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
TL;DR
Faint astronomical features remain difficult to recover amid correlated imaging noise and observation-specific artifacts. ASTERIS integrates multiple exposures with self-supervised spatiotemporal denoising, recovering structures and high-redshift candidates that co-addition does not reveal, while performance depends on training data matched to the observing environment.
Problem
Observation-specific noise and resolution degradation can hinder reliable recovery of faint astronomical sources.
Method
ASTERIS trains on temporally combined exposure cubes, normalizes faint signals, and recombines denoised faint data with the bright component.
Results
ASTERIS reveals faint spiral structure, gravitational-lensing arcs, and high-redshift candidates absent from corresponding co-addition images.
Takeaways & Limitations
Reliable faint-source recovery requires training models matched to the observing environment, as Subaru-trained ASTERIS produced few false positives whereas the JWST-trained model did not.
Takeaways & Limitations
The high-redshift sample cannot constrain the characteristic UV luminosity because it is dominated by faint-end galaxy candidates.
Abstract
from arXiv · showhide
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.
Methods summary
ASTERIS combines reduced, aligned multi-exposure imaging with self-supervised denoising to improve faint-source recovery and downstream high-redshift-galaxy analysis. The workflow benchmarks denoising against co-addition and N2N using mock and observational data.
- Model and training: ASTERIS training uses JWST imaging data and a spatiotemporal 3D U-Net with attention, trained separately for long- and short-wavelength NIRCam data.The network uses a weighted combination of MSE and MAE losses and is trained on segmented image patches.
- Data processing and analysis: ASTERIS processes reduced, astrometrically aligned exposures and applies source detection to denoised outputs before high-redshift candidate selection and photometry.The JOF analysis used a composite detection image, forced-aperture photometry, PSF-based aperture corrections, and SED-based candidate selection.
- ASTERIS pipeline: ASTERIS denoises the faint ≤ 3σ image component and recombines it with the co-added bright component to retain the full dynamic range.The faint component is randomly divided into input and target sets for self-supervised learning, while the bright component is median-combined.
- Benchmarking: Mock-source benchmarking compares co-addition, N2N, and ASTERIS using detected-source counts, background-noise distributions, S/N, completeness, purity, F-score, PSF profiles, and photometric accuracy.The mock evaluation uses 50,000 injected sources and identical source-extraction parameters across methods.
- Observational applications: The observational applications examine crowded JWST fields, low-surface-brightness structures, gravitational-lensing arcs, and faint high-redshift candidates in the JOF.The JOF analysis includes SED comparisons and rest-frame UV luminosity functions for dropout-selected candidates.
Funding:
The paper reports funding from national research agencies and university programs, alongside contributor roles spanning conceptualization, methodology, data preparation, analysis, visualization, supervision, and writing.
- Funding: Funding came from the National Natural Science Foundation of China, the National Key R&D Program of China, and Tsinghua University research programs.The funding statement lists grants supporting Q.D., Z.C., J.W., and Y.G.
- Contributions: Contributor roles covered conceptualization, methodology, investigation, project administration, supervision, and funding acquisition.These roles are assigned across the authors in the contribution statements.
- Contributions: The team contributed to JWST and Subaru data preparation, luminosity-function analysis, visualization, and manuscript drafting and revision.The contribution statements separately identify the data, analysis, visualization, and writing responsibilities.
- Declarations and availability: The authors declare no competing interests, and source code, demonstration data, and observational data are made available through public repositories and archives.The availability statement identifies GitHub, Zenodo, MAST, and Subaru data archives.
Deeper detection limits in astronomical imaging using self-supervised
The paper is titled “Deeper detection limits in astronomical imaging using self-supervised spatiotemporal denoising.”
- The paper title identifies self-supervised spatiotemporal denoising as the approach to deeper detection limits in astronomical imaging.
- The manuscript includes supplementary methods, figures, tables, and references.
Imaging datasets
The study uses multiple JWST and Subaru near-infrared imaging datasets for training, benchmarking, testing, and application of ASTERIS.
- JWST datasets: JWST datasets include GLIMPSE, JOF, the Ultra Deep Field Medium Band Survey, and Medium bands, Mega Science observations.The surveys provide different exposure counts and filter coverage for training, testing, validation, and high-redshift-galaxy searches.
- JWST datasets: The GLIMPSE dataset uses deep co-added long-exposure images as ground truth for denoising validation.Individual exposures and subsets support algorithm training and benchmarking against the deep co-addition.
