Paper deep dive

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

Sai Teja Erukude, Lior Shamir

Year: 2026Venue: arXiv preprintArea: astro-ph.IMType: PreprintEmbeddings: 42

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%

Last extracted: 8/24/2026, 5:01:40 AM

Summary

This paper presents a method to enhance the imaging quality of ground-based digital sky surveys (specifically DESI Legacy Survey) to match the resolution of space-based telescopes (Hubble Space Telescope) using a Conditional Generative Adversarial Network (cGAN). The authors trained a Pix2Pix model on paired images of 20,000 galaxies and generated a catalog of 63,202 enhanced galaxy images. Evaluation via SExtractor and Ganalyzer shows that the enhanced images preserve morphological features with less than 5% relative difference in key descriptors compared to actual HST images.

Entities (10)

DESI Legacy Survey · survey · 99%Hubble Space Telescope · instrument · 99%Conditional Generative Adversarial Network · model-architecture · 98%Adam optimizer · algorithm · 95%PatchGAN · neural-network-component · 95%Pix2Pix · model-variant · 95%SExtractor · software-tool · 95%U-Net · neural-network-component · 95%COSMOS · datasetfield · 90%Ganalyzer · software-tool · 90%

Relation Signals (7)

Pix2Pix → generated → 63,202 galaxy images

confidence 95% · The catalog is available for download... The method was applied to 63,202 galaxies from the DESI Legacy Survey.

Pix2Pix → uses → U-Net

confidence 95% · Model architecture Pix2Pix conditional GAN (U-Net generator, PatchGAN discriminator)

Pix2Pix → uses → PatchGAN

confidence 95% · Model architecture Pix2Pix conditional GAN (U-Net generator, PatchGAN discriminator)

Pix2Pix → trainedon → DESI Legacy Survey

confidence 90% · Each training sample is a pair of images, where one image is the ground-based telescope image provided by the DESI Legacy Survey.

Pix2Pix → trainedon → Hubble Space Telescope

confidence 90% · The other image is the image of the same galaxy captured by the space-based HST, which is the target image that the cGAN is trained to generate

SExtractor → usedforanalysisof → Hubble Space Telescope

confidence 85% · The analysis was first applied to the space-telescopes images... we used the mature and commonly used SExtractor

SExtractor → usedforanalysisof → Pix2Pix

confidence 85% · The analysis was first applied to the space-telescopes images, and then the results of each image were compared to the results when applying the same analysis to the AI-enhanced galaxy images.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract: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.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

39,694 characters extracted from source content.

