ProFound: Source Extraction & Photometry

• ProFound is an image analysis package that uses iterative dilation and watershed deblending to accurately extract sources from astronomical images.
• It performs photometry on dilated segmentation maps, ensuring robust sky subtraction and reliable initial parameter estimates for galaxy profiling.
• It adapts to optical and radio continuum data, effectively deblending complex morphologies and assigning flux in crowded or diffuse regions.

ProFound is a source finding and image analysis package introduced for modern, deep, wide-area imaging surveys, with the specific aim of producing source extraction, photometry, segmentation maps, sky and variance maps, and initial parameter estimates that can directly seed 2D galaxy profiling with ProFit (Robotham et al., 2018). Its defining methodological choice is to perform photometry on iteratively dilated segmentation maps that contain nearly all identifiable flux, rather than on circular or elliptical apertures, while deblending is carried out across saddle points in flux using a non-discretised watershed algorithm (Robotham et al., 2018). The package was later adapted to radio continuum source extraction, where its pixel-based treatment of irregular emission was compared with Gaussian-component finders (Hale et al., 2019).

1. Origins and design objectives

ProFound was created to address a specific limitation of aperture-based photometry in contemporary survey pipelines. Traditional methods such as Petrosian and Kron magnitudes, and SExtractor’s AUTO photometry, impose simple circular or elliptical geometries; these are efficient, but they can misrepresent the complex shapes of resolved galaxies, complicate flux sharing when apertures overlap, and can fail pathologically in crowded regions and near bright stars, including cases where apertures loop around saturation halos (Robotham et al., 2018). ProFound therefore targets robust, reproducible source extraction for images with substantial sky area, emphasizing robust sky subtraction, pixel-matched sigma maps, non-overlapping segmentation maps, and initial parameter estimates for effective radius, axis ratio, position angle, and centroid (Robotham et al., 2018).

The package was designed to work in unison with ProFit, so that source detection and photometric measurement can feed semi-automatic galaxy profiling for large samples (Robotham et al., 2018). A number of bulge-disc decomposition projects were already using the ProFound–ProFit pipeline at publication time, and the software was released as an open-source R package under an LGPL-3 license on GitHub (Robotham et al., 2018).

A common misunderstanding is to treat ProFound as an ellipse-based photometry code with additional diagnostics. That is inaccurate. ProFound does compute elliptical approximations from segments, including quantities such as the $R100$ semi-major axis, but these ellipses are not used for photometry; they serve as diagnostics and as seeds for downstream tools such as ProFit (Robotham et al., 2018).

2. Segmentation, watershed deblending, and iterative dilation

The highest-level function, also named ProFound, orchestrates a multi-stage pipeline consisting of rough sky estimation (MakeSkyGrid), initial threshold-based segmentation (MakeSegim), improved sky estimation, basic photometry (SegimStats), iterative dilation (MakeSegimDilate), convergence selection (selectCoG), final segmentation-map assembly, conservative masking for sky estimation, final sky estimation, and final comprehensive photometry (Robotham et al., 2018). The returned image products include segim, segim_orig, objects, objects_redo, sky, skyRMS, and SBlim, while the catalogue segstats includes flux, magnitude, size metrics, shape parameters, centroid, flags, and related diagnostics (Robotham et al., 2018).

Initial segmentation uses a threshold defined relative to the local sky RMS, typically written as

$T = \mathrm{sky} + n_\sigma \,\sigma .$

Default operation with skycut=1 therefore corresponds to a $1\sigma$ threshold (Robotham et al., 2018). Deblending is performed by a non-discretised watershed based on EBImage’s C implementation. Segments are grown from peaks while respecting saddle points in the 2D flux landscape, with three exposed controls: tolerance, the uphill growth tolerance in units of sky RMS; sigma, the Gaussian smoothing scale; and ext, the local search radius (Robotham et al., 2018). The default parameterization is tolerance=4, sigma=1 pixel, and ext=2 pixels (Robotham et al., 2018).

Conceptually, the watershed procedure identifies the brightest unassigned pixel above threshold, starts a new segment, grows that segment by adding neighboring pixels that are not significantly uphill compared with assigned pixels, stops when no further growth is possible, and then repeats the process for the next brightest unassigned pixel (Robotham et al., 2018). Because boundaries follow saddle points, segments are intended to reflect the water-flow basins of the inverted flux map; no islands can form within larger segments (Robotham et al., 2018). This differs from discretised watershed schemes and from 1D flux-cut heuristics such as those associated with SExtractor (Robotham et al., 2018).

