🔭 astrophysics

Amplifying the imaging power of digital sky surveys with space telescopes data and generative AI

This paper presents a generative AI method trained on space-based telescope images to enhance the resolution and detail of ground-based digital sky survey images, effectively combining the high throughput of ground surveys with the superior imaging quality of space telescopes.

Original authors: Sai Teja Erukude, Lior Shamir

Published 2026-08-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sai Teja Erukude, Lior Shamir

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

Astronomers have long relied on two very different ways to look at the universe. On one hand, there are massive ground-based surveys that sweep across the sky, capturing millions of images of galaxies every night. These surveys are like wide-angle cameras, excellent for covering vast territories and collecting huge amounts of data quickly. On the other hand, there are space telescopes orbiting above Earth's atmosphere. These instruments act like powerful zoom lenses, capable of seeing faint, distant objects with incredible sharpness and detail that ground telescopes simply cannot match. The problem is that space telescopes can only look at a tiny patch of sky at a time, while ground surveys can see everything but often with a blurry, fuzzy view. For decades, scientists have had to choose between seeing a lot of the sky or seeing it clearly, but not both.

A team of researchers at Kansas State University has found a way to bridge this gap using a type of artificial intelligence known as generative AI. Instead of building new, expensive hardware, they developed a software tool that takes the fuzzy images of galaxies taken by ground-based telescopes and transforms them into sharp, detailed pictures that look as if they were taken by a space telescope. The method works by teaching a computer to recognize the hidden patterns in galaxy shapes. By showing the computer thousands of pairs of images—where one is a blurry ground view and the other is a sharp space view of the exact same galaxy—the system learns how to fill in the missing details. It essentially guesses what the fine structures of a galaxy should look like based on the faint signals it can see, turning a weak, indistinct smudge into a clear, detailed portrait.

The researchers tested this idea using a specific set of data. They gathered 20,000 pairs of galaxy images. In each pair, one image came from the Dark Energy Spectroscopic Instrument Legacy Survey, which uses a ground-based telescope, and the other came from the Hubble Space Telescope, which orbits high above Earth. The ground images were often grainy and lacked fine structure, while the Hubble images showed intricate spirals, bars, and dust lanes. The team trained their artificial intelligence model on these pairs, teaching it to translate the ground-based view into the space-based view. The model uses a specific type of neural network architecture, which functions like a complex filter that amplifies weak signals. It does not just sharpen the image; it reconstructs the shape of the galaxy, adding details that were invisible in the original ground photo but are consistent with how galaxies actually look when viewed with high precision.

When the researchers applied this trained model to new ground-based images, the results were striking. The AI-generated images revealed morphological features, such as the spiral arms or elliptical shapes of galaxies, that were previously hidden in the noise of the ground-based data. To verify the accuracy, the team compared the AI-enhanced images against the actual Hubble images of the same objects. They measured specific properties, such as the length of the galaxy's major axis, its angle, and its overall shape. The measurements from the AI-enhanced images were remarkably close to the real space telescope data, with most differences falling within a five percent margin. While the AI did introduce a small amount of visual "noise," similar to static on an old television screen, this could be smoothed out without losing the important details of the galaxy's structure. In a broader test, the team successfully applied this method to a catalog of over 63,000 galaxies from the ground-based survey, creating a massive new library of high-quality images.

The researchers also explored the limits of this technology by trying to enhance images from the Sloan Digital Sky Survey to match the extreme clarity of the James Webb Space Telescope. In this case, the gap between the blurry ground image and the incredibly sharp space image was too large. The AI struggled to reconstruct the details accurately, showing that the method works best when the difference in quality between the source and the target is not too extreme. This suggests that while the tool is powerful, it is not a magic wand that can create perfect detail out of nothing; it relies on the ground image providing enough of a signal for the AI to recognize the underlying shape.

To make this technology accessible to other scientists, the team released a software tool called "Galaxy Enhancer." This program allows users to input the coordinates of any galaxy, automatically download the ground-based image, and instantly generate an enhanced version using the trained AI model. The researchers have also made the entire dataset of enhanced images and the code available to the public. This approach offers a new way to maximize the power of existing astronomical surveys. It means that the vast archives of ground-based images, which cover huge portions of the sky, can now be viewed with a level of detail that was previously reserved for the most expensive space missions. By combining the wide coverage of Earth-based telescopes with the sharp vision of space telescopes through software, astronomers can now study the shapes and structures of millions of galaxies with a clarity that was previously impossible to achieve.