Expand or collapse full text

Amplifying the imaging power of digital sky surveys with space telescopes data and generative AI2026Amplifying the imaging power of digital sky surveys with space telescopes data and generative AI–References Sai Teja Erukude Affiliation: Kansas State University, Manhattan, KS, 66506, USA Lior Shamir Thanks: E-mail: lshamir@mtu.edu Affiliation: Kansas State University, Manhattan, KS, 66506, USA Accepted x x x. Received x x x; in original form x x x Abstract 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. Keywords: techniques: image processing – methods: data analysis – telescopes. 1 Introduction Digital sky surveys have had a transformative impact on astronomy research (24; 27; 13; 20; 37). Powered by robotic telescopes, digital sky surveys image the sky continuously, collecting and storing image data. These data can be accessed by the public through the concept of virtual observatory. Sky surveys such as the Panoramic Survey Telescope and Rapid Response System (Pan-STARRS) (22), the Hyper Suprime-Cam (HSC) (1), Sloan Digital Sky Survey (SDSS) (38), the Vera Rubin Observatory (21), and the Dark Energy Survey (DES) (9) continuously image the sky and collect extremely large astronomical data. Another revolutionary astronomical research instrument is the space telescope. Space telescopes such as the Hubble Space Telescope (HST), the James Webb Space Telescope (JWST), Euclid (28), and Roman (36), have been providing image data of astronomical objects with unprecedented quality to transform our understanding of the Universe. While the imaging power of space-based telescopes cannot be met by Earth-based telescopes, the throughput of space-based telescopes is lower compared to Earth-based digital sky surveys such as the Vera Rubin Observatory. Therefore, an ideal astronomical imaging device would be a combination of the throughput and sky coverage of Earth-based digital sky surveys, with the image quality of space-based telescopes. Such a system will allow for imaging a large number of objects while also providing the ability to analyze their shape (32; 12; 31; 8; 15; 14). Here, we use the concept of generative AI to elevate the image quality of Earth-based telescopes to the level of space-based telescopes. Generative AI has been used in image computing for numerous purposes, ranging from the generation of “deep fake” images to the creation of computer art. Among other tasks, it allows for generating a synthetic image based on information learned directly from the target images. The concept of generative AI has also been used to generate synthetic galaxy images (7; 16; 26), or improving the resolution of low images (18). By using a very large set of images of the same objects taken by both Earth-based and space-based telescopes, we develop a generative AI model that elevates the quality of the Earth-based images. The model identifies the weak signal in galaxy images and amplifies that signal through a generative AI system that is sensitive to the patterns and the visual content of the object. In that sense, the generative AI is used as a complex filter that enhances weak signal to turn it into clear image details. That provides a computational solution to the enhancement of galaxy images taken by Earth-based telescopes, with no need for new hardware or optics. Its application to digital sky surveys can combine the footprint and throughput of Earth-based digital sky surveys with the imaging power of space telescopes. 2 Data The proposed solution is based on a Conditional Generative Adversarial Network (cGAN) that elevates a ground-based image to a quality equivalent to an image taken by a space telescope. Therefore, the data required for training should include paired images, where one image is acquired by a ground-based instrument and the other shows the exact same astronomical object, but taken by a space telescope. Here, the space-based image data are taken from the Hubble Space Telescope (HST), and the ground-based image data are taken by the Dark Energy Spectroscopic Instrument (DESI) Legacy Imaging Survey (11). Although modern digital sky surveys generate very large databases, the requirement to have galaxies imaged by both the space-based telescope and the ground-based telescope reduces the total size of images that can be used for such a training dataset. The HST images (34) were taken from the Cosmological Evolution Survey (COSMOS) (30). The sources were detected by applying SExtractor (5) and selecting sources with a magnitude of 4σ or higher compared to the background. The sources were then separated by using the Subimage tool of Montage (4). The images were FITS format images of dimensionality 122 × 122 pixels, and these images were converted to the simpler TIF format with a dynamic range of 16-bit for the image processing. The entire dataset contained 20,000 objects. Each HST image was paired with an image taken from the DESI Legacy Survey centered at the same coordinates. That was done by retrieving image data from the DESI Legacy Survey at the same coordinates as the COSMOS images. Images were retrieved in both the JPG and FITS formats using the cutout API of the ninth data release (DR9). The JPG format has the advantage of size and is also a highly common format used as input for GAN models. On the other hand, the FITS format provides much more detailed information that will