GPUPhot: Kernel-Based Algorithms for Point-Source Detection and Photometry with a Spatially Variable PSF

arXiv:2611.03753 · astro-ph.IM · Submitted 2026-11-02 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Astrophysics. The day in astrophysics: observations, models and what the data forces.

Vera: We're looking at the paper: "GPUPhot: Kernel-Based Algorithms for Point-Source Detection and Photometry with a Spatially Variable PSF".

Jocelyn: Here's the gist of the paper. Modern imaging surveys generate data at rates that exceed the throughput of photometry pipelines built on iterative, per-source fitting, a construct that does not map onto parallel hardware.

Vera: The paper presents GPUPhot, a set of algorithms for point-source detection and photometry reformulated to run natively on a Graphics Processing Unit (GPU) by casting every stage as a convolution with a fixed kernel or an element-wise operation on full-frame maps.

Jocelyn: This GPU-native framework is designed to handle wide-field images with a spatially variable Point Spread Function (PSF), which is described through principal-component eigenmodes, allowing the matched filter to decompose exactly into genuine convolutions while retaining its maximum-likelihood foundation.

Vera: The authors demonstrate that this framework achieves sub-pixel astrometric accuracy and fewmillimagnitude photometric precision even in fields with strong background gradients and markedly inhomogeneous PSFs.

Title and authors: Vera: They start with defect correction and background estimation before moving on to isolated-star selection, PSF modeling, matched-filter detection, and finally aperture photometry.

Jocelyn: Before we get to the detection part, they detail how they handle cosmic ray mitigation using a two stage filter involving a discrete Laplace operator convolution followed by an iterative Fast Fourier Transform interpolation scheme.

Subrahmanyan: That’s interesting because they make sure that an isolated cosmic ray may interpolate to the background level, whereas the surrounding light of a star may keep its reconstructed core above the threshold.

Vera: And for random telegraph noise, they use a median filter to create a smooth reference image and then use local statistical analysis on the residual map to flag outliers.

Jocelyn: That helps suppress erroneous detections of structured image features, like the wings of bright point sources or steep background gradients.

Subrahmanyan: Then they deal with background subtraction by combining tile based percentile statistics with FFT based convolution to produce a smooth, locally representative background map.

Vera: The goal there seems to be removing that large scale gradient across the field while still preserving those compact point sources for detection.

Jocelyn: So once we have the map cleaned up, how do they model that spatially varying PSF? That seems like the core challenge mentioned in their title.

The paper's summary: Vera: To model the spatially varying PSF, they construct an empirical model in three steps. First, they detect isolated stars using a fixed Gaussian kernel approximation.

Jocelyn: Then they build a position independent reference PSF P0, and finally characterize the spatial deviations from that reference using principal component analysis.

Subrahmanyan: They use this principal component analysis to find a small set of fixed templates, or eigenmodes, which capture the bulk of the star to star PSF variation.

Vera: These eigenmodes are then projected onto and interpolated into continuous spatial coefficient maps across the focal plane using FFT based interpolation.

Jocelyn: That’s how they build that continuous map that describes how the PSF changes depending on where you are in the image.

Subrahmanyan: The spatially varying PSF is then modeled as a sum of fixed templates with position dependent amplitudes, P of x; mu equals P0 of x plus sigma K k from one to K c k mu ek(x).

Vera: Those scalar coefficients, c k mu, encode all that spatial variation.

Jocelyn: So the core idea is they can describe the PSF as a sum of fixed templates where the amplitudes change based on position in a way that’s captured by these eigenmodes.

Subrahmanyan: This is much more flexible than just using one single, fixed PSF for the whole image.

The paper's improvements: Vera: The detection part uses a generalized matched filter of Equation twelve, which generalizes the fixed Gaussian kernel detection to the empirical PSF.

Jocelyn: This filter identifies local maxima of the resulting significance map, and that statistic is defined as nu mu equals M con P bar mu over sigma mu p A mu.

Subrahmanyan: And A mu is a position dependent effective kernel area derived from those coefficient maps, which shows how the detection area changes across the image.