Technical Summary: Amplifying the Imaging Power of Digital Sky Surveys with Space Telescopes Data and Generative AI

Problem Statement
Digital sky surveys (e.g., SDSS, DES, Vera Rubin Observatory) offer high throughput and extensive sky coverage but generally produce images with inferior resolution and detail compared to space-based telescopes (e.g., HST, JWST). Conversely, space telescopes provide exceptional imaging power and deep Universe coverage but lack the throughput to survey large areas of the sky efficiently. The authors posit that an ideal astronomical imaging system would combine the high throughput of ground-based surveys with the image quality of space-based telescopes. Currently, achieving this requires substantial hardware investment or the limitation of observing only specific deep fields.

Methodology
The paper proposes a computational solution using a Conditional Generative Adversarial Network (cGAN) to elevate the quality of ground-based galaxy images to the level of space-based observations. The method does not require new hardware or optics.

Data: The training dataset consists of 20,000 paired images. The target domain images are from the Hubble Space Telescope (HST) COSMOS survey, while the source domain images are from the Dark Energy Spectroscopic Instrument (DESI) Legacy Survey, centered on the same coordinates. The authors utilized both JPG and FITS formats, noting that while JPG is common for GAN inputs, FITS preserves the high dynamic range necessary for the AI to identify subtle patterns in galaxy shapes.
Pre-processing: Images were converted from FITS to 16-bit TIF format to preserve pixel value differences, resized to 256×256 pixels, and concatenated side-by-side (ground/space) for training.
Architecture: The system employs a Pix2Pix framework (Isola et al., 2018).
- Generator: Based on a U-Net architecture (Ronneberger et al., 2015) with a contracting encoder and expansive decoder connected by skip connections. This design allows the combination of high-level information with low-level details to amplify visual information.
- Discriminator: Based on a PatchGAN architecture (Demir and Unal, 2018), which classifies $N \times N$ patches rather than full images, aiding in the preservation of high-frequency details.
Training: The model was trained using the Adam optimizer (learning rate 0.0002, batch size 1) for 100 epochs. The loss function is a weighted sum of adversarial loss ( $L_{gan}$ ) and L1 loss ( $L_1$ ), with a weighting factor $\lambda = 0.5$ .
Inference: Once trained, the generator can take a ground-based image as input and output an enhanced image approximating the space-based target.

Key Contributions

Generative AI Framework: A trained cGAN model capable of converting weak signals from ground-based telescope images into detailed images resembling space-based observations.
Enhanced Galaxy Catalog: The application of this method to 63,202 galaxies from the DESI Legacy Survey, creating a public catalog of enhanced images available for download.
Software Tool: The release of "Galaxy Enhancer," a unified software tool that encapsulates the entire pipeline, including a module to download ground-based imagery via celestial coordinates and generate enhanced outputs in real-time.
Open Data and Code: Provision of the training dataset, the enhanced catalog, and the source code via public repositories (Figshare and GitHub).

Results

Visual Quality: Qualitative analysis shows the model successfully reconstructs morphological features of galaxies, making enhanced images visually similar to HST targets. However, the authors note the presence of "pepper noise" (background artifacts) typical of AI-generated images, which may require post-processing (e.g., low-pass filtering) that could potentially alter relevant visual content.
Quantitative Analysis: Using SExtractor and Ganalyzer on 1,000 test pairs, the authors measured descriptors such as major/minor axis, position angle, elongation, and ellipticity. The average relative difference between descriptors computed on HST images and AI-enhanced images ranged from approximately 3.1% to 5.6%, with Major axis (0.056) and Minor axis (0.052) exceeding the 5% threshold.
Morphological Accuracy: In the test set of 1,000 galaxies, 45 cases (4.5%) showed a mismatch in broad morphology (elliptical vs. spiral) between the AI-enhanced image and the HST ground truth. This corresponds to a relative difference of 0.06 (6%) for the broad morphology descriptor as listed in Table 4.
Limitations in Quality Gap: An experiment enhancing SDSS images to match JWST deep field images (Stephan's Quintet) demonstrated that the method's effectiveness diminishes as the quality gap between the source and target instruments widens. The AI struggled to match the extreme detail of JWST from the lower-resolution SDSS inputs, unlike the more moderate gap between DESI and HST.

Significance and Claims
The paper claims that this method provides a computational solution to enhance galaxy images without the need for new hardware. By combining the footprint and throughput of ground-based surveys with the imaging power of space telescopes, the method allows for the analysis of galaxy shapes on a large scale. The authors emphasize that the tool can be used for "on-the-fly" enhancement, allowing users of digital sky surveys to improve objects of interest in real-time.

The authors maintain a modest stance regarding the technology's limitations. They acknowledge that the system is trained on extended objects (galaxies) and may not be suitable for point sources. Furthermore, they caution that for rare objects or cases where the training data does not cover specific patterns, the generative model might produce images inconsistent with the true visual appearance of the object. The method is presented as a way to maximize the discovery power of existing and future digital sky surveys without the resource investment typical of space-based missions.

More like this