be needed to reconstruct the shape of the object from a weak signal. While the human eye might not always be sensitive to the high dynamic range enabled by the FITS format, the information can be used by the AI model to identify subtle patterns in the galaxy shapes that can be observed clearly when using space-telescope images. Because the physics of galaxy formation leads to repetitive shapes, the identification of such patterns can enable the cGAN to reconstruct the fine details of the shape of the galaxy. Both formats were tested, which involved using two separate datasets of paired images in two distinct experiments. Figure 1 shows examples of the galaxy images taken by DESI Legacy Survey and the corresponding galaxy image taken by the HST. As the figure shows, the HST images are far more detailed compared to the images taken by the DESI Legacy Survey, providing more information about the shape of the galaxies that cannot be seen by observing the images acquired by the Earth-based telescope. Figure 1: Pairs of training images used to train the cGAN. Each pair includes an image taken by the DESI Legacy Survey (left) and an image of the same object taken by the HST (right). 2.1 Data pre-processing pipeline The DESI Legacy Survey and Hubble Space Telescope images are first passed through a pre-processing pipeline to prepare them for training. This pipeline consists of several key steps, as illustrated in Figure 2. It begins with converting the original FITS files into 16-bit TIF format by averaging pixel values along the first dimension (axis 0) to produce a single 2D grayscale image. The purpose of the conversion to the TIF format was to use a format compatible with the architecture, while the 16-bit dynamic range ensured that subtle pixel value differences are preserved through the format conversion. Next, all images are resized to 256×256 pixels. Finally, each ground-based image is concatenated with its corresponding space telescope image counterpart side by side. This results in a training dataset comprising 20,000 image pairs used to train the cGAN. The training images are available in “train.zip” at https://doi.org/10.6084/m9.figshare.30226591. Figure 2: Overview of the data pre-processing pipeline, illustrating the inputs, the output, and key transformation steps involved. 3 A cGAN-based method for amplifying the imaging power of ground-based telescopes A Generative Adversarial Network 17 is a deep learning concept used to generate synthetic data designed to mimic original data. A GAN is made of two neural networks: the generator and the discriminator. The generator converts input data into synthetic data optimized to be indistinguishable from real data. The discriminator acts like a regulator, attempting to identify the synthetic data from the real data. It tries to classify between the generated sample and the real sample to estimate the differences between them. The generator learns from the discriminator’s feedback to produce samples that are more similar to the real samples. This is done iteratively, with the generator and discriminator both improving through the process. We use the Conditional Generative Adversarial Network (cGAN) to improve the quality of ground-based images. The GAN can learn from the patterns of galaxy shapes in the ground-based telescope images by comparing them to the detailed space-based images of the same objects. By pairing each ground-based image with the space-based image, the GAN learns how to convert weak signal that is difficult to observe by eye into the detailed full shape as observed by a powerful space telescope. For that purpose, we use the paired cGAN approach of image-to-image conversion 19. The networks are trained such that the input pair is an image taken by the ground-based telescope, paired with an image of the same object taken by the space telescope. Through training, the system learns the links between the weaker patterns identified in the Earth-based images and their corresponding detailed shapes in the space telescope images. After the system is trained to identify the links between the paired images, it can enhance new ground-based images and turn them into detailed images of quality typical of space telescopes. 3.1 cGAN Architecture The generator is based on the U-Net-based architecture (29), which has been found effective also for astronomy images (15). The U-shaped structure is made of a contracting path (encoder) and an expansive path (decoder), with skip connections between the corresponding layers and the encoder and decoder paths, as shown in Figure 3. This design allows the combination of high-level information with low-level details, and can therefore amplify the visual information into the fine details of space-based telescopes. Figure 3: U-Net architecture of the generator used in the enhancement of ground-based telescope images. The discriminator is based on a PatchGAN architecture (10), as shown in Figure 4. The PatchGAN discriminator is trained and classifies N×N patches of the input images rather than the full images. This design is effective for enhancing high-frequency information, helping to preserve the fine details in the generated images. Figure 4: PatchGAN-based architecture of the discriminator. The loss LgL_g is the commonly used weighted sum of the adversarial loss LganL_gan and the L1L_1 loss, as shown by Equation 1. LganL_gan is