🔭 astrophysics

DeepDISC-Euclid: Source Classification and Photometric Redshifts in Euclid Deep Field North With a Pixel-Level Deep Learning Approach

This paper presents DeepDISC, a pixel-level deep learning framework that achieves high accuracy in source detection, classification, and photometric redshift estimation for the Euclid Deep Field North, demonstrating superior or comparable performance to existing Euclid Quick Data Release products, particularly for quasars.

Original authors: Yuanzhe Jiang, Yue Shen, Grant Merz, Shurui Lin, Xin Liu, Zhiwei Pan, Mingyang Zhuang, William Roster, Mara Salvato, Malgorzata Siudek, Grant Stevens

Published 2026-04-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuanzhe Jiang, Yue Shen, Grant Merz, Shurui Lin, Xin Liu, Zhiwei Pan, Mingyang Zhuang, William Roster, Mara Salvato, Malgorzata Siudek, Grant Stevens

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 as a giant, incredibly crowded city at night. The Euclid Space Telescope is like a new, ultra-high-definition security camera system launched to take a picture of this city. Its job is to count every single light (star, galaxy, or quasar) and figure out what kind of building it is and how far away it is.

However, there's a problem: the city is so crowded that many lights are smashed together, some are very faint, and some look like other things. The standard software the telescope team uses (called "PHZ") is good, but it sometimes gets confused, especially when trying to tell the difference between a bright star and a distant, active black hole (a quasar).

Enter DeepDISC, the "super-smart AI detective" created by the authors of this paper.

The Problem: A Messy Photo

The Euclid telescope took a picture of a specific patch of sky called the North Ecliptic Pole. It captured light in 9 different colors (like taking a photo with red, blue, green, and infrared filters all at once).

The standard software tries to sort these lights, but it struggles with:

Crowded streets: When two galaxies are right next to each other, the software sees them as one big blob.
Imposters: It's hard to tell a distant, active black hole (quasar) apart from a normal galaxy.
Distance guessing: Figuring out how far away a light is just by its color (photometric redshift) is tricky and often inaccurate for the weird objects.

The Solution: The Deep Learning Detective

The authors built a Deep Learning system (a type of AI that learns by looking at examples, not by following strict rules). Think of it like teaching a child to recognize animals. Instead of telling the child, "If it has four legs and a tail, it's a dog," you show them thousands of pictures of dogs, cats, and birds until they learn the patterns themselves.

DeepDISC does this with space images. Here is how they trained their AI:

The Training School: They didn't just guess. They used a "textbook" called DESI (a ground-based telescope that has already measured the exact distance of millions of objects using lasers). They showed the AI: "Here is a picture of a star, here is a galaxy, and here is a quasar. And here is the exact distance we know for each."
Three Specialized Detectives: Instead of one AI trying to do everything, they trained three specific models:
- Model 1 (The Spotter): Looks at the 9-band image and says, "There is an object here! Is it a star, a galaxy, or a quasar?" It also learns to "un-stick" (deblend) objects that are stuck together.
- Model 2 (The Galaxy Ranger): Specifically trained to guess the distance of galaxies.
- Model 3 (The Quasar Hunter): Specifically trained to guess the distance of quasars.

The Results: A Better Map

When they tested DeepDISC on the Euclid data, the results were impressive:

Spotting Power: It found 93% of all the objects the standard software found, but it was even better at finding the faint, hidden ones. When they checked against a super-powerful telescope (JWST), they realized DeepDISC was actually 90% pure (very few false alarms).
The "Quasar" Breakthrough: This is the big win. Standard software often mistakes quasars for galaxies. DeepDISC correctly identified 85% of the quasars, whereas the standard software struggled. It's like the AI finally learned to spot the "wolf in sheep's clothing."
Distance Accuracy: For galaxies, the distance guesses were just as good as the standard method. But for quasars? DeepDISC was much more accurate. It reduced the "guessing error" by a huge amount.

The Analogy: The Party Guest List

Imagine you are at a massive party (the universe) and you need to make a guest list.