After initial segmentation, ProFound performs photometry on iteratively dilated segments. The default dilation kernel is a circular top-hat of diameter 9 pixels, with up to 6 iterations (Robotham et al., 2018). If dilations overlap, contested pixels are assigned to the segment with more flux in the current iteration; per-object convergence is declared when the flux change relative to the previous iteration is within 5%, and the final segmentation map is assembled object by object at the converged iteration (Robotham et al., 2018). Empirically, stars typically converge in about 2 dilations, galaxies in about 4, and very extended low-surface-brightness galaxies can require the default maximum of 6 (Robotham et al., 2018). At the faintest limit, dilation can add up to about 60% of the total flux before noise halts growth (Robotham et al., 2018).

3. Photometric formalism, sky modelling, and catalogue products

Photometry in ProFound is defined directly on the segment footprint. For a segment $S$, the total flux is

$F = \sum_{i \in S} \left(I_i - B_i\right),$

where $I_i$ is the pixel intensity and $B_i$ is the estimated background from the final sky map (Robotham et al., 2018). The corresponding signal-to-noise ratio is

$$\mathrm{S/N} = \frac{F}{\sigma_F}, \qquad \sigma_F^2 = \sum_{i \in S} \sigma_i^2,$$

with $\sigma_i$ obtained from the pixel-matched skyRMS map; if gain is known, shot noise and read noise can be folded into $\sigma_i$ (Robotham et al., 2018).

The catalogue records total flux and magnitude, sizes $T = \mathrm{sky} + n_\sigma \,\sigma .$0, $T = \mathrm{sky} + n_\sigma \,\sigma .$1, and $T = \mathrm{sky} + n_\sigma \,\sigma .$2, mean surface brightnesses within those regions, axis ratio, position angle, centroid, concentration, flags, and quality diagnostics (Robotham et al., 2018). Shape summaries are based on second moments and ellipse approximations, with

$T = \mathrm{sky} + n_\sigma \,\sigma .$3

for semi-major and semi-minor axes $T = \mathrm{sky} + n_\sigma \,\sigma .$4 and $T = \mathrm{sky} + n_\sigma \,\sigma .$5 (Robotham et al., 2018). Mean surface brightness inside the $T = \mathrm{sky} + n_\sigma \,\sigma .$6 flux region is reported as

$T = \mathrm{sky} + n_\sigma \,\sigma .$7

where $T = \mathrm{sky} + n_\sigma \,\sigma .$8 is the magnitude within that region and $T = \mathrm{sky} + n_\sigma \,\sigma .$9 is the area in arcsec$1\sigma$0 (Robotham et al., 2018). For blended systems, ProFound can also report flux_reflect, a symmetry-based flux reconstruction obtained by rotating the segment about its centre and estimating missing flux where the mirrored pixel lies outside the segment; the differences are typically modest, about $1\sigma$1 mag on VIKING-like data (Robotham et al., 2018).

Sky estimation is a central part of the design. MakeSkyGrid divides the image into box-car regions, removes masked pixels and objects, applies iterative clipping within each region, and interpolates the resulting sky and skyRMS images via bilinear or bicubic interpolation, with bilinear interpolation noted as safer when sky varies rapidly (Robotham et al., 2018). The full pipeline uses three passes of sky estimation with increasingly aggressive masks (Robotham et al., 2018). This repeated sky–segmentation coupling is intended to stabilize both the background model and the final segment footprints.

These outputs integrate directly with ProFit. Effective radius, axis ratio, position angle, centroid, and total flux provide initial guesses for Sérsic or Moffat components; segmentation maps define which pixels enter the likelihood; and sky and sigma maps are pixel-matched to the model noise (Robotham et al., 2018). Demonstration on VIKING-like data showed that ProFound and ProFit fluxes agree within about $1\sigma$2 mag typically, at worst within $1\sigma$3 mag, that total modelled flux is within about 1% of ProFound, and that $1\sigma$4 increases by about 5% for extended sources after fitting (Robotham et al., 2018).