Vera: They then filter these initial candidates by isolation criteria, a stricter significance threshold, and a saturation criterion to reject non linear response.

Jocelyn: The integer pixel peak positions are refined to subpixel accuracy using the first moment centroid of nu within a square window around each maximum.

Subrahmanyan: For photometry, they use optimized circular aperture photometry that decomposes the process into flux measurement, aperture correction, noise modeling, and brightness dependent rule selection.

Vera: The crucial part is the aperture correction C of r; mu which they estimate empirically from growth curves of isolated stars grouped into spatial clusters to capture smooth spatial PSF variation across the focal plane.

Jocelyn: They also calibrate the optimal aperture radius ropt as a function of the convolutional detection significance nu through an empirical power law fit, so it’s adaptive.

Conclusion: Vera: So, looking at these results, we see subpixel astrometric residuals of zero point eleven to zero point thirty-eight arcseconds and a constant photometric offset of about zero point zero five magnitude.

Jocelyn: That offset is fully absorbed into the photometric zero point, which reaches a few millimagnitude floor for bright sources.

Subrahmanyan: This means that even in those tricky fields with strong background gradients, the precision holds up at a few millimagnitude level.

Vera: Overall, GPUPhot achieves high throughput at a low computational cost; reductions take under seven seconds for a two thousand times two thousand CCD frame on a consumer grade GPU.

Jocelyn: It’s already in routine operation at the Teide Observatory processing over one point two petabytes of data since twenty twenty three.

Subrahmanyan: The whole point of GPUPhot is showing that you can get high precision with complex spatial variation by moving the heavy lifting onto parallel hardware using these kernel based algorithms.

Vera: So, this paper on GPUPhot shows how to make photometric pipelines fast and accurate without sacrificing detail, even when the sky itself is messy.

Miguel R. Alarcon, Samuel Lemes-Perera, Miquel Serra-Ricart, Javier Licandro

Light Bridges · Instituto de Astrofísia de Canarias Instituto de Astrofísia de Canarias Departamento de Astrofísia Universidad de La Laguna Departamento de IngenierƑa InformƓtica y de Sistemas Universidad de La Laguna

astro-ph.IM

Submitted: 2026-11-02

Updated: 2026-11-02

Comments: 25 pages, 14 figures, accepted for publication in ApJS

Code: https://github.com/Light-Bridges/GPUPhot

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 83/100

The gist: Modern imaging surveys generate data at rates that exceed the throughput of photometry pipelines built on iterative, per-source fitting, a construct that does not map onto parallel hardware.

Key concepts

GPU-Native Framework
This approach redesigns the entire image processing pipeline so that every step—from noise reduction to detection—is performed using standard GPU operations like convolutions or element-wise calculations on full maps. This bypasses traditional, iterative fitting methods that don't map well to parallel hardware, allowing for extremely fast processing of large datasets.
Spatially Varying PSF Modeling
The paper models how the star's blur changes across the image using principal-component eigenmodes. By decomposing this variation into a small set of fixed templates and position-dependent coefficients, the system can exactly represent complex PSFs while keeping the core maximum-likelihood detection foundation intact.
Matched Filter Detection
Source detection uses a generalized matched filter that is built from the empirical PSF model. This filter is evaluated as a sum of genuine convolutions multiplied by precomputed maps, enabling it to efficiently find local peaks in significance maps even when the PSF changes across the field.
Adaptive Aperture Photometry
Instead of using a fixed measurement area, this method uses an adaptive aperture scheme. The optimal aperture radius for each source is calculated dynamically based on maximizing the signal-to-noise ratio, ensuring accurate flux measurement regardless of local background gradients or PSF variations.

