PaperlayerView exact revision on arXiv

Source-linked AI summary

Deeper detection limits in astronomical imaging using self-supervised spatiotemporal denoising

Yuduo Guo, Hao Zhang, Mingyu Li, Fujiang Yu, Yunjing Wu, Yuhan Hao, Song Huang, Yongming Liang, Xiaojing Lin, Xinyang Li, Jiamin Wu, Zheng Cai, Qionghai Dai

arXiv:2602.17205v2astro-ph.IMastro-ph.COastro-ph.GAcs.AI

TL;DR

Astronomical imaging detection limits are constrained by noise, including correlations across pixels and exposures. ASTERIS exploits those correlations across multiple exposures, improving detection limits by 1.0 mag at 90% completeness versus co-addition and by 1.0 mag at 90% purity versus N2N.

  • Problem

    Astronomical imaging detection limits are constrained by noise, including correlations across pixels and exposures that could in principle be learned and corrected.

    Source: Concept, design, and implementation · p. 3

  • Method

    ASTERIS is a self-supervised denoising algorithm that exploits spatiotemporal correlations across multiple exposures to produce a higher-S/N image during co-addition.

    Source: Concept, design, and implementation · p. 3

  • Results

    At 90% completeness, ASTERIS improves the detection limit by 1.0 mag over co-addition; at 90% purity, it improves by 1.0 mag over N2N.

    Source: Algorithmic performance · p. 4

  • Takeaways & Limitations

    162 high-redshift galaxy candidates at zphot ≳ 9 were identified, approximately three times the previous number, with the faintest at MUV ≈ -16 mag, one magnitude fainter than previous work.

    Source: Application to faint high-redshift… · p. 6

    ASTERIS supports only an even number of input exposures; its 4-exposure variant improves the detection limit by 0.7 mag at 90% completeness, versus 1.0 mag for the 8-exposure version.

    Source: Extension to different numbers of… · p. 7

Abstract

from arXiv · show

The detection limit of astronomical imaging observations is limited by several noise sources. Some of that noise is correlated between neighbouring image pixels and exposures, so in principle could be learned and corrected. We present an astronomical self-supervised transformer-based denoising algorithm (ASTERIS), that integrates spatiotemporal information across multiple exposures. Benchmarking on mock data indicates that ASTERIS improves detection limits by 1.0 magnitude at 90% completeness and purity, while preserving the point spread function and photometric accuracy. Observational validation using data from the James Webb Space Telescope (JWST) and Subaru telescope identifies previously undetectable features, including low-surface-brightness galaxy structures and gravitationally-lensed arcs. Applied to deep JWST images, ASTERIS identifies three times more redshift > 9 galaxy candidates, with rest-frame ultraviolet luminosity 1.0 magnitude fainter, than previous methods.

Concept, design, and implementation

ASTERIS denoises aligned exposures with self-supervised spatiotemporal learning, improving detection depth while preserving image and photometric fidelity. Tests on JWST and Subaru data recover faint sources and structures, and its application to JWST deep imaging yields more high-redshift galaxy candidates.

Concept, design, and implementation

  • ASTERIS extends single-image denoising to aligned multi-exposure stacks, using self-supervised learning to estimate shared signal from independent noise realizations.Training uses 16 exposures split into independent input and target sets of eight, with composite MSE and MAE losses; a dedicated attention mechanism fuses neighboring pixels across exposures. After training, it denoises any eight input exposures without additional reference data. It clips pixels above 3σ for median combination and reintegrates them, concentrating learning on faint regions while preserving dynamic range.Source: Concept, design, and implementation · p. 3

Observational validation and generalization

  • In an 8-exposure JWST field, ASTERIS recovered 93 of 168 deep-stack sources versus 50 for co-addition; on all 168 exposures it found 229 sources, 35% more than standard methods.For an 8-exposure image, ASTERIS raised SSIM from 0.37 to 0.67 for a spiral galaxy and from 0.59 to 0.81 for a lensed arc, also improving identification of faint diffuse galaxy groups.Source: Observational validation and… · p. 5
  • A dedicated Subaru-trained ASTERIS model recovered faint sources from eight 3-second exposures against a 2088-second co-added reference; direct transfer of the JWST-trained model produced some false positives.The test used Subaru MOIRCS data in the Ks band.Source: Observational validation and… · p. 6

