🔭 astrophysics

From DES to KiDS: Domain adaptation for cross-survey detection of low-surface-brightness galaxies

This paper demonstrates that domain adaptation techniques, utilizing deep learning models trained on Dark Energy Survey data, can robustly identify and characterize over 20,000 low-surface-brightness galaxies and 434 ultra-diffuse galaxies in the KiDS DR5 survey, establishing a scalable pathway for creating homogeneous galaxy catalogues for future large-scale surveys like LSST and Euclid.

Original authors: Hareesh Thuruthipilly, Krzysztof Lisiecki, Junais, Katarzyna Małek, Agnieszka Pollo, William J. Pearson, Antonio Vanzanella, Saptarshi Pal, Miguel Figueira, Pratik Dabhade, Anna Durkalec, Aidan P. Co

Published 2026-05-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hareesh Thuruthipilly, Krzysztof Lisiecki, Junais, Katarzyna Małek, Agnieszka Pollo, William J. Pearson, Antonio Vanzanella, Saptarshi Pal, Miguel Figueira, Pratik Dabhade, Anna Durkalec, Aidan P. Cotter, Unnikrishnan Sureshkumar, Nandini Hazra, Patryk Matera, Subhrata Dey, Michal Vrábel, Anirban Dutta, Henry Willems, Nicola Principi Cavaterra, Natalia Dobrowolska, Wojciech Knop

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 the universe is a giant, dark room filled with thousands of different types of furniture. Most of the furniture is shiny, bright, and easy to spot—like a gleaming chandelier or a polished table. These are the "normal" galaxies we usually see in the sky. But hidden in the corners, barely visible against the dark walls, are faint, ghostly objects made of thin, wispy fabric. These are Low-Surface-Brightness Galaxies (LSBGs). They are so dim and spread out that they are incredibly hard to find, yet they might make up a huge chunk of all the furniture in the room.

This paper is about a team of astronomers who built a special "smart camera" to find these ghostly galaxies in a new map of the sky, even though the camera was originally trained on a different map.

Here is the story of how they did it, broken down into simple steps:

1. The Problem: Training a Dog to Find Cats in a Different Forest

The astronomers had a very smart computer program (a type of Artificial Intelligence called Deep Learning) that was excellent at finding these faint galaxies in one specific sky survey called DES (Dark Energy Survey). Think of this program as a dog trained to find a specific type of rabbit in a specific forest.

However, they wanted to use this "dog" to find rabbits in a different forest called KiDS (Kilo-Degree Survey). The problem? The two forests look different. The lighting is different, the trees are a different size, and the ground texture is different. Usually, if you take a dog trained in one forest and drop it in another, it gets confused and stops working.

2. The Solution: The "Translator" Trick

Instead of retraining the dog from scratch (which would take a long time and require a lot of new training data), the team used a clever trick called Domain Adaptation.

Imagine you are trying to show a picture of a rabbit to someone who speaks a different language. Instead of teaching them the whole language, you just translate the concept of the picture.

The team took the images from the new forest (KiDS) and adjusted them so they looked physically similar to the old forest (DES).
They converted the brightness of the stars into a universal "physical unit" (like changing from "looks bright to my eye" to "exact amount of light energy").
They resized the images so the "pixels" (the tiny dots making up the picture) matched the same scale.

By doing this, they made the new forest look familiar to the old dog. The dog didn't need new training; it just recognized the rabbits because they looked the same in this new, translated language.

3. The Hunt: Finding the Ghosts

With their "translated" images, they ran the AI models over the KiDS data.

The Result: The AI found 20,180 of these faint, ghostly galaxies!
The Super-Faint Ones: They also found 434 "Ultra-Diffuse Galaxies" (UDGs). These are like the most transparent, wispy ghosts of all—huge in size but incredibly light in weight.

4. The Cleanup: The Human Eye

Computers are great, but they aren't perfect. Sometimes they get tricked by weird patterns in the sky, like a smudge on a lens or a distant star that looks like a fuzzy blob.