Terminology

Summary

Modern imaging surveys generate data at rates that exceed the throughput of photometry pipelines built on iterative, per-source fitting, a construct that does not map onto parallel hardware. p. 0

Where the paper says it

Alarcon , 1, 2, 3 Samuel Lemes-Perera , 1, 4 Miquel Serra-Ricart , 1, 2, 3 and Javier Licandro 2, 3 1Light Bridges, Observatorio Astron´omico del Teide. Carretera del Observatorio del Teide, s/n, G¨u´ımar, Canarias, Spain 2 Instituto de Astrof´ısica de Canarias (IAC), C/ V´ıa L´actea, s/n, E-38205, La Laguna, Spain 3Departamento de Astrof´ısica, Universidad de La Laguna (ULL), E-38206 La Laguna, Canarias, Spain 4Departamento de Ingenier´ıa Inform´atica y de Sistemas, Universidad de La Laguna (ULL), E-38206 La Laguna, Canarias, Spain (Received September 6, 2026; Revised September 17, 2026; Accepted October 8, 2026) ABSTRACT Modern imaging surveys generate data at rates that exceed the throughput of photometry pipelines built on iterative, per-source fitting, a construct that does not map onto parallel hardware.

The paper presents GPUPhot, a set of algorithms for point-source detection and photometry reformulated to run natively on a Graphics Processing Unit (GPU) by casting every stage as a convolution with a fixed kernel or an element-wise operation on full-frame maps. p. 0
Where the paper says it

We present a set of algorithms for point-source detection and photometry reformulated so that the entire reduction runs natively on a Graphics Processing Unit (GPU), implemented in GPUPhot. Their design rests on a single principle: every stage is cast as a convolution with a fixed kernel or an element-wise operation on full-frame maps.

This GPU-native framework is designed to handle wide-field images with a spatially variable Point Spread Function (PSF), which is described through principal-component eigenmodes, allowing the matched filter to decompose exactly into genuine convolutions while retaining its maximum-likelihood foundation. p. 0
Where the paper says it

The algorithms are designed to operate on wide-field images with a spatially variable Point Spread Function (PSF), where the field-dependent PSF is described through a small set of fixed, flux-conserving principal-component eigenmodes with position-dependent coefficients, so that the spatially varying matched filter decomposes exactly into a few genuine convolutions while retaining its maximum-likelihood foundation. We show that, even in fields with strong background gradients and markedly inhomogeneous PSFs, the framework achieves sub-pixel astrometric accuracy and fewmillimagnitude photometric precision for bright sources.

The authors demonstrate that this framework achieves sub-pixel astrometric accuracy and fewmillimagnitude photometric precision even in fields with strong background gradients and markedly inhomogeneous PSFs. p. 0
Where the paper says it

The algorithms are designed to operate on wide-field images with a spatially variable Point Spread Function (PSF), where the field-dependent PSF is described through a small set of fixed, flux-conserving principal-component eigenmodes with position-dependent coefficients, so that the spatially varying matched filter decomposes exactly into a few genuine convolutions while retaining its maximum-likelihood foundation. We show that, even in fields with strong background gradients and markedly inhomogeneous PSFs, the framework achieves sub-pixel astrometric accuracy and fewmillimagnitude photometric precision for bright sources.

Pipeline Architecture

The GPUPhot pipeline is structured sequentially, beginning with defect correction and background estimation before moving to isolated-star selection, PSF modeling, matched-filter detection, and finally aperture photometry. The initial steps include cosmic ray mitigation using a discrete Laplace operator convolution followed by an iterative FFT-based interpolation scheme to reconstruct affected regions. p. 2

Where the paper says it

This fixed kernel size is motivated by the sub-pixel to few-pixel spatial extent of typical cosmic ray events on the detectors employed; a larger or parametric kernel would risk encroaching on the wings of point sources in well-sampled images. Pixels within the dilated mask are set to NaN in a working copy of the image, and an iterative Fast Fourier Transform (FFT)-based interpolation scheme (fill nan fft) reconstructs the affected regions by iteratively filling the masked pixels, repeating until no NaN values remain or the count of unfilled pixels is unchanged between successive iterations.

Random Telegraph Noise is handled via a median filter and statistical characterization of the residual map, followed by a morphological isolation criterion to suppress structured image features like source wings. Background subtraction employs tile-based percentile statistics combined with FFT-based convolution to produce a smooth, locally representative background map, intentionally masking only compact point sources to estimate the background. p. 3
Where the paper says it

BACKGROUND SUBTRACTION Accurate background estimation is a critical prerequisite for reliable source detection and photometry, particularly in crowded fields or images with spatially varying sky levels. Our algorithm produces a smooth, locally representative background map by combining tile-based percentile statistics with FFT-based convolution, and is designed to handle images of arbitrary size and pixel scale.