The Standard Method: You walk around with a clipboard. You see a group of people huddled together and write down "One Big Group." You see a person in a suit and write "Guest." You see a person in a tuxedo and write "Guest" (even though they might be the DJ).
DeepDISC: You have a super-intelligent assistant who looks at the same party.
- They see the huddled group and say, "Actually, that's three people standing close together."
- They see the person in the tuxedo and say, "That's not a guest; that's the DJ (Quasar)!"
- They look at the lighting and the shadows to guess exactly how far away each person is standing.

The Final Product

The authors released a new catalog containing 13 million objects from this patch of sky.

It includes the location of every object.
It tells you what it is (Star, Galaxy, or Quasar).
It gives a probability distribution for how far away it is (not just a single number, but a range of likely distances).

Why Does This Matter?

This isn't just about making a prettier picture.

Cosmic History: By finding more quasars (which are ancient beacons), we can study the early universe better.
Dark Energy: By getting better distance measurements for galaxies, we can understand how the universe is expanding and what "Dark Energy" is doing.
Future Proofing: This AI approach is a blueprint for future telescopes (like the Roman Space Telescope). As we get more data, we can feed it to the AI to make it even smarter.

In short: The authors taught a computer to look at the universe's "night lights" with 9 different colored filters, taught it to spot the tricky ones, and gave us a much clearer, more accurate map of the cosmos than we had before.

1. Problem Statement

The Euclid mission's first Quick Data Release (Q1) provides extensive multi-band imaging and spectroscopic data for hundreds of millions of objects. However, the standard Euclid processing pipeline (PHZ) has limitations:

Quasar Identification: The PHZ pipeline struggles to distinguish quasars (QSOs) from galaxies, leading to incomplete or impure quasar samples.
Photometric Redshifts (Photo-z): The PHZ photo-z estimates are optimized for galaxies and are suboptimal for quasars.
Deblending: Traditional catalog-based methods often fail to separate closely blended sources, particularly in crowded fields.

The paper addresses the need for a more robust, pixel-level deep learning approach to simultaneously perform source detection, deblending, classification (stars, galaxies, quasars), and photo-z estimation with full probability density functions (PDFs).

2. Methodology

The authors utilize DeepDISC, a supervised deep learning framework based on the Swin Transformer architecture and the Detectron2 framework. The approach involves a staged training strategy using three dedicated models to overcome the scarcity of spectroscopic labels for photo-z training compared to the abundance of segmentation data.

Data Sources

Input Images: 9-band mosaicked images (Euclid VIS, NISP bands $I_E, Y_E, J_E, H_E$ + ground-based UNIONS $u, g, r, i, z$ ) from the Euclid Deep Field North (EDF-N).
Ground Truth for Detection: Euclid MER PF segmentation maps (derived from $I_E$ or stacked NIR bands).
Ground Truth for Classification & Photo-z: Dark Energy Spectroscopic Instrument Data Release 1 (DESI DR1) spectroscopic classifications and redshifts.
- Training Set: ~40,000 unique DESI-matched sources across 9,051 image cutouts.
- Validation/Test Sets: Split to ensure unique sources and balanced distributions.

Model Architecture & Training Strategy

The pipeline consists of three models trained sequentially:

Model 1 (Detection & Classification):
- Task: Source detection, instance segmentation (deblending), and classification (Star/Galaxy/QSO).
- Architecture: Swin Transformer backbone + Region Proposal Network (RPN) + Cascade ROI Head (for detection) + MLP (for classification).
- Training: Jointly trained on Euclid segmentation maps and DESI classifications. Initialized from ImageNet/COCO pre-trained weights.
Model 2 (Galaxy Photo-z):
- Task: Estimate photometric redshifts for galaxies.
- Architecture: Inherits weights from Model 1 (backbone and classification frozen). Adds a Mixture Density Network (MDN) ROI head outputting a 5-component Gaussian mixture PDF.
- Training: Trained exclusively on DESI galaxy spectroscopic redshifts.
Model 3 (Quasar Photo-z):
- Task: Estimate photometric redshifts for quasars.
- Architecture: Similar to Model 2, with an MDN head trained exclusively on DESI quasar spectroscopic redshifts.