Application to faint high-redshift galaxies

  • ASTERIS identified 162 zphot ≳ 9 candidates—about three times the number previously reported [52, 53]—with faintest candidates one magnitude fainter than earlier work.The 5σ depth improved from 30.3–30.8 mag in previous work to 30.9–31.6 mag; ASTERIS recovered all previously spectroscopically confirmed galaxies and 75% of its candidates were not previously reported [52, 53].Source: Application to faint high-redshift… · p. 6
  • ASTERIS improves high-redshift selection by detecting fainter sources and sharpening photometric-redshift constraints for marginal detections.Deeper dropout-band limits help distinguish high-redshift Lyman breaks from similar low-redshift Balmer breaks, particularly for sources near S/N ~ 2 in standard co-additions.Source: Application to faint high-redshift… · pp. 6–7
  • ASTERIS identified four previously unreported z ~ 16 to 22.5 candidates and UV luminosity-function slopes of –2.45 ± 0.03 at z ~ 9 to 12 and –2.28 ± 0.02 at z ~ 12 to 16.The slopes indicate higher number densities of faint galaxies down to MUV ~ -16 than predicted by theoretical models.Source: Application to faint high-redshift… · p. 7

Extension to different numbers of exposures

  • The 4-exposure ASTERIS variant improves the 90%-completeness detection limit by 0.7 mag, versus 1.0 mag for the 8-exposure version.The authors discourage duplicating exposures because it increases false positives; when more exposures are available, they can be grouped and co-added into the required input count.Source: Extension to different numbers of… · p. 7

Methods summary

ASTERIS is a self-supervised denoising method that learns spatiotemporal correlations across aligned exposures, then combines the denoised faint signal with the bright image component. The methods also describe mock-data benchmarking, real-image validation, and a pipeline for selecting and analyzing high-redshift galaxy candidates.

ASTERIS workflow

  • ASTERIS aligns repeated exposures, separates bright and faint signals at a 3σ threshold, learns from faint input–target exposure pairs, and recombines the denoised faint component with the bright median.The faint component is Z-score normalized before network input; the final image retains full dynamic range.Source: Methods summary · pp. 8, 10, Figure 1 · p. 9

Network and training

  • The model uses a spatiotemporal 3D U-Net with multi-deconvolved-head-transposed attention, trained separately on short- and long-wavelength data with weighted MSE and MAE losses.Training used 120,000 patches formatted as 8 × 128 × 128 pixels and converged in approximately 10 iterations (~26 hours).Source: Methods summary · p. 8

Mock-data benchmarking

  • Mock-source benchmarks used JWST NIRCam F115W images and compared completeness, purity, F-score, PSF profiles, and forced-aperture photometric accuracy across denoising methods.The mock sources ranged from 27.5 to 31.5 mag; 50,000 injected sources were used for the S/N and photometric-accuracy comparisons.Source: Methods summary · pp. 8, 12, Figure 2 · p. 11

High-redshift candidate analysis

  • For JOF candidate analysis, ASTERIS images supported source detection, multi-filter forced-aperture photometry, photometric-redshift selection, SED-based property inference, and luminosity-function estimation.The workflow used EAZY for preliminary SED fitting, BEAGLE for physical properties, and mock-source injection and recovery to assess completeness.Source: Methods summary · p. 8

Observational validation

  • Real-image validation covered JWST and Subaru observations, including crowded fields, spiral arms, gravitational-lensing arcs, and faint diffuse sources.The comparisons included co-additions with many exposures serving as ground-truth images.Source: Figure 3 · p. 13, Methods summary · p. 14

High-redshift candidate examples

  • ASTERIS raised F090W S/N from < 2.0 to 4.7 for an ambiguous-redshift source, shifting best-fitting zphot from 9.45 to 2.08; it also revealed an F200W-dropout candidate absent from co-addition.For a faint F115W-dropout candidate, F090W and F115W photometric depth changed by > 1.0 mag.Source: Figure 5 · p. 16, Methods summary · p. 17

Funding:

The authors report support from Chinese national and institutional funding programs, a postdoctoral science foundation, and the New Cornerstone Science Foundation.

Funding:

  • Q.D. received National Natural Science Foundation of China funding (62088102); Z.C. received National Key R&D Program funding (2023YFA1605600) and National Natural Science Foundation funding (12525303).Z.C. also received support from the Tsinghua University Initiative Scientific Research Program.Source: Funding · p. 27
  • J.W. received National Natural Science Foundation of China grants 62222508 and 62525506; Y.G. received grants 62505156, 2024M761675, and 2025T180226.Y.G.'s support included the China Postdoctoral Science Foundation grants 2024M761675 and 2025T180226.Source: Funding · p. 27
  • H.Z. received National Natural Science Foundation of China grant 62505157, while Y.G., J.W., and Q.D. were supported by the Beijing Key Laboratory of Cognitive Intelligence.Z.C. and J.W. also received New Cornerstone Science Foundation support through the XPLORER PRIZE.Source: Funding · p. 27

Deeper detection limits in astronomical imaging using self-supervised

The supplied passages identify Yuduo Guo et al. and provide the paper’s DOI, but do not describe its methods or findings.

spatiotemporal denoising

Materials and Methods · Imaging datasets

The study uses JWST surveys for benchmarking, training, testing and candidate searches, and Subaru MOIRCS observations divided between model training and testing.

