🔭 astrophysics

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

This work presents ASTERIS, a self-supervised, transformer-based algorithm that leverages spatiotemporal correlations across multiple exposures to significantly improve astrophysical detection limits, thereby enabling the discovery of fainter galaxies and low-surface-brightness structures in deep imaging data from telescopes such as JWST and Subaru.

Original authors: 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

Published 2026-05-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: 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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to hear a single, tiny whisper in a crowded stadium. The murmur of the crowd is the "noise," and the whisper is the faint light of a distant, ancient galaxy. For decades, astronomers have tried to hear these whispers simply by asking the crowd to be quieter or by recording the stadium for longer. Yet sometimes the noise is not just random murmuring; it is a complex, swirling pattern standing in the way.

This post introduces a new tool called ASTERIS, an intelligent computer program designed to hear these whispers much better than before. Here is how it works, simply explained:

The Problem: The "Noise" in the Image

When telescopes like the James Webb Space Telescope (JWST) take images of the universe, they do not simply take a single photo. They take many photos of the same location and stack them on top of each other, as if layering tracing paper. This is called "co-addition."

Imagine taking a blurry photo of a moving car. If you take 10 photos and stack them, the car becomes clearer, but the static background noise also becomes slightly clearer. Normally, astronomers assume this noise is just random "snow" on a television screen. In reality, however, the noise is often "structured"—it has patterns, like ripples in a pond or the hum of an engine. Conventional stacking methods cannot fix this structured noise because they treat every pixel as if it were independent.

The Solution: ASTERIS (The "Time-Traveling Detective")

The authors developed ASTERIS, which stands for Astronomical Self-supervised Transformer-based Enhanced Reduction of Image Static (a fancy name for an intelligent noise reducer).

Instead of simply stacking photos, ASTERIS acts like a detective examining the same scene from slightly different angles and at slightly different times.

The Analogy: Imagine trying to find a specific person in a crowd. You have 8 different photos of the same crowd, taken a fraction of a second apart. In each photo, the crowd shifts slightly, but the person you are looking for remains in the same spot.
How it works: ASTERIS considers all 8 photos together. It learns that the "person" (the real galaxy) remains consistent, while the "crowd movement" (the noise) varies in a specific, predictable way. It uses a special type of artificial intelligence (a "Transformer," similar to the technology behind modern chatbots) to figure out exactly what the noise looks like and subtract it, leaving the galaxy crystal clear.

The Magic of "Self-Supervision"

Normally, to teach a computer to remove noise, you need a "clean" photo to show it what the result should look like. Yet in astronomy, we never have a clean photo of a deep, faint galaxy; that is precisely why we take images in the first place!

ASTERIS is self-supervised. It teaches itself.

The Analogy: Imagine you have two groups, each with 8 photos. You tell the computer: "Use Group A to guess what the real image looks like, and check your guess against Group B." Since the real galaxy is present in both groups but the noise is different, the computer learns to ignore the noise and keep the galaxy. It needs no "perfect" answer key; it figures it out by comparing its own work with the other photos.

What Did They Find?

The team tested this on real data from the JWST and the Subaru Telescope (a massive telescope on a mountain in Hawaii).

Seeing the Invisible: They found that ASTERIS could see galaxies that were previously invisible. It improved the "detection limit" by about 1.0 magnitude. In astronomy, this is a big deal—it is like being able to see a candle from twice the distance.
More Galaxies, Fewer Errors: When they searched for the oldest, most distant galaxies (from the time when the universe was still a baby), ASTERIS found three times more candidates than previous methods.
No Blur: Some older methods would smooth out the noise but also blur the image, making stars look like fuzzy smudges. ASTERIS kept the images sharp and preserved the true shape of the galaxies.
Proof in Practice: They did not test this only on fake data. They applied it to the "JADES Origins Field," a very deep patch of sky. They found faint, red galaxies hiding in the noise. One of these galaxies is so distant (redshift ~17) that its light has been traveling for over 13 billion years.

Why This Matters

This is not just about making prettier images. By cleaning the noise without blurring the image, ASTERIS enables astronomers to:

See fainter objects without waiting months to get more telescope time.
Study the very first galaxies formed after the Big Bang with much greater confidence.
Distinguish between a real faint galaxy and a random error in the camera.

In short: ASTERIS is like a pair of glasses for astronomers that filters out the "noise" of the universe, finally allowing them to hear the faint whispers of the earliest moments of the cosmos.

Technical Summary: Deeper Detection Limits in Astronomical Imaging via Self-Supervised Spatiotemporal Denoising

Problem Statement

The detection limit in observational astronomy is fundamentally constrained by noise sources, including instrumental readout noise, sky background, and photon statistics. While hardware advancements (e.g., larger apertures) improve the signal-to-noise ratio (S/N), they cannot correct correlated background noise arising from zodiacal emission, galactic foreground sources, and atmospheric scattering.

Conventional software approaches face significant limitations:

Co-addition (Stacking): Averaging multiple exposures improves S/N only by a factor of $\sqrt{M}$ if pixel noise is independent and identically distributed (i.i.d.). In reality, noise is often correlated due to the point spread function (PSF), structured background emission, and instrumental effects, leading to diminishing returns in deep surveys.
Smoothing: Convolution with kernels suppresses high-frequency noise but degrades spatial resolution.
Supervised Deep Learning: Methods requiring pairs of noisy and clean images struggle with the gap between synthetic and real domains, as well as the scarcity of reliable paired training data (e.g., long/short exposure pairs).
Existing Self-Supervised Methods: Current single-image or pair-based self-supervised denoising methods (e.g., Noise2Noise) often rely on general image S/N metrics that do not translate into practical astronomical gains. These methods can introduce artifacts, miss faint sources, or compromise photometric accuracy.