the discriminator loss, and l1l_1 is the generator loss. λ is a hyperparameter set to 0.5. Lg=Lgan+λ⋅L1.L_g=L_gan+λ· L_1. (1) 3.2 Training the model The model was trained using the dataset of image pairs described in Section 2. Each training sample is a pair of images, where one image is the ground-based telescope image provided by the DESI Legacy Survey. The other image is the image of the same galaxy captured by the space-based HST, which is the target image that the cGAN is trained to generate automatically from the source image. We trained the model using the Adam optimizer (23) with a learning rate of 0.0002, a batch size of 1, and a total of 100 epochs. For transparency, all training hyperparameters are listed in Table 1. Table 3 shows the number of parameters used in each of the models. Unless otherwise noted, we report results using the model checkpoint from epoch 60, as discussed in Section 4. All experiments were executed on a high-performance computing system, with the hardware and software details provided in Table 2. Table 1: Training configuration summary for the proposed Pix2Pix-based generative AI framework. Component Setting Model architecture Pix2Pix conditional GAN (U-Net generator, PatchGAN discriminator) (19) Generator backbone U-Net encoder C64–C512; decoder CD512–C64 (29) Discriminator PatchGAN, C64–C512, Sigmoid patch output (10) Input resolution 256×256×3256× 256× 3 RGB tiles (ground/space pairs) Dataset type Paired image translation (Ground → Space telescope) Normalization Pixel values scaled from [0, 255] to [-1, 1] Batch size 1 Optimizer Adam, learning rate 0.00020.0002, β1=0.5 _1=0.5 Loss functions Adversarial BCE + L1 (MAE) reconstruction Loss weighting Generator loss: BCE : L1 = 1 : 100 Training epochs 100 Model saving Periodic generator checkpoints (every 10 epochs) Implementation framework TensorFlow / Keras Repository https://github.com/SaiTeja-Erukude/Enhancing-Ground-Based-Astronomy-using-GenAI Table 2: Hardware and software configuration of the experimental server. Component Specification Operating System Debian GNU/Linux 12 (bookworm) Kernel Version 6.1.0-28-amd64 CPU Model 2 × Intel Xeon Gold 5317, 3.0 GHz CPU Cores / Threads 48 cores / 96 threads System Memory 125 GiB RAM Primary Storage 223 GB SSD (system) Secondary Storage 5.2 TB HDD (data) Table 3: Number of parameters in the models. Model Trainable Non-trainable Total parameters parameters parameters Generator 54,419,459 9,856 54,429,315 / U-Net Discriminator 6,965,441 2,816 6,968,257 / PatchGAN Full cGAN 54,422,275 6,975,297 61,397,572 3.3 Algorithms The complete training and inference procedures of the proposed generative AI framework are summarized using two stage-wise algorithms. Algorithm 1 describes the standalone training procedure for the Pix2Pix Conditional GAN model that amplifies the imaging power, while Algorithm 2 outlines the inference pipeline used to generate the space-quality enhanced outputs from ground-telescope images. Algorithm 1 Training procedure for the Pix2Pix conditional GAN model. 1: 2: Ground–space paired dataset (A,B)(A,B) where: A=A= ground–telescope images (input domain) and B=B= space–telescope images (target domain) 3: Discriminator D, Generator G, Adversarial composite model GANGAN 4: Epochs E, Batch size b 5: Compute number of batches N←⌊|A|/b⌋N← |A|/b 6: Compute total update steps T←E⋅NT← E· N 7: for t=1t=1 to T do 8: Select real sample pairs (xA,xB)(x_A,x_B) from dataset 9: Compute real labels yrealy_real 10: Generate fake target samples: x^B←G⁡(xA) x_B← G(x_A) 11: Compute fake labels yfakey_fake 12: Update discriminator on real: ℒD1←D.train_on_batch([xA,xB],yreal)L_D1← D.train\_on\_batch([x_A,x_B],y_real) 13: Update discriminator on fake: ℒD2←D.train_on_batch([xA,x^B],yfake)L_D2← D.train\_on\_batch([x_A, x_B],y_fake) 14: Update generator using composite objective: ℒG←GAN.train_on_batch(xA,[yreal,xB])L_G← GAN.train\_on\_batch(x_A,[y_real,x_B]) 15: if tmod(10⋅N)=0t (10· N)=0 then 16: Evaluate model on held-out samples and store results 17: Save the trained generator model G∗G 18: end if 19: end for Algorithm 2 Inference Procedure for Enhancing Ground Telescope Images 1: 2: Trained generator G∗G 3: input image xAx_A from domain A (ground–telescope image) 4: Preprocess xAx_A (resize, grayscale conversion, normalization) 5: Generate enhanced output: x^B←G∗(xA) x_B← G (x_A) ⊳ x^B x_B approximates the target domain B: space–telescope image 6: Postprocess x^B x_B to valid pixel range 7: return enhanced image x^B x_B (space–quality reconstruction) 4 Results Evaluation of the effectiveness of GANs is known to be a challenging task 35; 6. GANs are often used to generate complex data, making it difficult to assess the quality of the output quantitatively. In some cases, cognitive tests are used to determine the ability of the GAN to generate images that seem to a person similar to natural images, or other target images that the GAN is expected to generate (2; 3; 39; 25). The analysis of the GAN is performed here using the loss function, as well as through manual inspection by comparing the generated images to the same images captured by the Hubble Space Telescope. Another method of evaluation is applying basic galaxy image analysis on the galaxy images acquired by the space telescope and the AI-generated galaxy images, and comparing the differences between the results. Figure 5 shows examples of DESI