Source-linked AI summary
Amplifying the imaging power of digital sky surveys with space telescopes data and generative AI
Sai Teja Erukude, Lior Shamir
TL;DR
Ground-based digital sky surveys provide high-throughput, wide-coverage imaging but lower image quality than space telescopes. The paper uses generative AI trained on paired observations to enhance galaxy images, producing detailed reconstructions while combining survey throughput with space-telescope imaging quality.
Problem
Ground-based surveys have high data-collection bandwidth but imaging quality that remains below space-based telescopes, whereas space telescopes have lower throughput.
Method
A cGAN learns from paired ground-based and space-based images of the same galaxies to amplify weak signals and generate enhanced galaxy images.
Results
45 of 1,000 tested galaxies had broad morphology that did not match the corresponding space-telescope image, while descriptors normally stayed within a 5% difference.
Takeaways & Limitations
The method can generate large catalogs of enhanced galaxy images and is provided as software for on-the-fly transformation during digital sky-survey browsing.
Takeaways & Limitations
Performance is limited when the image-quality gap between source and target increases, and filtering generated pepper noise can alter relevant visual content.
Abstract
from arXiv · showhide
While Digital sky surveys provide excellent throughput of image data and can cover a large footprint, their imaging power is normally inferior to that of space-based telescopes. Space-based telescopes, on the other hand, provide excellent imaging power and can image the deep Universe, but cannot provide the same throughput as advanced ground-based sky surveys. Here, we utilize generative AI to elevate the quality of galaxy images taken by ground-based telescopes to the level of details enabled by space telescopes. The solution is based on the nature of galaxy shapes, allowing generative AI trained on space-based images to convert weak signal into detailed and clear galaxy images. The method allows for combining the high throughput of ground-based sky surveys with the image quality of space-based telescopes. The source code for the method is available, as well as paired training data and a catalog of 63,202 galaxy images enhanced by the proposed method. We also provide a software tool that encapsulates the entire pipeline and the custom generative AI model to generate galaxy images with enhanced quality.
1 INTRODUCTION
Digital sky surveys offer broad, high-throughput imaging, while space telescopes provide substantially higher image quality. The paper proposes generative AI to enhance ground-based galaxy images by recovering weak signals and detailed morphology.
- Digital sky surveys continuously collect extremely large astronomical datasets across broad sky coverage.Examples include Pan-STARRS, HSC, SDSS, Vera Rubin Observatory, and DES.
- Space telescopes provide unprecedented-quality astronomical images, but their throughput is lower than that of Earth-based digital sky surveys.
- The paper seeks to combine ground-based survey throughput and coverage with space-telescope image quality for analyzing galaxy shapes.
- Generative AI is used to elevate Earth-based galaxy images by identifying weak signals and amplifying them into clear image details.The model learns from images of the same objects acquired by Earth-based and space-based telescopes.
- The proposed approach provides a computational enhancement solution without requiring new hardware or optics.Applied to digital sky surveys, it is intended to combine their footprint and throughput with space-telescope imaging power.
2 DATA
The training data pair DESI Legacy Survey images with HST images of the same galaxies. The data are prepared through source detection, image extraction, format conversion, and preprocessing steps for cGAN training.
- The cGAN is trained to generate an HST image from a paired DESI Legacy Survey image of the same astronomical object.The HST image serves as the target image, while the ground-based image serves as the source.
- 20,000 galaxy objects were assembled from HST COSMOS data and paired ground-based observations.Sources were detected with SExtractor, separated using Montage, and converted from 122 × 122-pixel FITS images to 16-bit TIF format.
- Figure 1 presents paired DESI Legacy Survey and HST images of the same objects, with the HST images providing more visible galaxy-shape detail.
- Figure 2 summarizes the preprocessing pipeline's inputs, output, and key transformation steps.
- The preprocessing pipeline converts FITS files into 16-bit TIF images by averaging pixel values along axis 0 to produce single 2D grayscale images.The 16-bit range preserves subtle pixel-value differences during format conversion.
3 A CGAN-BASED METHOD FOR AMPLIFYING THE IMAGING POWER OF GROUND-BASED TELESCOPES
The method uses a conditional GAN to learn mappings from ground-based galaxy images to paired space-based images, combining adversarial training with pixel-level reconstruction. Its U-Net generator and PatchGAN discriminator are trained on paired data to enhance weak signals and preserve fine details.
- cGAN architecture: A conditional GAN learns to convert ground-based telescope images into detailed images of the same galaxies captured by space telescopes.Each training sample pairs a DESI Legacy Survey image with its corresponding HST image.