The team had a group of 18 human volunteers act as "quality control inspectors."
They looked at thousands of images one by one to make sure the AI wasn't making mistakes.
After this human review, they confirmed the final list of 20,180 galaxies.

5. What Did They Learn?

Once they had their list of ghosts, they started asking questions about them:

Are they related to normal galaxies? Yes. The team found that these faint galaxies sit on a smooth sliding scale between small, normal dwarf galaxies and the huge, ultra-diffuse ones. It's like finding a smooth ramp connecting a small stool to a giant beanbag chair, rather than a gap between them. They are all part of the same family.
Are they blue or red? The galaxies came in two main colors, like a split personality:
- Blue (73%): These are "teenagers" still making new stars. They are active and forming new life.
- Red (27%): These are "retirees" that have stopped making stars. They are quiet and old.
Where do they live? The team noticed a strong pattern based on where the galaxies live.
- Galaxies living alone in the "countryside" (isolated space) are mostly blue and active.
- Galaxies living in crowded "cities" (galaxy clusters) are mostly red and quiet.
- The Analogy: It's like living in a busy city. The noise and pressure of the crowd (other galaxies) seem to "quench" or stop the star-making process, turning the blue, active galaxies into red, quiet ones. The environment is the reason they changed.

6. The Big Picture

The paper concludes that this "translator" method works incredibly well. It proves that we can use AI trained on one survey to find galaxies in another without starting over.

This is a huge deal because, in the near future, massive new telescopes (like the LSST and Euclid) will take pictures of the entire sky, finding millions of these faint galaxies. We won't have time to train a new AI for every single telescope. This paper shows us a roadmap: train once, translate the data, and find the ghosts everywhere.

In short, the team built a universal key that unlocks the door to the universe's most hidden, faint inhabitants, proving that even the dimmest lights in the cosmos can be found if you know how to look.

Technical Summary: From DES to KiDS: Domain adaptation for cross-survey detection of low-surface-brightness galaxies

Problem Statement
Low-surface-brightness galaxies (LSBGs) are critical for understanding galaxy formation and constitute a significant fraction of the galaxy population and dynamical mass of the Universe. However, their diffuse nature makes them difficult to detect, leading to severe underrepresentation in current catalogues. A major challenge in the field is the lack of a universally adopted observational definition and the difficulty of applying detection methods trained on one survey to another due to differences in photometric zero-points, pixel scales, point-spread functions (PSFs), and noise properties. As the astronomical community prepares for the Legacy Survey of Space and Time (LSST) and Euclid, which will discover millions of diffuse galaxies, there is an urgent need for robust, automated, and scalable methods to construct homogeneous LSBG catalogues across heterogeneous datasets without requiring labor-intensive retraining or manual re-labeling for every new survey.

Methodology
The authors investigate the efficacy of domain adaptation techniques to identify LSBGs in the Kilo-Degree Survey Data Release 5 (KiDS DR5) using models trained exclusively on the Dark Energy Survey (DES).