Legacy Survey images, and images generated by the GAN when these images are used as input. The generated images can be compared to the images of the same objects taken using the Hubble Space Telescope. As the figure shows, the GAN is capable of reconstructing the morphological features of the galaxy, making the enhanced galaxy similar in detail to the image taken by HST. Figure 5: Examples of ground-based telescope images, the space-based images of the same objects, and the images generated by the GAN from the ground-based images. The residual maps are created from the differences between generated images and the space telescope images. The objects are at coordinates (α=150.6532946o,δ=1.6253657o)(α=150.6532946^o,δ=1.6253657^o), (α=150.3196276o,δ=1.7540747o)(α=150.3196276^o,δ=1.7540747^o), (α=150.3742923o,δ=1.6191228o)(α=150.3742923^o,δ=1.6191228^o), (α=150.0445515o,δ=1.6114927o)(α=150.0445515^o,δ=1.6114927^o). While the figure shows just a few examples, these examples are representative of the entire catalog. The entire catalog is available online and described in Section 4.1. Figure 6: Examples from the catalog where the model demonstrates outstanding performance. Even when the input image lacks visible details of the galaxy, the model generates an enhanced version that closely matches the ground truth. The residual maps also show the good agreement between the generated image and space telescope image. The objects are at coordinates (α=149.68237o,δ=2.4202963o)(α=149.68237^o,δ=2.4202963^o), (α=149.97054o,δ=2.6837897o)(α=149.97054^o,δ=2.6837897^o), (α=149.8200861o,δ=2.4452223o)(α=149.8200861^o,δ=2.4452223^o), (α=149.8254184o,δ=2.1643785o)(α=149.8254184^o,δ=2.1643785^o). Figure 7: Catalog examples highlighting cases where the model underperforms. These issues might be addressed by training the model on a larger and more diverse set of images. The objects are at coordinates (α=150.7538768o,δ=1.957399o)(α=150.7538768^o,δ=1.957399^o), (α=150.7714556o,δ=2.0616442o)(α=150.7714556^o,δ=2.0616442^o), (α=150.7619236o,δ=2.6261117o)(α=150.7619236^o,δ=2.6261117^o), (α=150.7552252o,δ=2.2725501o)(α=150.7552252^o,δ=2.2725501^o). As Figures 6 and 7 show, the enhanced images also contain some background “pepper noise” added to the images, which is typical to AI-generated images. That can be removed from the images by using a final step of pre-processing. For instance, a simple low-pass filter can smooth that noise. That, however, can also change some of the relevant visual content. Therefore, removing the “pepper noise” should be done under the awareness that visual content can be affected by applying such filters. Another test was done by applying automatic image analysis to the galaxy images. The analysis was first applied to the space-telescopes images, and then the results of each image were compared to the results when applying the same analysis to the AI-enhanced galaxy image. If the AI-enhanced galaxy images are the same as the space telescope images, the results are expected to be identical. Therefore, any difference between the analysis of the space telescope image and the corresponding AI-generated image shows that the enhanced images are different. For that purpose we used the mature and commonly used SExtractor (5). SExtractor is a software tool for basic analysis of astronomical images. Since it is designed for both point sources and extended sources, it can also be applied to galaxy images. Here we used 1,000 images of galaxies taken by HST, paired with AI enhanced galaxy images of the same galaxies. These pairs of images are taken from the catalog described in Section 4.1. The analysis included several descriptors computed by SExtractor (5), which are the major axis, minor axis, position angle, elongation, and ellipticity. Additionally, we used the Ganalyzer tool (33) to determine whether the galaxy is elliptical or spiral, and compare the results computed on the space telescope images to the results computed on the AI-enhanced images. The analysis was done such that elliptical shape was assigned 1, and spiral shape was assigned the value 0. Table 4 shows the average relative difference Δμ μ between the descriptors computed from the space telescope images and the descriptors computed from the corresponding AI-enhanced images of the same galaxies. As the table shows, the descriptors as measured on the AI-enhanced images are not identical to those determined by the space telescope images, but normally stay within the 5% difference. The broad morphology reflects 45 galaxies out of the 1,000 that were tested in which the morphology of the AI-enhanced image did not match the morphology determined by using the space telescope image. Table 4: The differences between galaxy descriptors computed from the space telescope images and the same attributed computed from the AI-enhanced images of the same galaxies. Descriptors Relative difference¯ Relative difference Major axis (A) 0.056 Minor axis (B) 0.052 Position angle (θ) 0.031 Elongation (A/B) 0.05 Ellipticity 1-(B/A) 0.05 Broad morphology 0.06 4.1 A catalog of enhanced galaxy images from DESI Legacy Survey To test the ability of the method to provide catalogs of enhanced galaxy images, the method was applied to 63,202 galaxies from the DESI Legacy Survey. The galaxies were imaged by the Dark Energy Camera (DECam) of the Blanco telescope in Cerro Tololo, Chile. The catalog is