- JWST datasets: The JOF covers approximately 9 arcmin² with 14 filters and is used to search for high-redshift galaxy candidates.
- Data reduction: NIRCam data were processed with the JWST Science Calibration Pipeline and additional procedures addressing flicker noise, scattered light, background estimation, and astrometric alignment.The processing used pipeline calibration, 1/f-noise suppression, artifact mitigation, and catalog-based astrometric correction.
- Subaru dataset: Subaru MOIRCS imaging comprises Ks, BrG, and NB2083 near-infrared observations from an 8.2 m telescope.The observations were taken under photometric conditions with seeing FWHM of approximately 0.7″.
ASTERIS: model architecture and training strategy
ASTERIS uses spatiotemporal neural-network processing and self-supervised losses to denoise aligned multi-exposure astronomical images. Its preprocessing and training choices are designed to balance detection completeness, purity, and applicability across observing strategies and instruments.
- Model architecture: ASTERIS uses a 3D U-Net with spatiotemporal feature extraction and Restormer-inspired attention modules to process correlated astronomical background noise.The architecture combines multi-scale feature extraction, temporal and spatial skip connections, and adaptive feature weighting.
- Loss design: ASTERIS combines average and frame losses so the output approximates the underlying signal while remaining robust to outliers in individual exposures.The frame loss uses MAE-based robustness, while the weighting factor k balances completeness and purity; k = 0.125 was selected for training.
- Preprocessing and optimization: 3σ clipping outperformed alternative clipping thresholds and Poisson-based losses across the tested completeness and purity measurements.Without sigma-clipping, both completeness and purity were severely degraded.
- Instrument adaptation: Instrument-specific training was effective for Subaru, whereas applying the JWST-trained model to Subaru introduced false-positive detections.The authors attribute this difference to stronger atmospheric scattering noise and turbulence-induced resolution degradation in ground-based observations.
Quantitative mock tests
The mock-test pipeline benchmarks ASTERIS and competing denoising methods using injected sources in astronomical imaging data. Performance is assessed through source recovery metrics based on true- and false-positive detections.
- Mock-data construction: Mock sources span 27.5–31.5 mag and are injected into realistic imaging data with empirical JWST point-spread-function properties.The source distribution follows a third-order power law, and the mock-test parameters are summarized in Table S2.
- Evaluation metrics: Completeness is the fraction of injected sources correctly detected, while purity is evaluated from false-positive and positive detections.The pipeline also computes the F-score to combine completeness and purity.
- Source detection: Sources are identified with Source Extractor using a threshold of at least eight connected pixels above 0.5σ significance.Cross-matching against injected ground-truth sources determines true-positive and false-positive detections.
High-redshift Galaxies Identification and UV Luminosity Function
ASTERIS-based photometry and color selection were used to identify high-redshift galaxy candidates in JWST JOF imaging and derive UV luminosity functions. The faintest luminosity-function bins remain incompletely sampled, and the characteristic luminosity was fixed because the sample is dominated by faint candidates.
- Candidate selection: Source detection combined ASTERIS-denoised F277W, F356W, and F444W mosaics before Source Extractor and forced-aperture photometry.Photometry used 0.1″ radius circular apertures.
- Candidate selection: High-redshift candidates were selected with Lyman-break dropout criteria spanning z ∼ 9 to 22.5 across F115W, F150W, and F200W.The selection also required robust redder-band detections, non-detections in bluer bands, color constraints, and photometric-redshift probabilities.
- Candidate selection: Photometric redshifts were estimated with EAZY SED fitting across all 14 bands, followed by visual inspection to remove imaging artifacts and contamination from nearby bright sources.The adopted redshift solution corresponded to the minimum χ2 SED fit subject to the stated probability thresholds.
- UV luminosity function: The faintest luminosity-function bins had about 10% mean completeness at MUV = -16.1 for F115W dropouts and MUV = -16.5 for F150W dropouts.No luminosity function was derived for F200W dropouts because the sample size was insufficient for statistical constraints.
- UV luminosity function: The selected samples had an estimated cosmic variance of ~20%, numerically comparable to the relative Poisson uncertainty for bins containing fewer than 25 sources.The authors state that this did not affect their conclusions.
- UV luminosity function: The fitted faint-end slopes were αS = -2.45 ± 0.03 for F115W dropouts and αS = -2.28 ± 0.02 for F150W dropouts.The analysis fixed the characteristic luminosity because the sample was dominated by candidates fainter than MUV > -18 and could not constrain it.