PSF Modeling and Spatial Variation

Accurate photometry requires modeling the spatially varying PSF by constructing an empirical model in three steps: detection of isolated stars using a fixed Gaussian kernel approximation, construction of a position-independent reference PSF P0, and characterization of spatial deviations from the reference via Principal Component Analysis (PCA). The spatially varying PSF is modeled as a sum of fixed templates with position-dependent amplitudes, P(x; µ) = P0(x) + Σ Kk k=1 ck(µ)ek(x), where the basis functions are determined empirically and the scalar coefficient maps encode all spatial variation. p. 7

Where the paper says it

Equation (10) cannot be generalized by simply replacing P0 by a position-dependent kernel P(x; µ), because the resulting operation is no longer a convolution: a kernel that varies with µ does not admit the FFT-based evaluation on which efficient detection relies. To resolve this tension, we model the spatially varying PSF as a sum of fixed templates with position-dependent amplitudes, P(x; µ) = P0(x) +X K k=1 ck(µ)ek(x), (11) where the basis functions {ek} K k=1 (hereafter PSF deviation eigenmodes) are fixed templates determined empirically from the data, and the scalar coefficient maps ck(µ) encode all of the spatial variation.

The PCA is applied to noise-normalized deviations of isolated stars to find a small set of fixed templates, or eigenmodes ek, which capture the bulk of the star-to-star PSF variation. p. 7
Where the paper says it

The goal is to find a small set of fixed templates that, combined linearly, reproduce the bulk of the star-to-star PSF variation. Concretely, the eigenmodes {ek} K k=1 are the K leading principal components of the centered deviations { ˜di − ⟨ ˜d⟩}: a set of mutually orthonormal patterns, ordered so that each successively captures the largest possible fraction of the remaining sample variance.

These eigenmodes are then projected onto and interpolated into continuous spatial coefficient maps ck(µ) across the focal plane using FFT-based interpolation. p. 7
Where the paper says it

Projection coefficients and spatial interpolation —Each star’s deviation is projected onto the eigenmode basis. In the normalized space, the coefficient of ˜di along ek is c˜k,i = X x ˜di(x)ek(x).

Source Detection and Centroiding

Source detection is performed by applying the matched filter of Equation (12), which generalizes the fixed Gaussian kernel detection to the empirical PSF, identifying local maxima of the resulting significance map. The detection statistic is ν(µ) = (M ⊛ P¯)µ / σµp A(µ), where A(µ) is the position-dependent effective kernel area derived from the coefficient maps. p. 8

Where the paper says it

1 from a fixed Gaussian kernel to the empirical PSF, retaining the same maximum-likelihood foundation under the locally stationary noise approximation. The detection statistic at each pixel is ν(µ) = (M ⊛ P¯)µ σµ p A(µ) , (21) where (M ⊛ P¯)µ is the spatially varying matched filter of Equation (12), and the position-dependent effective kernel area is A(µ) = X x P 2 (x; µ) = AP0 + 2X K k=1 ck(µ) πk + X K k=1 c 2 k (µ), (22)

Initial candidates are filtered by isolation criteria, a stricter significance threshold, and a saturation criterion to reject non-linear response. p. 6
Where the paper says it

Second, a stricter significance threshold Nσ,iso = 10 is applied to νG to retain only high-confidence detections. Third, a saturation criterion rejects candidates whose peak pixel value exceeds 80 % of the detector full-well capacity, providing a conservative margin against non-linear response.

The integer-pixel peak positions are refined to sub-pixel accuracy using the first-moment centroid of ν within a square window around each maximum. p. 9
Where the paper says it

Two candidates are required to be separated by at least 2 × FWHM pixels to avoid duplicate detections of the same source split by noise fluctuations. The integer-pixel peak positions are subsequently refined to sub-pixel accuracy by computing the first-moment centroid of ν within a square window of side 1 × FWHM around each maximum, µˆ = µ + P x′∈W ν(x ′ )(x ′ − µ) P x′∈W ν(x ′) , (23) where x ′ , µ and ˆµ are two-dimensional position vectors and W is the centroid window.