available for download at https://doi.org/10.6084/m9.figshare.30226591. The catalog is organized into two directories: “enhanced_galaxies” and “galaxy_comparisons”. The “enhanced_galaxies” folder contains the output images produced by our generative AI model, each with a resolution of 256×256 pixels, as illustrated in Figure 8. Figure 8: Sample enhanced images from the catalog of enhanced galaxy images. The “galaxy_comparisons” folder includes image collages that provide a visual qualitative assessment of the model’s performance. Each collage consists of three images: the original input from a ground-based telescope (DESI), the AI-enhanced output, and the corresponding target image captured by a space-based telescope (ground-truth). These side-by-side comparisons help evaluate how closely the generated images resemble the true observations, as depicted in Figure 9. Figure 9: Example images from the “galaxy_comparisons” of the catalog. To reduce the overall size of the catalog, all images in the catalog are saved in PNG format. To be able to provide the “ground truth” for each enhanced galaxy image, the catalog is based on galaxies from the footprint covered by the COSMOS field. Each galaxy in the catalog is identified by the equatorial celestial coordinates, which are embedded in the filenames. These filenames consist of two components separated by an underscore: the first represents the Right Ascension (RA), and the second denotes the Declination (Dec) of the galaxy. For instance, an enhanced galaxy image in the catalog could be named “150.3295405_1.6032845.png”. In this case, the coordinates are (α=150.3295405o,δ=1.6032845o)(α=150.3295405^o,δ=1.6032845^o). 4.2 Galaxy Enhancer Software To simplify the user experience, we have bundled multiple components into a unified tool called “Galaxy Enhancer”. As depicted in Figure 10, this tool is composed of several integrated modules: the “User Input Module”, the “Ground-Based Imagery Download Module”, and the “Enhancer Module”. Figure 10: Modules and data flow involved in the galaxy enhancer tool. The user input module prompts users to provide the celestial coordinates of their target galaxy, specifically, the Right Ascension (RA) and Declination (Dec). The corresponding galaxy image is then retrieved in FITS format. After downloading, the FITS file is converted to a 16-bit TIF image. This TIF image is subsequently processed by the Enhancer Module, which houses our pre-trained custom generative AI model, as described in Section 3. The module processes the input, generates the enhanced output, saves the result, and returns the path to the enhanced image. The output file is typically named using the input coordinates, in the format: “RA_Dec.png”. Figure 11 demonstrates how the Galaxy Enhancer tool operates within a terminal interface. The full code base for downloading and using this tool is publicly accessible at https://github.com/SaiTeja-Erukude/Enhancing-Ground-Based-Astronomy-using-GenAI/tree/main/galaxy_enhancer. Figure 11: Terminal output illustrating each stage of the galaxy enhancer execution. 4.3 Experiments with SDSS and JWST galaxies Another experiment was done by enhancing galaxies imaged by SDSS, and comparing the results to images taken by JWST deep field. Figure 12 shows five galaxies imaged by SDSS, the enhanced galaxies, and their comparison to JWST imaged. The JWST field that was used for this experiment is the Stephan’s Quintet, at (RA=22h 35m 57.49s, Dec=33o 57’ 36”). The field is highly detailed, with over 150 million pixels, made from almost 1,000 separate images. Is is also within the footprint of SDSS, allowing to compare the two instruments. Figure 12: Examples of enhanced SDSS galaxies in comparison to JWST deep field. As the figure shows, the AI-enhanced SDSS images do not fit the JWST images at the same quality as the DESI Legacy Survey images enhanced to the quality of COSMOS images. That can be explained by the relatively low details of the SDSS images, compared to the very high level of details of JWST. While the AI can enhance images well, it is limited when the image quality gap between the source and target increases. The JWST Stephan’s Quintet image is obviously of far higher quality compared to HST COSMOS images, while SDSS images or of lower quality compared to DESI Legacy Survey. That high difference in image quality makes it far more challenging for the AI to enhance the images, showing the limitation of the method. 5 Conclusion Autonomous digital sky surveys are among the most powerful and most productive research instruments of our time. Earth-based digital sky surveys have a high bandwidth of data collection, but the quality of the imaging is still not comparable to space-based telescopes. Here, we used generative AI to enhance the quality of the images taken by Earth-based digital sky surveys. The enhancement is done by using a pix2pix GAN as a comprehensive filter. It is trained by pairs of images of the same galaxies taken by both Earth-based and space-based telescopes. The GAN is then trained to bridge between the images and can then transform images taken by Earth-based telescopes into the quality typical of space-based telescopes. In that sense, it can transform digital sky surveys into much more powerful instruments, combining their ability to cover large parts of the sky with the imaging quality of space-based telescopes. The