Deeper detection limits in astronomical imaging using self-supervised spatiotemporal denoising.
Saved in:
| Title: | Deeper detection limits in astronomical imaging using self-supervised spatiotemporal denoising. |
|---|---|
| Authors: | Guo, Yuduo (AUTHOR), Zhang, Hao (AUTHOR), Li, Mingyu (AUTHOR), Yu, Fujiang (AUTHOR), Wu, Yunjing (AUTHOR), Hao, Yuhan (AUTHOR), Huang, Song (AUTHOR), Liang, Yongming (AUTHOR), Lin, Xiaojing (AUTHOR), Li, Xinyang (AUTHOR), Wu, Jiamin (AUTHOR), Cai, Zheng (AUTHOR), Dai, Qionghai (AUTHOR) |
| Source: | Science. 4/30/2026, Vol. 392 Issue 6797, p1-11. 11p. |
| Subjects: | Signal denoising, Spatiotemporal processes, Deep learning, Space telescopes, Imaging systems in astronomy, Galactic redshift, James Webb Space Telescope (Spacecraft), Machine learning |
| Abstract: | The detection limit of astronomical imaging observations is limited by several noise sources. Some of that noise is correlated between neighboring pixels and exposures, so in principle it could be learned and corrected. We present the Astronomical Self-supervised Transformer-based Denoising (ASTERIS) algorithm, which integrates spatiotemporal information across multiple exposures. Benchmarking on mock data indicated 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 the Subaru Telescope identified previously undetectable features, including low-surface-brightness galaxy structures and gravitationally lensed arcs. Applied to deep JWST images, ASTERIS identified three times more redshift ≳9 galaxy candidates than previous methods, with rest-frame ultraviolet luminosity 1.0 magnitude fainter. Editor's summary: Astronomical imaging typically combines multiple exposures (coaddition) to improve the signal-to-noise ratio. That procedure is effective at reducing some types of noise but not others. Guo et al. developed a machine learning algorithm for the coaddition process that reduces systematic noise in the faint parts of the image. They verified the performance of this approach using mock sources injected into real observations and real images of the same fields taken with different exposure times. The detection of faint sources improved by about one astronomical magnitude (a factor of about 2.5) without introducing false positives, which was demonstrated in an example application to high-redshift galaxies. —Keith T. Smith INTRODUCTION: The detection limit in observational astronomy determines how faint and distant an object can be observed in an image. It is limited by the signal-to-noise ratio (S/N) produced by the limited exposure time and various noise sources. Coaddition methods can suppress independent noise sources by averaging M exposures (where M is the number of individual exposures), theoretically increasing the S/N by √M. However, this approach produces diminishing returns in deep surveys. Deep learning–based methods have been developed to produce visually cleaner denoised images, but they offer marginal gains in practical astronomical detection of faint sources and have not been tailored to astronomical observations. Improved algorithms and evaluation pipelines could potentially enhance the completeness of faint-source detection in astronomical imaging while preserving high data fidelity. RATIONALE: We developed the Astronomical Self-supervised Transformer-based Denoising (ASTERIS) algorithm, which integrates spatiotemporal information across multiframe exposures with the goal of improving detection limits. Before denoising, the astronomical images are divided into bright and faint parts using a clipping threshold at three SDs (3σ), and ASTERIS focuses exclusively on the faint parts (≤3σ). The adopted self-supervised training scheme does not require an explicit noise model, allowing ASTERIS to be generalized to different telescopes and instruments. To optimize both detection completeness and purity, we adopted a loss function that combines mean squared error and mean absolute error applied to the averaged frame and the corresponding individual frames, respectively. The performance of ASTERIS was tested using both mock source injection and real observational data taken by the James Webb Space Telescope (JWST) and the Subaru Telescope. RESULTS: We quantitatively evaluated the effect of ASTERIS on practical detection limits in astronomical applications using multiple metrics: the photometric S/N, source completeness, source purity, point spread function (PSF) fidelity, and photometric accuracy. Compared with previous methods, ASTERIS improved the detection limit by more than 1.0 magnitude at 90% completeness and 90% purity while preserving the PSF fidelity and photometric accuracy. ASTERIS revealed previously undetectable features in the observational data, and we confirmed that they were real using deeper observations of the same fields. As an example scientific application of ASTERIS, we applied it to a deep JWST imaging field and searched for high-redshift galaxies using the same procedures as previous studies. This process identified 162 candidate galaxies at redshift ≳9, approximately three times more than previous studies using conventional coaddition of the same data. We also constructed the ultraviolet luminosity function of these galaxy candidates, which extended 1.0 magnitude deeper than in previous work. CONCLUSION: Our evaluation pipeline provides quantitative assessment of deep learning–assisted astronomical imaging by considering the practical gains in scientific applications, not just visual improvement. By jointly exploiting temporal and spatial information, ASTERIS reduces noise and extends the detection limits achieved in real JWST data by 1.0 magnitude. Schematic of the ASTERIS denoising algorithm and performance.: Standard coaddition of astronomical images (left) produces noisy images. ASTERIS uses a self-supervised neural network applied to the faint parts of the image. The network integrates spatiotemporal information within a three-dimensional attention–convolution hybrid framework. It is optimized for faint astronomical sources and trained on real data. The resulting completeness and purity of ASTERIS (red lines) were compared with standard coaddition (black lines). [ABSTRACT FROM AUTHOR] |
| Copyright of Science is the property of American Association for the Advancement of Science and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.) | |
| Database: | Psychology and Behavioral Sciences Collection |
|
Full text is not displayed to guests.