Imaging datasets

  • GLIMPSE (34) provides the benchmarking data: F090W, F115W and F444W each have 168 independent exposures, while deep co-adds serve as validation ground truth.The survey’s single massive-galaxy-cluster field has seven broad-band filters at wavelengths ≥ 0.9 µm and two medium-band filters, F410M and F480M.Source: Imaging datasets · p. 30
  • JOF (50, 51) covers approximately 9 arcmin2 in 14 filters and was used to search for candidate high-redshift galaxies with ASTERIS.Independent exposures range from 12 in F200W and F356W to 60 in F182M and F250M.Source: Imaging datasets · p. 30
  • Ultra Deep Field Medium Band Survey data (59) provided ASTERIS training images in F182M, F210M and F480M, each with 24 independent exposures.Effective exposure times across the survey’s filters range from approximately 1.1 to 7.7 hours.Source: Imaging datasets · p. 30
  • Medium bands, Mega Science data (41) were used for ASTERIS testing, not training, because the filters have only 8 or 16 independent exposures.F090W has 16 independent exposures; all other bands have eight.Source: Imaging datasets · p. 30

JWST NIRCam imaging data reduction

  • JWST NIRCam data were processed with pipeline jwst v1.15.1 (60) and CRDS context jwst_1318.pmap, with added flicker-noise suppression, scattered-light mitigation and background estimation.The added procedures follow previous work (61, 62).Source: JWST NIRCam imaging data reduction · p. 30
  • Stage 1 jump rejection used a 3σ threshold; after default Stage 2 processing, 1/f noise was subtracted per amplifier and astrometry was aligned to survey catalogs.JOF images used GOODS-S Deep v2.0 reference positions (50), while other datasets used DESI DR10 (63).Source: JWST NIRCam imaging data reduction · pp. 30–31
  • Visual inspection excluded 27 artifact-affected JOF exposures; remaining images were mosaicked onto a unified WCS target grid and resampled to 0.04″ pixels.Outlier rejection used default parameters.Source: JWST NIRCam imaging data reduction · p. 31

Subaru MOIRCS imaging data and reduction

  • Subaru MOIRCS (6, 47, 48, 88) Ks, BrG and NB2083 images were reduced with MCSRED (90), aligned using SCAMP (91), and coadded with SWARP (92); 2017/2024 data trained the model and 2025 data tested it.Observations were taken under photometric conditions with seeing FWHM ∼ 0.7″; programs were S17A-198S (49), S24A-091 and S24B-135.Source: Subaru MOIRCS imaging data and reduction · p. 31

ASTERIS: model architecture and training strategy

ASTERIS combines spatiotemporal attention with a 3D U-Net and complementary average and frame losses to denoise aligned multi-exposure images. Its training pipeline uses 3σ clipping, and the paper reports that alternative clipping thresholds and loss strategies did not outperform this configuration.

ASTERIS: model architecture and training strategy

  • ASTERIS uses a 3D U-Net with spatial and temporal skip connections, 3D attention and feed-forward blocks, and output-stage attention to fuse multi-exposure features.The 3D-MDTA block combines CNN local connectivity and translation equivariance with transformer-like long-range dependencies; the 3D-GDFN adaptively weights features.Source: ASTERIS: model architecture and… · pp. 31–32
  • Average MSE loss targets the mean signal to reduce background fluctuations, while frame MAE targets the median to suppress outlier-driven false positives.The losses balance detection completeness and purity; the selected weight was k = 0.125, and tested weights from 0.1k₀ to 10k₀ had similar F-scores that outperformed either single-loss model.Source: ASTERIS: model architecture and… · p. 32

Data pre-processing for ASTERIS

  • The pipeline forms aligned exposure cubes, trains and infers on the faint component below 3σ, median-coadds the bright component, then recombines it after denoising.Z-score normalization standardizes dynamic range across filters and programs while preserving faint-signal distributions and mitigating global brightness shifts.Source: Data pre-processing for ASTERIS · p. 33

Sigma-clipping threshold

  • 3σ clipping outperformed tested alternatives in completeness and purity; 2σ left more outliers, while 5σ had slightly lower purity.Poisson-based and flux-weighted loss alternatives also did not outperform the 3σ-clipping configuration.Source: Sigma-clipping threshold · pp. 33–34

Training strategy

  • JWST training used separate long- and short-wavelength NIRCam models, each trained on 120,000 patches and converging within 10 epochs.Eight exposures were randomly selected from sets of at least 16 per patch each epoch, with rotations and flips for augmentation; each model required about 26 hours to train.Source: Training strategy · p. 34
  • On an 8 × 1650 × 1650-pixel cube, inference produced a denoised 1650 × 1650-pixel image in approximately 18.1 seconds.The trained model was generally applicable across the adopted JWST NIRCam datasets from different programs, including datasets with different exposure times.Source: Training strategy · p. 34
  • A Subaru-trained ASTERIS model recovered faint sources with few false positives, whereas a JWST-trained model introduced false-positive detections on Subaru data.The authors attribute the transfer issue primarily to atmospheric scattering and turbulence-related resolution degradation, and recommend instrument-specific training.Source: Training strategy · p. 34

Quantitative mock tests

The mock tests benchmarked ASTERIS against prior denoising methods on JWST NIRCam GLIMPSE data and evaluated source recovery and photometric behavior using injected sources.