4. Optical and near-infrared survey performance

The original evaluation included both simulated and real survey data. In VIKING-like simulations, the test set comprised 100 frames of $1\sigma$5 pixels, each containing 200 stars and 200 galaxies with magnitudes 15–23, a PSF FWHM of 5 pixels, and galaxies spanning Sérsic indices $1\sigma$6–4, axial ratios 0.3–1, and boxiness $1\sigma$7 (Robotham et al., 2018). With skycut=1, completeness remained high until near the $1\sigma$8 surface-brightness limit while false positives remained far below true positives; pushing skycut below about 0.8 increased spurious detections (Robotham et al., 2018).

In the same simulations, stars showed unbiased fluxes and slightly underestimated sizes at the faint limit because dilation halted on noise, whereas galaxies showed a small negative flux bias that increased toward the faint end (Robotham et al., 2018). Even so, ProFound recovered more flux than SExtractor’s AUTO across magnitudes (Robotham et al., 2018). Size estimates remained within a factor of two for the running median, and 99% of stars and about 70% of galaxies lay within a factor of two in flux or size, which the paper judged adequate for ProFit initialization (Robotham et al., 2018).

On LSST Deep-32/36 simulations, ProFound recovered 1548/1571 good sources with 91.3%/88.3% true matches under a 3D match criterion in position and flux, while SExtractor recovered 1441/1386 with 86.5%/85.2% true matches (Robotham et al., 2018). Restricting ProFound to the brightest approximately 1433/1375 sources increased the true-positive rates to 93.8%/91.3% (Robotham et al., 2018).

Comparisons on UltraVISTA data showed that ProFound’s dilated segments qualitatively agreed with SExtractor AUTO apertures in isolated regions, while avoiding mis-groupings in crowded fields and detecting additional faint sources (Robotham et al., 2018). ProFound magnitudes were typically brighter by less than 0.2 mag up to about 23 mag, with many deviations attributed to improved deblending of close stars that AUTO merged (Robotham et al., 2018). In colour photometry, matched-segment colours derived from pre-dilation segments at about a $1\sigma$9 threshold were tighter than UltraVISTA 2-arcsec aperture colours, indicating reduced aperture-mismatch noise (Robotham et al., 2018). The same analysis also used ProFound mean surface brightness and size, together with NIR colours, for star–galaxy separation (Robotham et al., 2018).

5. Adaptation to radio continuum source extraction

A later study investigated ProFound as an alternative to Gaussian-component fitting for radio continuum source extraction, especially for complex, extended morphologies that are increasingly important for interferometers rich in short baselines (Hale et al., 2019). In this setting, ProFound retains its segmentation-based logic: it constructs an initial smooth sky model, seeds segments from pixels above threshold, refines the sky after source removal, measures fluxes and shape statistics, and dilates segments until flux convergence (Hale et al., 2019). The photometry is still pixel-based and isophotal rather than Gaussian-model based.

Radio images, however, are in Jy/beam, so the summed segment flux must be converted to integrated flux density via a beam-area correction (Hale et al., 2019). The study also applied a PSF-based correction for faint unresolved sources whose threshold-limited segments do not fully sample the synthesized beam, with correction factors typically near 1.0 and rising to about 1.2 for faint unresolved sources (Hale et al., 2019). For the VLA XMM-LSS mosaic, the inverted-image analysis indicated that skycut=3.5 yielded about 98% real detections (Hale et al., 2019).

The radio evaluation reported flux recovery for simulated Gaussian sources comparable to PyBDSF and AEGEAN, with median recovered-to-injected flux ratios of about 1.02, 1.06, and 1.01 respectively (Hale et al., 2019). For extended sources re-injected from the real VLA image, ProFound yielded $S$0, whereas AEGEAN underestimated the flux at about 0.76 with a large negative tail, and PyBDSF with atrous_do was closer to unity at about 0.97 but showed much larger scatter (Hale et al., 2019). Residual analyses further showed that ProFound produced noise-like residuals with minimal positive or negative tails, whereas PyBDSF and AEGEAN often left significant positive residuals and, in multiscale settings, negative residuals (Hale et al., 2019).