method was applied to a catalog of a large number of galaxies, demonstrating that it can be used to generate catalogs of enhanced images. It is also provided in the form of a software tool that can transform input galaxy images, and therefore can be used for on-the-fly transformation of the images. The ability to transform the images in real time can be used by digital sky surveys to provide users with the ability to enhance objects of their choice as they browse through the user interface. While the method can provide additional power to the digital sky survey, it has several limitations. Firstly, it is trained and tested on extended objects only, as point sources are not part of this study. The enhancement is based on repetitive patterns of galaxy shapes, and in some cases of rare objects, the enhancement might lead to an image that is different from what that rare object is. In any case, machine learning systems are normally expected to have a certain degree of errors, and therefore some galaxies might be transformed in a manner that is not consistent with the true visual appearance of the object. But despite the limitations, the method can be used to enhance the quality of Earth-based images without the need to make the substantial resource investment typical of space-based telescopes. It is also fast and can therefore be used for on-the-fly enhancement of images without necessarily generating dedicated catalogs. Due to its availability and low footprint, such a method can be added to existing and future digital sky surveys to maximize their discovery power. Acknowledgments We would like to thank the knowledgeable anonymous reviewer for the helpful comments. Funding The research was supported in part by NSF grant number OIA-2148878. Conflicts of Interest The authors declare that there is no conflict of interest regarding the publication of this article. Data Availability The catalog of enhanced galaxy images, as well as training data for the cGAN, can be downloaded at https://doi.org/10.6084/m9.figshare.30226591. Code used in this project is available at https://github.com/SaiTeja-Erukude/Enhancing-Ground-Based-Astronomy-using-GenAI. References Aihara et al. (2018) H. Aihara, N. Arimoto, R. Armstrong, S. Arnouts, N. A. Bahcall, S. Bickerton, J. Bosch, K. Bundy, P. L. Capak, J. H. Chan, et al. The hyper suprime-cam ssp survey: overview and survey design. Publications of the Astronomical Society of Japan 70 (SP1), p. S4. Cited by: §1. Arora and Soni (2021) T. Arora and R. Soni A review of techniques to detect the gan-generated fake images. Generative Adversarial Networks for Image-to-Image Translation, p. 125–159. Cited by: §4. Ben Aissa et al. (2024) F. Ben Aissa, M. Hamdi, M. Zaied, and M. Mejdoub An overview of gan-deepfakes detection: proposal, improvement, and evaluation. Multimedia Tools and Applications 83 (11), p. 32343–32365. Cited by: §4. Berriman et al. (2004) G. Berriman, J. Good, A. Laity, A. Bergou, J. Jacob, D. Katz, E. Deelman, C. Kesselman, G. Singh, M. Su, et al. Montage: a grid enabled image mosaic service for the national virtual observatory. In Astronomical Data Analysis Software and Systems (ADASS) XIII, Vol. 314, p. 593. Cited by: §2. Bertin and Arnouts (1996) E. Bertin and S. Arnouts SExtractor: software for source extraction. Astronomy and Astrophysics Supplement Series 117 (2), p. 393–404. Cited by: §2, §4, §4. Borji (2022) A. Borji Pros and cons of gan evaluation measures: new developments. Computer Vision and Image Understanding 215, p. 103329. Cited by: §4. Campagne (2025) J. Campagne Galaxy imaging with generative models: insights from a two-models framework. Monthly Notices of the Royal Astronomical Society 539 (4), p. 3445–3458. Cited by: §1. Cecotti (2020) H. Cecotti Rotation invariant descriptors for galaxy morphological classification. International Journal of Machine Learning and Cybernetics 11 (8), p. 1839–1853. Cited by: §1. Collaboration: et al. (2016) D. E. S. Collaboration:, T. Abbott, F. Abdalla, J. Aleksić, S. Allam, A. Amara, D. Bacon, E. Balbinot, M. Banerji, K. Bechtol, et al. The dark energy survey: more than dark energy–an overview. Monthly Notices of the Royal Astronomical Society 460 (2), p. 1270–1299. Cited by: §1. Demir and Unal (2018) U. Demir and G. Unal Patch-based image inpainting with generative adversarial networks. arXiv:1803.07422. Cited by: §3.1, Table 1. Dey et al. (2019) A. Dey, D. J. Schlegel, D. Lang, R. Blum, K. Burleigh, X. Fan, J. R. Findlay, D. Finkbeiner, D. Herrera, S. Juneau, et al. Overview of the desi legacy imaging surveys. Astronomical Journal 157 (5), p. 168. Cited by: §2. Dieleman et al. (2015) S. Dieleman, K. W. Willett, and J. Dambre Rotation-invariant convolutional neural networks for galaxy morphology prediction. Monthly notices of the royal astronomical society 450 (2), p. 1441–1459. Cited by: §1. Djorgovski et al. (2001) S. Djorgovski, R. Brunner, A. Mahabal, S. Odewahn, R. d. Carvalho, R. Gal, P. Stolorz, R. Granat, D. Curkendall, J. Jacob, et al. Exploration of large digital sky surveys. In Mining the Sky: Proceedings of the MPA/ESO/MPE Workshop Held at Garching, Germany, July 31-August 4, 2000, p. 305–322. Cited by: §1. Elfattah et al. (2012) M. A. Elfattah, M. A. A. ELsoud, A. E. Hassanien, and T. Kim Automated classification of galaxies using invariant moments. In International Conference on Future Generation Information Technology, p. 