Login for full access.
|
|
| Abstract: | The detection limit of astronomical imaging observations is limited by several noise sources. Some of that noise is correlated between neighboring pixels and exposures, so in principle it could be learned and corrected. We present the Astronomical Self-supervised Transformer-based Denoising (ASTERIS) algorithm, which integrates spatiotemporal information across multiple exposures. Benchmarking on mock data indicated 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 the Subaru Telescope identified previously undetectable features, including low-surface-brightness galaxy structures and gravitationally lensed arcs. Applied to deep JWST images, ASTERIS identified three times more redshift ≳9 galaxy candidates than previous methods, with rest-frame ultraviolet luminosity 1.0 magnitude fainter. Editor's summary: Astronomical imaging typically combines multiple exposures (coaddition) to improve the signal-to-noise ratio. That procedure is effective at reducing some types of noise but not others. Guo et al. developed a machine learning algorithm for the coaddition process that reduces systematic noise in the faint parts of the image. They verified the performance of this approach using mock sources injected into real observations and real images of the same fields taken with different exposure times. The detection of faint sources improved by about one astronomical magnitude (a factor of about 2.5) without introducing false positives, which was demonstrated in an example application to high-redshift galaxies. —Keith T. Smith INTRODUCTION: The detection limit in observational astronomy determines how faint and distant an object can be observed in an image. It is limited by the signal-to-noise ratio (S/N) produced by the limited exposure time and various noise sources. Coaddition methods can suppress independent noise sources by averaging M exposures (where M is the number of individual exposures), theoretically increasing the S/N by √M. However, this approach produces diminishing returns in deep surveys. Deep learning–based methods have been developed to produce visually cleaner denoised images, but they offer marginal gains in practical astronomical detection of faint sources and have not been tailored to astronomical observations. Improved algorithms and evaluation pipelines could potentially enhance the completeness of faint-source detection in astronomical imaging while preserving high data fidelity. RATIONALE: We developed the Astronomical Self-supervised Transformer-based Denoising (ASTERIS) algorithm, which integrates spatiotemporal information across multiframe exposures with the goal of improving detection limits. Before denoising, the astronomical images are divided into bright and faint parts using a clipping threshold at three SDs (3σ), and ASTERIS focuses exclusively on the faint parts (≤3σ). The adopted self-supervised training scheme does not require an explicit noise model, allowing ASTERIS to be generalized to different telescopes and instruments. To optimize both detection completeness and purity, we adopted a loss function that combines mean squared error and mean absolute error applied to the averaged frame and the corresponding individual frames, respectively. The performance of ASTERIS was tested using both mock source injection and real observational data taken by the James Webb Space Telescope (JWST) and the Subaru Telescope. RESULTS: We quantitatively evaluated the effect of ASTERIS on practical detection limits in astronomical applications using multiple metrics: the photometric S/N, source completeness, source purity, point spread function (PSF) fidelity, and photometric accuracy. Compared with previous methods, ASTERIS improved the detection limit by more than 1.0 magnitude at 90% completeness and 90% purity while preserving the PSF fidelity and photometric accuracy. ASTERIS revealed previously undetectable features in the observational data, and we confirmed that they were real using deeper observations of the same fields. As an example scientific application of ASTERIS, we applied it to a deep JWST imaging field and searched for high-redshift galaxies using the same procedures as previous studies. This process identified 162 candidate galaxies at redshift ≳9, approximately three times more than previous studies using conventional coaddition of the same data. We also constructed the ultraviolet luminosity function of these galaxy candidates, which extended 1.0 magnitude deeper than in previous work. CONCLUSION: Our evaluation pipeline provides quantitative assessment of deep learning–assisted astronomical imaging by considering the practical gains in scientific applications, not just visual improvement. By jointly exploiting temporal and spatial information, ASTERIS reduces noise and extends the detection limits achieved in real JWST data by 1.0 magnitude. Schematic of the ASTERIS denoising algorithm and performance.: Standard coaddition of astronomical images (left) produces noisy images. ASTERIS uses a self-supervised neural network applied to the faint parts of the image. The network integrates spatiotemporal information within a three-dimensional attention–convolution hybrid framework. It is optimized for faint astronomical sources and trained on real data. The resulting completeness and purity of ASTERIS (red lines) were compared with standard coaddition (black lines). [ABSTRACT FROM AUTHOR] |
|---|---|
| ISSN: | 00368075 |
| DOI: | 10.1126/science.ady9404 |