- Generator: The generator uses a U-Net with encoder–decoder paths and skip connections to combine high-level information with low-level details.This design is intended to amplify visual information into fine details characteristic of space-based images.
- Discriminator: The discriminator uses PatchGAN classification of N×N image patches to enhance high-frequency information and preserve fine details.Unlike full-image classification, the discriminator evaluates local patches.
- Training objective: The generator objective combines adversarial loss and L1 loss as Lg = Lgan + λ · L1, with λ set to 0.5.Lgan is the discriminator loss and L1 is the generator loss as described in the method.
- Training procedure: Training uses paired ground-based and space-based images, with the Pix2Pix cGAN procedure iteratively updating the discriminator and generator.The framework includes separate training and inference algorithms for producing enhanced outputs from ground-telescope images.
4 RESULTS
The results evaluate enhanced galaxy images through visual comparison, automated morphology measurements, a large catalog, and an integrated software tool. The method reconstructs galaxy features, generally preserves measured descriptors within 5%, but shows limitations from noise and large source-to-target quality gaps.
- Visual evaluation: The GAN reconstructs galaxy morphological features in DESI Legacy Survey images, producing detail similar to corresponding Hubble Space Telescope images.Figure 5 compares ground-based inputs, generated outputs, HST images, and residual maps.
- Limitations: Enhancement can add background “pepper noise,” while smoothing it with a low-pass filter may alter relevant visual content.The paper advises applying noise removal with awareness that visual information can be affected.
- Automated evaluation: Descriptors from 1,000 AI-enhanced images normally stayed within 5% of measurements from corresponding space-telescope images.The evaluated descriptors included major axis, minor axis, position angle, elongation, and ellipticity.
- Automated evaluation: 45 of 1,000 tested galaxies had a broad morphology classification that differed between the AI-enhanced and space-telescope images.Morphology was assessed using elliptical-versus-spiral classifications.
- Catalog: The catalog contains enhanced images and side-by-side comparisons of DESI inputs, AI outputs, and corresponding space-based ground-truth images.Enhanced images are stored as 256×256-pixel PNG files, and the catalog is based on galaxies within the COSMOS footprint.
- Galaxy Enhancer Software: The Galaxy Enhancer combines user input, ground-based imagery download, and enhancement modules into one integrated tool.The tool is intended to simplify use of the complete enhancement pipeline.
- Experiments with SDSS and JWST galaxies: AI-enhanced SDSS images matched JWST images less closely than DESI Legacy Survey images enhanced toward COSMOS images when the source-to-target quality gap was larger.The paper identifies the increasing image-quality gap as a limitation on enhancement performance.
5 CONCLUSION
Generative AI enhances Earth-based galaxy images while preserving the throughput and broad coverage of digital sky surveys. The method supports large enhanced-image catalogs and fast, on-the-fly use, but remains limited by source-image quality and galaxy type.
- Generative AI enhances Earth-based images toward space-telescope image quality, addressing the lower imaging quality of digital sky surveys.The conclusion describes a pix2pix GAN trained on paired Earth-based and space-based galaxy images.
- Enhancement quality decreases when the source-to-target image-quality gap is large, as illustrated by the weaker SDSS-to-JWST result.The authors attribute this limitation to the relatively low detail in SDSS images compared with JWST images.
- The method was applied to a large galaxy catalog and can transform selected objects in real time through a software tool.This supports both pre-generated enhanced catalogs and interactive image enhancement while browsing survey data.
- The method is limited to extended objects and may produce visually inconsistent results for rare galaxies because it relies on repetitive galaxy-shape patterns.The authors also note that machine-learning errors can make transformed images differ from an object’s true visual appearance.
- The approach combines digital sky surveys’ broad coverage and throughput with the imaging quality of space-based telescopes without requiring new hardware or optics.It is presented as a computational alternative to the resource investment associated with space-based telescopes.
- The method’s speed and low footprint allow integration into existing and future digital sky surveys for on-the-fly enhancement without dedicated catalogs.The conclusion frames this as a way to maximize survey discovery power.
Funding
The research received partial support from NSF grant OIA-2148878.
- The research was supported in part by NSF grant OIA-2148878.
Conflicts of Interest
The authors report no conflict of interest regarding publication of the article.
- The authors declare no conflict of interest regarding publication of the article.
Data Availability
The enhanced-galaxy catalog, cGAN training data, and project code are publicly available through linked repositories.
- The catalog of enhanced galaxy images and cGAN training data can be downloaded from Figshare.
- The project code is available on GitHub.