Adaptive Photometry and Calibration

Photometry utilizes optimized circular aperture photometry, decomposing the process into flux measurement, aperture correction, noise modeling, and brightness-dependent rule selection. The aperture correction C(r; µ) is empirically estimated from growth curves of isolated stars grouped into spatial clusters to capture smooth spatial PSF variation across the focal plane. p. 1

Where the paper says it

A difficulty common to all of them is that, across a wide field, the PSF is not constant: optical aberrations, detector tilt and atmospheric seeing gradients make it vary systematically over the focal plane. Capturing this variation is essential for both unbiased detection and precise photometry, and has been addressed either by expanding the PSF on a spatial polynomial basis (E.

The optimal aperture radius ropt is calibrated as a function of the convolutional detection significance ν through an empirical power-law fit, which allows for an adaptive scheme that selects the aperture maximizing signal-to-noise on a source-by-source basis. p. 9
Where the paper says it

The signal at each detected source is the sum of the pixel values inside a circular aperture of radius r centered on the refined position ˆµ, plus a small fraction of the flux that falls outside the aperture and which has to be recovered. The procedure can be cleanly decomposed into four parts: the aperture flux measurement itself, the aperture correction that recovers the missing flux, the noise model that propagates the photometric uncertainty and the brightness-dependent rule that selects the optimal aperture radius for each source.

Validation and Performance

The pipeline was validated through a controlled injection–recovery experiment on four instruments spanning an order of magnitude in plate scale, confirming complete detection and pure false positives over a broad operating range. The results show sub-pixel astrometric residuals of 0. p. 0

Where the paper says it

The algorithms are designed to operate on wide-field images with a spatially variable Point Spread Function (PSF), where the field-dependent PSF is described through a small set of fixed, flux-conserving principal-component eigenmodes with position-dependent coefficients, so that the spatially varying matched filter decomposes exactly into a few genuine convolutions while retaining its maximum-likelihood foundation. We show that, even in fields with strong background gradients and markedly inhomogeneous PSFs, the framework achieves sub-pixel astrometric accuracy and fewmillimagnitude photometric precision for bright sources.

11–0.38′′ and a constant photometric offset of ∼ 0. p. 8
Where the paper says it

For all but the brightest star the residuals exhibit no systematic structure beyond pixellevel photon noise. This shows a faint dipolar residual, a positive and a negative lobe on opposite sides of the center, which is the first-order signature of a small, subpixel offset between the model and the stellar centroid, rather than a deficiency of the eigenmode basis.

05 mag, which is fully absorbed into the zero point, reaching a few-millimagnitude floor for bright sources. p. 0
Where the paper says it

The algorithms are designed to operate on wide-field images with a spatially variable Point Spread Function (PSF), where the field-dependent PSF is described through a small set of fixed, flux-conserving principal-component eigenmodes with position-dependent coefficients, so that the spatially varying matched filter decomposes exactly into a few genuine convolutions while retaining its maximum-likelihood foundation. We show that, even in fields with strong background gradients and markedly inhomogeneous PSFs, the framework achieves sub-pixel astrometric accuracy and fewmillimagnitude photometric precision for bright sources.

The overall system achieves high throughput at low computational cost, with reductions taking under 7 s for a 2k×2k CCD frame on a consumer-grade GPU. The pipeline is already in routine operation at the Teide Observatory, processing over 1. p. 1
Where the paper says it

Its single design principle is that every stage of the reduction is cast either as a convolution with a fixed kernel or as an element-wise operation on full-frame maps, so that the whole reduction runs natively on a GPU even where the underlying problem is not naturally a convolution. GPUPhot is already in routine operation as the realtime reduction pipeline of the Two-meter Twin Telescope (TTT5 ), the Transient Survey Telescope (TST6 ) and ATLAS-Teide (J.