Key Innovation: The "staged" approach freezes the detection/deblending features learned in Stage 1. This prevents the sparse DESI photo-z labels (~42k sources) from degrading the robust detection features learned from the much larger Euclid segmentation map dataset.

3. Key Contributions

Integrated Pixel-Level Framework: Demonstrates a unified deep learning pipeline that performs detection, deblending, classification, and probabilistic photo-z estimation simultaneously.
Specialized Quasar Handling: Achieves superior quasar classification and redshift estimation compared to standard Euclid PHZ products, which are galaxy-centric.
Large-Scale Catalog Release: Releases a comprehensive catalog of ~13.3 million sources for EDF-N, including classifications, photo-z PDFs, and segmentation map indices.
Staged Training Strategy: Proposes a practical compromise for current data releases where spectroscopic labels are limited, protecting detection performance while enabling accurate photo-z estimation.

4. Results

The models were evaluated on a held-out test set and compared against Euclid PHZ (MT25), Diffusion Model (GS25), and X-ray CTP (WR25) catalogs.

Detection Performance

Completeness (Recall): ~93% overall (vs. Euclid MER catalog).
Purity (Precision): ~80% overall.
True Purity: Estimated at ~90% when validated against a deep JWST source catalog (NEXUS) in a sub-region.
Deblending: DeepDISC shows improved ability to separate blended sources, particularly those with different colors, compared to Euclid MER.

Classification Performance

Stars: 99.2% completeness, 98.5% purity.
Galaxies: 99.0% completeness, 98.9% purity.
Quasars: 84.8% completeness, 90.6% purity.
- Note: Performance drops at bright magnitudes due to confusion between low-z quasars and their host galaxies, and at faint magnitudes due to low SNR.
Comparison: Outperforms Euclid PHZ in quasar classification and matches or exceeds the Diffusion Model (GS25) while automatically distinguishing stars from quasars (unlike GS25).

Photometric Redshift Performance

Galaxies (Model 2):
- Median bias ( $e_z$ ): $2.0 \times 10^{-3}$ .
- Scatter ( $\sigma_{IQR}$ ): $0.0393$.
- Outlier fraction ( $\eta$ ): 5.77%.
- Comparison: Comparable scatter to Euclid PHZ but with significantly lower bias and fewer outliers. Matches AstroPT (MS25) performance without requiring pre-selected galaxy samples.
Quasars (Model 3):
- Median bias: $-5.7 \times 10^{-3}$ .
- Scatter: $0.110$.
- Outlier fraction: 3.58%.
- Comparison: Significantly outperforms Euclid PHZ (which has 5x larger scatter) and achieves accuracy comparable to the X-ray CTP catalog (WR25), but with a much larger sample size as it does not require pre-selection.

5. Significance and Future Outlook

Scientific Impact: The released catalog enables immediate science applications, including cleaner quasar samples for reionization studies and improved weak lensing samples.
Methodological Advancement: Validates that pixel-level deep learning can outperform traditional template-fitting and catalog-based ML methods for large-scale surveys like Euclid and the upcoming Roman Space Telescope.
Limitations & Future Work:
- Current photo-z coverage for galaxies is limited to $z \lesssim 1.8$ due to DESI training data limits.
- Future improvements will leverage larger spectroscopic datasets (Euclid DR1, Subaru-PFS, JWST NEXUS) to enable fully joint training (detection + photo-z simultaneously) and extend redshift coverage to higher $z$ .
- Semi-supervised learning and generative augmentation are proposed to address labeled-data bottlenecks.

In conclusion, DeepDISC-EUCLID demonstrates that deep learning approaches can provide robust, high-fidelity source catalogs that complement and enhance standard pipeline products, particularly for challenging populations like quasars.