Methodology: ASTERIS

The authors present ASTERIS (Astronomical Self-supervised Transformer-based denoising algorithm), a self-supervised deep learning framework designed to exploit spatiotemporal correlations across multiple dithered exposures of the same field.

Core Architecture and Strategy

Spatiotemporal Learning: Unlike standard single-image denoising, ASTERIS processes a stack of $M$ aligned exposures as a 3D data cube (spatial $x, y$ + temporal $t$ ). It utilizes a 3D U-Net architecture extended by 3D Multi-Deconvolved-Head Transposed Attention (3D-MDTA) blocks and 3D Gated-Deconvolved Feed-Forward Network (3D-GDFN) blocks. This hybrid design captures both local spatial correlations and long-range dependencies across exposures.
Self-Supervised Training: The algorithm requires no clean ground truth. It splits a set of $M$ exposures (typically 16) into two independent subsets: an input set and a target set (e.g., 8 exposures each). Since the exposures are astrometrically aligned, they share the same underlying signal but possess independent noise realizations. The network learns to predict the signal expectation of the input set using the target set as a proxy.
Loss Function: The training objective combines two components:
1. Mean Loss (MSE): Minimizes the mean squared error between the co-addition of the input set and the target set, improving source detection completeness by reducing background fluctuations.
2. Image Loss (MAE): Minimizes the mean absolute error between individual corresponding exposures, providing robustness against outliers and improving detection purity.
Flux Thresholding (3 $\sigma$ -Clipping): To handle the large dynamic range of astronomical images, ASTERIS selectively operates on the faint parts of the image. Pixels below a $3\sigma$ significance threshold are processed by the neural network. Pixels above this threshold (bright sources) are temporarily masked (clipped), median-combined, and reintegrated into the final image. This focuses the network's capacity on low S/N regions without introducing flux discontinuities.
Instrument Adaptation: Separate models are trained for different instruments (JWST NIRCam short/long wavelengths and Subaru MOIRCS) to account for varying noise characteristics and PSFs.

Key Results

1. Benchmarking with Simulated Data

Using JWST NIRCam F115W data with injected simulated sources (27.5–31.5 mag), ASTERIS was compared against standard co-addition, Gaussian smoothing, BM3D/BM4D, Neighbor2Neighbor, and Noise2Noise (N2N).

Detection Limits: At 90% completeness and 90% purity, ASTERIS improves the detection limit by 1.0 magnitude compared to standard co-addition and by 1.0 magnitude compared to N2N.
F-Score: ASTERIS improves the limiting depth (at F-Score = 0.9) by approximately 1.7 mag over co-addition.
Photometric and Spatial Accuracy: ASTERIS preserves the PSF profile (statistically indistinguishable from co-addition, $p=0.9$ ) and shows no systematic photometric distortion, unlike N2N, which exhibited degraded resolution ( $p < 0.05$ ).

2. Observational Validation

JWST Data (GLIMPSE Program): Applied to an 8-exposure stack, ASTERIS recovered 93 of 168 sources detected in a deep ground truth with 168 exposures, whereas standard co-addition with 8 exposures recovered only 50. Applied to the full set of 168 exposures, ASTERIS identified 35% more sources than standard methods.
Low Surface Brightness (LSB) Structures: ASTERIS significantly improved the Structural Similarity (SSIM) metric for extended structures. For the outer arms of a spiral galaxy, SSIM increased from 0.37 (co-addition) to 0.67. For a gravitationally lensed arc, SSIM improved from 0.59 to 0.81.
Generalization (Subaru Telescope): A model specifically trained on Subaru MOIRCS data successfully recovered faint sources in ground-based Ks-band images, demonstrating adaptability to atmospheric turbulence and higher sky background, although a JWST-trained model applied directly to Subaru data produced false positives.

3. Application to High-Redshift Galaxies (JADES Origins Field)

ASTERIS was applied to the JADES Origins Field (JOF), a deep JWST survey.

Candidate Discovery: The method identified 162 high-redshift galaxy candidates at $z_{phot} \gtrsim 9$ , approximately three times as many as found in previous studies using standard co-addition.
Fainter Limits: The faintest identified candidates had rest-frame absolute UV magnitudes ( $M_{UV}$ ) of $\approx -16$ mag, 1.0 magnitude fainter than previous limits.
Resolution of Redshift Ambiguities: By improving S/N in "dropout" bands (filters where the Lyman break leads to non-detection), ASTERIS helped resolve photometric redshift ambiguities, distinguishing high-redshift Lyman breaks from low-redshift Balmer breaks.
Luminosity Functions: The derived UV luminosity functions for $z \sim 9-12$ and $z \sim 12-16$ exhibit steep faint-end slopes ( $\alpha \approx -2.45$ and $-2.28$, respectively), indicating higher number densities of faint galaxies than predicted by some theoretical models.

Significance and Claims

The work claims that ASTERIS represents a significant advance in astronomical imaging by exploiting spatiotemporal correlations inherent in multiple exposures, which are typically ignored by standard stacking or single-image denoising methods.

Efficiency: The algorithm enables the extraction of deeper scientific results from existing datasets without requiring additional telescope time or hardware upgrades.
Robustness: It effectively suppresses correlated noise while preserving the PSF and photometric accuracy, which is crucial for morphological and flux measurements.
Scientific Impact: By pushing detection limits deeper, ASTERIS enables the discovery of previously undetectable low surface brightness structures and a substantial population of faint high-redshift galaxies, providing new constraints on early galaxy formation and the epoch of reionization.

The authors emphasize that the method is self-supervised and requires only the observational data itself for training, making it applicable to diverse survey strategies and instruments, provided suitable training data are available.