.
- Simulation design: The study evaluates the source-finding algorithms on simulated data whose image properties and input catalogue are controlled.The controlled input enables recovered sources to be matched against a known true source list.
- Source population: The master catalogue uses peak fluxes distributed as N(S) ∝ S^-2.3 over 25 µJy to 10 Jy.This is the stated flux constraint for the simulated source population.
- Source population: Sources are randomly positioned within ten sky regions, with source clustering excluded from the simulations.The positions were generated within regions similar to the study’s example sky field.
- Source population: Source major and minor axes are randomly distributed from 0 to 52′′, with position angles spanning −90° to +90°.These are the stated morphology constraints used to construct the master catalogue.
- Image construction: The simulated image has a 30′′ synthesised beam, 25 µJy/beam Gaussian rms noise, 4801×4801 pixels, and 6′′ pixels.The resulting beam sampling is five pixels per beam, and regions outside the catalogue contain noise but no sources.
5 SOURCE-FINDING EVALUATION
The evaluation compares source-finder catalogues against controlled simulated inputs using completeness, reliability, flux distributions, and parameter accuracy. Most packages perform well overall, but algorithm-specific failures—including Selavy fragmentation and IMSAD island handling—explain important residual errors.
- Evaluation setup: Each package was run at a 5σ threshold and evaluated against the input catalogue using completeness, reliability, and flux-distribution metrics.Cross-identification used positional and flux information, with a 30′′ matching radius and simulated-image ground truth.
- Flux distribution: Except for Selavy, catalogues have flux distributions consistent with the input catalogue; Selavy reports around twice as many sources because of source fragmentation.The excess occurs across flux levels and indicates that fragmented components are being counted separately.
- Completeness: Above SNR 6, all source finders exceed 99% completeness, while below SNR 7 sfind is less complete because it prioritizes minimizing false detections.Selavy and Aegean are the most complete across SNRs; Aegean maintains high completeness without sacrificing reliability.
- False detection rate: SExtractor, IMSAD, and Selavy have higher false-detection rates than sfind and Aegean because they cannot accurately characterize pixel islands.Selavy fragments low-SNR sources into three or four components, whereas IMSAD does not split islands containing multiple sources; all IMSAD false detections above SNR 20 arise from this flaw.
- Measured parameter correctness: Position accuracy generally improves with flux and agrees with ideal Gaussian-fitting expectations, with high-SNR deviations partly caused by package reporting precision.IMSAD’s 0.1′′ reporting resolution limits its measured median absolute deviation in position.
- Measured parameter correctness: Flux measurements generally agree with ideal Gaussian fits, except for SExtractor above SNR 50 and Selavy below SNR 50, where fragmentation reduces flux accuracy.Selavy’s flux accuracy is about half of ideal because each fragmented component contains only part of the true flux.
- Initial evaluation summary: All packages produce robust catalogues, but Aegean matches or exceeds the others in completeness without sacrificing reliability, making small residual errors consequential for very large surveys.For EMU’s expected 70 million sources, a 1% false-detection rate would yield 700,000 false sources.
6 MISSED SOURCES
The paper identifies two main populations of missed sources: isolated faint sources affected by noise and multiple components within one pixel island. Algorithmic failures in characterising blended islands can also generate false detections and impair transient searches.
- Missed-source populations: Two populations are missed: isolated faint sources and sources inside islands containing multiple components.The first population is noise-affected; the second arises when algorithms cannot correctly characterise multiple sources within one island.
- 6.1 Isolated faint sources: Noise can hide sources whose true flux exceeds the threshold but whose pixels rarely exceed it.Only lowering the detection threshold, increasing image sensitivity, or accepting more false detections can recover all sources above a flux limit.
- 6.2.1 Iterative fitting: Iterative fitting can fragment faint single sources when the allowed component count is poorly chosen, making their overall flux difficult to recover.The flux of a source is not simply the sum of the fluxes of its fitted fragments.
- 6.2.1 Iterative fitting: Flux-ratio extremes cause iterative fitting failures because residual structure is dominated by FWHM errors rather than amplitude errors.Even highly accurate flux and major-axis estimates can leave residual peaks above SNR 500 in the 9Jy and 1.7mJy example.
- 6.2.1 Iterative fitting: At an SNR as low as 11, fitting residuals can exceed 5σ, causing iterative methods to miss 5σ sources near a source of ⩾11σ.Unconstrained fitting can also drive a faint component toward a residual maximum.
- 6.2.2 Sectioning or de-blending: Missed sources within multi-component islands are difficult to correct statistically and can compromise reliable detection and characterisation of variable events.Such a correction would require the algorithm, source-flux distribution, and flux-dependent two-point correlation function.
7 THE NEW SOURCE FINDING PROGRAM: Aegean
Aegean is a new source finder designed to retain existing completeness and reliability while improving the characterisation of multi-component pixel islands. It uses curvature-derived summits to determine and jointly fit Gaussian components with constraints that prevent faint components from migrating into bright-source residuals.
- 7 THE NEW SOURCE FINDING PROGRAM: Aegean: Aegean was created to preserve existing completeness and reliability while improving multi-component island characterisation.The program was developed from analysis of how earlier algorithms affect source finding and characterisation.
- Curvature analysis: Aegean uses a single curvature map, where negative curvature identifies local maxima and helps characterise pixel islands.The curvature is derived from image derivatives measured in four directions and combined into mean curvature.
- Component estimation: Curvature and clipping masks divide islands into summits, which determine the number of Gaussian components and their initial parameters.Summits are contiguous regions with negative curvature and flux above the threshold.
- Component fitting: Aegean fits one Gaussian component per summit, fitting components in an island simultaneously.The workflow constructs a curvature map, estimates curvature noise, flood-fills islands, and jointly fits the resulting components.
- Demonstration: In the Figure 12 example, two summits enable Aegean to detect and characterise both components, while single-source islands receive one component.This summit-based behaviour links the curvature analysis directly to the fitted component count.
- Fitting constraints: Component positions are constrained within their corresponding summits, preventing faint components from migrating toward bright-source fitting residuals.Component fluxes must exceed 5σ, while size constraints use summit dimensions and the synthesised beam.
8 CONCLUSIONS
The analysis identifies faint near-threshold sources and multi-component pixel islands as the main difficult cases for source-finding packages. Aegean uses curvature-based constrained fitting to improve component characterization and catalogue completeness and reliability.
- Performance and failure modes: Source-finding packages detect isolated compact sources reliably but struggle with faint sources near the detection limit and islands containing multiple components.These are the two populations identified as poorly detected in the simulated-data assessment.
- Performance and failure modes: Iterative fitting approaches are more likely to mis-characterise multi-component islands than methods that estimate component counts before fitting.The comparison concerns how packages determine the number of components within an island before characterization.
- Aegean: Aegean derives a curvature image with a Laplacian transform to determine compact-component counts and initialize constrained multiple-Gaussian fits.The curvature image supplies initial parameters and limits for fitting multiple elliptical Gaussians.
- Aegean: Aegean achieves 5σ completeness better than the estimated ideal source finder without sacrificing reliability and is the most reliable tested algorithm.It also accurately characterizes islands containing multiple compact components.
- Conclusion: Constrained multiple-Gaussian fitting, accurate initial parameters, and sensible constraints are necessary to improve catalogue reliability and completeness.The authors anticipate that Aegean can support more complete, reliable, and accurate catalogues with limited manual intervention.