2 PB of data since 2023. p. 0
Where the paper says it

7 million images since 2023, amounting to 1. 2 PB of data, at effective end-to-end rates below 15 s per 150-Mpx image, from ingestion of the raw frame to a source catalog with astrometric and photometric solutions.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. GPUPhot can perform point-source detection and photometry natively on a Graphics Processing Unit by reformulating every stage as either a convolution with a fixed kernel or an element-wise operation on full-frame maps. This allows the entire reduction to run natively on a GPU even where the underlying problem is not naturally a convolution.

  2. The system can handle spatially varying Point Spread Functions (PSF) by describing their variation through a small set of fixed, flux-conserving principal-component eigenmodes with position-dependent coefficients, which allows the matched filter to decompose exactly into a few genuine convolutions while retaining its maximum-likelihood foundation.

  3. The pipeline can achieve subpixel astrometric accuracy and fewmillimagnitude photometric precision even in fields with strong background gradients and markedly inhomogeneous PSFs, as demonstrated by the validation against external catalogs.

  4. Defect correction can mitigate cosmic rays using a two-stage filter involving a discrete Laplace operator convolution followed by an iterative Fast Fourier Transform interpolation scheme, ensuring that an isolated cosmic ray may interpolate to the background level, whereas the surrounding light of a star may keep its reconstructed core above the threshold.

  5. The system can handle random telegraph noise by applying a median filter to produce a smooth reference image and then using local statistical analysis on the residual map to flag outliers, which suppresses the erroneous detection of structured image features–such as the wings of bright point sources or steep background gradients.

  6. Background estimation can produce a smooth, locally representative background map by combining tile-based percentile statistics with FFT-based convolution, effectively removing the large-scale gradient across the field while preserving compact point sources for detection.

  7. Source detection uses a generalized matched filter of Equation (12), which is evaluated as a sum of K plus one genuine convolutions multiplied by precomputed coefficient maps, allowing it to operate efficiently on the GPU even with spatially varying PSFs.

  8. Photometry utilizes an adaptive aperture scheme where the optimal radius is determined by maximizing the signal-to-noise ratio, calibrated via an empirical power-law fit of Equation (36), ensuring that the optimal radius for every detection is determined from a single quantity already available at detection time.

  9. The system can perform astrometric calibration by transforming pixel centroids into celestial coordinates using a Simple Imaging Polynomial (SIP) of order three simultaneously with linear WCS terms, achieving per-axis residuals of 0.11 to 0.38 arcseconds in validation tests.

  10. Photometric calibration establishes a constant zero point of approximately +0.05 mag across the entire dynamic range, independent of brightness and position, which is fully absorbed into the photometric zero point, resulting in a few-millimagnitude floor for bright sources.

Abstract

Modern imaging surveys generate data at rates that exceed the throughput of photometry pipelines built on iterative, per-source fitting, a construct that does not map onto parallel hardware. We present a set of algorithms for point-source detection and photometry reformulated so that the entire reduction runs natively on a Graphics Processing Unit (GPU), implemented in GPUPhot. Their design rests on a single principle: every stage is cast as a convolution with a fixed kernel or an element-wise operation on full-frame maps. The algorithms are designed to operate on wide-field images with a spatially variable Point Spread Function (PSF), where the field-dependent PSF is described through a small set of fixed, flux-conserving principal-component eigenmodes with position-dependent coefficients, so that the spatially varying matched filter decomposes exactly into a few genuine convolutions while retaining its maximum-likelihood foundation. We show that, even in fields with strong background gradients and markedly inhomogeneous PSFs, the framework achieves sub-pixel astrometric accuracy and few-millimagnitude photometric precision for bright sources. The pipeline is already in production at the telescopes of the Teide Observatory, where it has processed 3.7 million images since 2023, amounting to 1.2 PB of data, at effective end-to-end rates below 15 s per 150-Mpx image, from ingestion of the raw frame to a source catalog with astrometric and photometric solutions. Because the same primitives underlie every stage, the individual algorithms are self-contained and applicable well beyond photometry. They form the basis of a complete, openly developed software package, released publicly with this paper.

Sources

Related papers