Training Data: The study utilizes a labelled dataset of 37,000 objects (18,532 LSBGs and 18,468 contaminants) from DES DR1, identified by Tanoglidis et al. (2021b) and updated by Thuruthipilly et al. (2024b).
Model Architecture: An ensemble of three deep learning models is employed: one Convolutional Neural Network (CNN) and two Transformer-based models (LSBG-DETR and LSBG-ViT). These models were originally developed for DES but are retrained using two-band (g and r) inputs to reduce computational costs while maintaining performance.
Domain Adaptation Strategy: To bridge the distributional differences between the source domain (DES) and the target domain (KiDS), the authors apply a physics-motivated unsupervised domain adaptation approach. This involves:
1. Converting pixel values to physical surface-brightness units ( $\mu$ Jy arcsec $^{-2}$ ) to normalize photometric zero-points.
2. Resizing images to a common resolution (64 $\times$ 64 pixels) to account for differences in pixel scales while conserving flux.
3. Applying these normalizations allows the DES-trained models to be applied directly to KiDS imaging without retraining.
Selection Pipeline:
1. Pre-selection: From the KiDS DR5 parent catalogue ( $\sim$ 1.38 $\times$ 10 $^8$ sources), candidates are filtered based on effective radius ( $2'' < r_{eff} < 20''$ ), axis ratio ( $q > 0.3$ ), and surface brightness ( $\mu_{eff,r} > 23$ mag arcsec $^{-2}$ ), alongside color cuts.
2. Deep Learning Classification: The ensemble classifies sources as LSBG candidates if at least one model assigns a probability $P \ge 0.5$ .
3. Structural Fitting: Candidates undergo simultaneous g, r, and i-band Sérsic fitting using galfitm to refine structural parameters and identify nucleated systems.
4. Visual Inspection: A team of 18 co-authors visually inspects the candidates to remove contaminants and misfits, ensuring high sample purity.
5. Redshift and Physical Properties: The final sample is cross-matched with spectroscopic surveys (2dFGRS, GAMA, SDSS, DESI) and H I surveys (ALFALFA, FAST) to obtain redshifts. For sources lacking spectroscopic redshifts, photometric redshifts (photo-z), stellar masses ( $M_*$ ), and star formation rates (SFR) are derived via Spectral Energy Distribution (SED) fitting using CIGALE.

Key Results

Catalogue Construction: The methodology successfully identifies 20,180 LSBGs and 434 Ultra-Diffuse Galaxies (UDGs) in KiDS DR5.
Model Performance: The ensemble achieves a recall of 97% on the DES test set with a contamination rate of $\sim$ 6%. Validation against known LSBGs in the overlapping KiDS footprint shows a re-identification rate of 95.1% for DES LSBGs and 92.2% for known UDGs, confirming the models' ability to generalize across surveys.
Structural Properties: The KiDS-LSBGs exhibit structural parameters (Sérsic index, axis ratio, effective radius) consistent with LSBGs from DES and the Hyper Suprime-Cam SSP Survey (HSC-SSP).
Size-Luminosity Relation: The study reveals a continuous size-luminosity relation connecting classical dwarf galaxies, LSBGs, and UDGs, suggesting they represent a diffuse extension of the dwarf galaxy population rather than a distinct class.
Photometric Redshifts: SED fitting yields photo-z estimates with a normalized median absolute deviation (NMAD) of $\sigma_{NMAD} = 0.031$ , comparable to general KiDS galaxy populations. However, photo-z values are systematically overestimated ( $\Delta z \simeq 0.024$ ), leading to systematic offsets in stellar mass and SFR estimates. Despite these biases, the specific star formation rate (sSFR) shifts are small ( $\sim$ 0.13–0.22 dex), preserving the structure of the star-forming main sequence.
Color Bimodality and Environment: The LSBG sample shows a bimodal color distribution ( $\sim$ 73% blue, $\sim$ 27% red). Strong environmental trends are observed: cluster LSBGs and UDGs are redder and exhibit reduced star formation compared to non-cluster systems. In the lowest redshift bin ( $z \le 0.03$ ), 72.7% of cluster LSBGs are quenched, compared to only 11.1% in non-cluster environments. UDGs are more strongly quenched than non-UDGs in clusters.

Significance and Claims
The paper demonstrates that domain adaptation, specifically through surface-brightness normalization and image resizing, enables deep learning models trained on one survey (DES) to effectively identify LSBGs in a different, independent survey (KiDS) without retraining. This establishes a scalable and computationally efficient pathway for constructing homogeneous LSBG catalogues across heterogeneous datasets.

The authors claim this framework is particularly well-suited for the upcoming era of LSST and Euclid, where the sheer volume of data will make survey-specific model training and manual inspection impractical. By proving that cross-survey identification is feasible, the study provides a foundation for systematic, large-scale studies of the structural, stellar population, and environmental properties of diffuse galaxies across cosmic time. The work also highlights that while photo-z derived properties have systematic biases, they remain useful for population-level statistical analyses, such as characterizing the star-forming main sequence of LSBGs.