The same work emphasized a multi-wavelength advantage: segments defined in one band can be reused for forced photometry in radio maps, supporting cross-band catalogues with consistent apertures and deblending (Hale et al., 2019). A plausible implication is that ProFound’s segmentation formalism generalizes naturally when morphology is more informative than any single parametric component model.

6. Comparative assessments, scaling behaviour, and limitations

Hydra II provided a later comparison of ProFound with Aegean, Caesar, PyBDSF, and Selavy on a $S$1 deg cutout from the EMU Phase 1 Pilot Survey (Boyce et al., 2023). In that study, ProFound was characterized as a segmentation-based finder originally developed for optical imaging and adapted to radio data; it identified contiguous segments using $S$2-clipped background/noise estimation and reported total flux densities via pixel sums rather than Gaussian fits (Boyce et al., 2023). Hydra optimized thresholds to a 90% PRD cutoff; for the EMU deep image ProFound used $S$3 and $S$4, while for the shallow image it used $S$5 and $S$6 (Boyce et al., 2023).

The resulting trade-offs were explicit. ProFound reported the largest number of deep detections, $S$7, compared with 8,538 for Aegean, 8,292 for PyBDSF, 7,838 for Caesar, and 5,880 for Selavy, but it also had the lowest shallow-to-deep recovery rate on EMU, 6.8% (Boyce et al., 2023). The paper interpreted this as a mixture of extra real detections, splitting into multiple components, and spurious detections near threshold (Boyce et al., 2023). No systematic flux bias was seen in flux-ratio plots, but ProFound showed the largest scatter among the five finders: its EMU $S$8 scatter averaged $S$9, higher than the other methods (Boyce et al., 2023). ProFound also showed an excess of very small sizes below the beam, attributed partly to noise spikes and partly to faint sources for which only a few pixels exceeded threshold (Boyce et al., 2023).

These radio-comparison results qualify, rather than overturn, the original optical design claims. ProFound remains strong for extended and diffuse morphology, total-flux measurement of complex systems, and edge-truncated sources, but it is more vulnerable near the detection threshold, where segment-bound sizes can shrink and low-SNR neighbors can blend into single components (Boyce et al., 2023). Hydra II therefore recommended a hybrid strategy for EMU-scale catalogues: use a Gaussian-fit finder such as Aegean for compact-source characterization and residual generation, then apply ProFound or Caesar to the residual images to capture extended or diffuse structures (Boyce et al., 2023).

From a computational perspective, the original package reported runtime dominated by watershed segmentation at about 50%, followed by sky maps at about 20%, dilation at about 20%, and photometry at about 10% (Robotham et al., 2018). Scaling is roughly linear with pixel count until RAM limits are reached; machines with about 8–16 GB can handle roughly 10k–20k square images in-core (Robotham et al., 2018). The main function cannot handle images with more than $F = \sum_{i \in S} \left(I_i - B_i\right),$0 pixels, so ProFoundLarge processes very large FITS files by tiling them with overlaps and recombining them, and is embarrassingly parallel (Robotham et al., 2018). The implementation is single-threaded by default, and optimization gains are modest because the watershed is already C-based (Robotham et al., 2018).

The limitations stated for the optical package remain central. Flux assignment in blends is pragmatic rather than a sophisticated flux-sharing heuristic, with optimal deblending deferred to generative modelling in ProFit (Robotham et al., 2018). Saddle-point segmentation cannot form nested islands within segments, so very extended sources with embedded clumps may require multiple segmentation passes or combined maps (Robotham et al., 2018). Crowded or confusion-limited fields and saturation halos remain challenging, although hard segment boundaries prevent the extreme aperture growth that motivated the software in the first place (Robotham et al., 2018).

7. Later reuse of the name

The title “ProFound” was also used in 2026 for a separate domain: a moderate-sized vision foundation model for volumetric prostate mpMRI (Wang et al., 4 Mar 2026). That work described a domain-specialised, volumetric, multi-modal model pretrained with self-supervised masked autoencoding on 5,000 patients and more than 22,000 unique 3D MRI volumes, and evaluated across 11 downstream clinical tasks including prostate cancer detection, Gleason grading, lesion localisation, gland volume estimation, and anatomical segmentation (Wang et al., 4 Mar 2026). Its architectures, training objectives, and applications are distinct from the astronomical source-extraction package, although the shared name can cause bibliographic ambiguity.