103–111. Cited by: §1. Erukude and Shamir (2025) S. T. Erukude and L. Shamir Galaxy image simplification using generative ai. Astronomy and Computing, p. 100990. Cited by: §1, §3.1. Fussell and Moews (2019) L. Fussell and B. Moews Forging new worlds: high-resolution synthetic galaxies with chained generative adversarial networks. Monthly Notices of the Royal Astronomical Society 485 (3), p. 3203–3214. Cited by: §1. Goodfellow et al. (2014) I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio Generative adversarial nets. Advances in Neural Information Processing Systems 27. Cited by: §3. Hettiarachchi et al. (2024) A. Hettiarachchi, S. Rathnayake, and K. Dissanayaka A generative adversarial network to upscale the resolution of low-resolution galaxy images. In 2024 6th International Conference on Advancements in Computing, p. 55–60. Cited by: §1. Isola et al. (2018) P. Isola, J. Zhu, T. Zhou, and A. A. Efros Image-to-image translation with conditional adversarial networks. External Links: 1611.07004, Link Cited by: Table 1, §3. Ivezić et al. (2012) Ž. Ivezić, T. C. Beers, and M. Jurić Galactic stellar populations in the era of the sloan digital sky survey and other large surveys. Annual Review of Astronomy and Astrophysics 50 (1), p. 251–304. Cited by: §1. Ivezić et al. (2019) Ž. Ivezić, S. M. Kahn, J. A. Tyson, B. Abel, E. Acosta, R. Allsman, D. Alonso, Y. AlSayyad, S. F. Anderson, J. Andrew, et al. LSST: from science drivers to reference design and anticipated data products. Astrophysical Journal 873 (2), p. 111. Cited by: §1. Kaiser et al. (2002) N. Kaiser, H. Aussel, B. E. Burke, H. Boesgaard, K. Chambers, M. R. Chun, J. N. Heasley, K. Hodapp, B. Hunt, R. Jedicke, et al. Pan-starrs: a large synoptic survey telescope array. Proceedings of the SPIE 4836, p. 154–164. Cited by: §1. Kingma and Ba (2014) D. P. Kingma and J. Ba Adam: a method for stochastic optimization. arXiv:1412.6980. Cited by: §3.2. Kron (1995) R. G. Kron Digital optical sky surveys. Publications of the Astronomical Society of the Pacific 107 (714), p. 766. Cited by: §1. Lang et al. (2021) O. Lang, Y. Gandelsman, M. Yarom, Y. Wald, G. Elidan, A. Hassidim, W. T. Freeman, P. Isola, A. Globerson, M. Irani, et al. Explaining in style: training a gan to explain a classifier in stylespace. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 693–702. Cited by: §4. Lanusse et al. (2021) F. Lanusse, R. Mandelbaum, S. Ravanbakhsh, C. Li, P. Freeman, and B. Póczos Deep generative models for galaxy image simulations. Monthly Notices of the Royal Astronomical Society 504 (4), p. 5543–5555. Cited by: §1. Margony (1999) B. Margony The sloan digital sky survey. Philosophical Transactions of the Royal Society of London. Series A: Mathematical, Physical and Engineering Sciences 357 (1750), p. 93–103. Cited by: §1. Mellier et al. (2024) Y. Mellier, A. Abdurroúf, J. A. Barroso, A. Achúcarro, J. Adamek, R. Adam, G. Addison, N. Aghanim, M. Aguena, V. Ajani, et al. Euclid. i. overview of the euclid mission. Astronomy & Astrophysics. Cited by: §1. Ronneberger et al. (2015) O. Ronneberger, P. Fischer, and T. Brox U-net: convolutional networks for biomedical image segmentation. In 18th International Conference on Medical image computing and Computer-assisted Intervention, Vol. 3, p. 234–241. Cited by: §3.1, Table 1. Scoville et al. (2007) N. Scoville, R. Abraham, H. Aussel, J. Barnes, A. Benson, A. Blain, D. Calzetti, A. Comastri, P. Capak, C. Carilli, et al. COSMOS: hubble space telescope observations. Astrophysical Journal Supplement Series 172 (1), p. 38. Cited by: §2. Semenov et al. (2025) V. Semenov, V. Tymchyshyn, V. Bezguba, M. Tsizh, and A. Khlevniuk Galaxy morphological classification with manifold learning. Astronomy and Computing 52, p. 100963. Cited by: §1. Shamir (2009) L. Shamir Automatic morphological classification of galaxy images. Monthly Notices of the Royal Astronomical Society 399 (3), p. 1367–1372. Cited by: §1. Shamir (2011) L. Shamir Ganalyzer: a tool for automatic galaxy image analysis. Astrophysical Journal 736 (2), p. 141. Cited by: §4. Shamir (2021) L. Shamir Automatic identification of outliers in hubble space telescope galaxy images. Monthly Notices of the Royal Astronomical Society 501 (4), p. 5229–5238. Cited by: §2. Shmelkov et al. (2018) K. Shmelkov, C. Schmid, and K. Alahari How good is my gan?. In Proceedings of the European conference on Computer Vision, p. 213–229. Cited by: §4. Spergel et al. (2015) D. Spergel, N. Gehrels, C. Baltay, D. Bennett, J. Breckinridge, M. Donahue, A. Dressler, B. Gaudi, T. Greene, O. Guyon, et al. Wide-field infrarred survey telescope-astrophysics focused telescope assets wfirst-afta 2015 report. arXiv:1503.03757. Cited by: §1. Tyson and Borne (2012) J. A. Tyson and K. D. Borne Future sky surveys: new discovery frontiers. Advances in Machine Learning and Data Mining for Astronomy, p. 161–181. Cited by: §1. York et al. (2000) D. G. York, J. Adelman, J. E. Anderson Jr, S. F. Anderson, J. Annis, N. A. Bahcall, J. Bakken, R. Barkhouser, S. Bastian, E. Berman, et al. The sloan digital sky survey: technical summary. Astronomical Journal 120 (3), p. 1579. Cited by: §1. Zhang et al. (2017) L. Zhang, Y. Ji, X. Lin, and C. Liu Style transfer for anime sketches with enhanced residual u-net and auxiliary classifier gan. In Asian Conference on Pattern Recognition, p. 506–511. Cited by: §4.

Instant research discovery