Quantitative mock tests

  • ASTERIS was benchmarked against prior denoising methods on mock JWST NIRCam GLIMPSE images, with co-addition baselines using temporally 3σ-clipped aligned exposures and N2N using Restormer (37).Co-addition baselines averaged the aligned exposures after outlier rejection.Source: Quantitative mock tests · p. 34
  • Injected sources used an empirical STPSF PSF (FWHM approximately 0.036″; EE80 approximately 0.136″) and 27.5–31.5 mag third-order power-law counts consistent with observed source counts.Pixels at or below 30 mag clipped before ASTERIS were recombined after denoising.Source: Quantitative mock tests · pp. 34–35
  • Evaluation used completeness (true-positive detections of injected sources divided by total mock sources) and purity (false-positive detections divided by total positive detections).Source: Quantitative mock tests · p. 35
  • Source Extractor detections required at least eight connected pixels above 0.5σ and matched injected sources within 0.04″; PSF and photometric accuracy were assessed with PSFEx and 0.14″-aperture forced photometry.Source: Quantitative mock tests · p. 35

High-redshift Galaxies Identification and UV Luminosity Function

This section identifies high-redshift galaxy candidates using ASTERIS input configurations, dropout colors, and SED fitting, and measures their photometric redshifts and rest-frame UV luminosities. It derives UV luminosity functions for the F115W and F150W dropout samples.

Adopted Cosmology and Conventions

Source Extraction and Photometry

  • Source extraction used a combined ASTERIS-denoised F277W, F356W, and F444W detection image, followed by Source Extractor and 0.1″-radius forced-aperture photometry.Photometric uncertainties came from 10,000 random apertures in source-free regions.Source: Source Extraction and Photometry · p. 35

Sample Selection

  • Candidate selection required a dropout color break of at least 1.3 mag, robust redward-band detections, no bluer-band detections, constrained UV colors, and a high-redshift SED probability threshold.Candidates also underwent visual inspection to exclude imaging artifacts and nearby bright sources.Source: Sample Selection · pp. 36–37

UV Luminosity Measurement

  • UV luminosities were measured by fitting NIRCam photometry with BEAGLE at EAZY redshift solutions, integrating best-fitting spectra over rest-frame 1450–1550 Å, and converting to MUV.Completeness was assessed by injecting mock galaxies into real images and applying the same detection and selection procedures; the maximum total completeness was about 78%.Source: UV Luminosity Measurement · pp. 37–38
  • About 10% completeness in the faintest bins occurs at MUV = -16.1 for F115W and MUV = -16.5 for F150W; no F200W luminosity function was derived.Estimated cosmic variance was about 20%, comparable to the relative Poisson uncertainty for bins with fewer than 25 sources.Source: UV Luminosity Measurement · p. 39
  • For F115W dropouts, Φ* = 7.09 × 10^-5 Mpc^-3 mag^-1 and αS = -2.45 ± 0.03; for F150W dropouts, Φ* = 3.82 × 10^-5 Mpc^-3 mag^-1 and αS = -2.28 ± 0.02.Because the sample is dominated by candidates fainter than MUV = -18, the characteristic luminosity was fixed at MUV* = -20.32.Source: UV Luminosity Measurement · p. 39
  • Eight independent exposures achieved high detection purity, whereas duplicating a single exposure produced numerous false positives by amplifying local noise fluctuations.The comparison used a crowded JWST GLIMPSE NIRCam F115W field; duplication of 4×t0 exposures also produced many false positives.Source: Figure S20 · p. 62
  • Four-exposure ASTERIS inputs formed by subset co-addition reduced completeness through information loss, while eight independent exposures achieved higher completeness.The comparison used a crowded JWST GLIMPSE NIRCam F090W field and identical Source Extractor parameters to identify true-positive sources.Source: Figure S21 · p. 63
  • Table S3 lists ASTERIS-identified high-redshift galaxy candidates with photometric redshift, rest-frame UV absolute magnitude, and dropout filter, plus one low-redshift candidate.The table defines zphot as the median BEAGLE photometric redshift and MUV as the median rest-frame UV absolute AB magnitude.Source: Table S3 · p. 67

Related papers

Loading 2602.17205v2…