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

Yuanzhe Jiang 1,∗, Yue Shen 1,2,∗, Grant Merz 1, Shurui Lin 1, Xin Liu 1,2,3,∗, Zhiwei Pan 1, Mingyang Zhuang 1, William Roster 4, Mara Salvato 4, Malgorzata Siudek 5, Grant Stevens 6 · 2026

Read DeepDISC-Euclid: Source Classification and Photometric Redshifts in Euclid Deep Field North With a Pixel-Level Deep Learning Approach by Yuanzhe Jiang 1,∗, Yue Shen 1,2,∗, Grant Merz 1, Shurui Lin 1, Xin Liu 1,2,3,∗, Zhiwei Pan 1, Mingyang Zhuang 1, William Roster 4, Mara Salvato 4, Malgorzata Siudek 5, Grant Stevens 6 (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-03182.txt.

Introduction

Jiang and colleagues apply pixel-level deep learning to classify sources and estimate photometric redshifts in Euclid's Deep Field North. Their DeepDISC system trains three models for detection, galaxy redshifts, and quasar redshifts, reaching about 93 percent detection completeness and 80 percent purity. The team releases a catalog of about 13 million objects with classifications and redshift estimates for the field.

Catalog record

CREATORYuanzhe Jiang 1,∗, Yue Shen 1,2,∗, Grant Merz 1, Shurui Lin 1, Xin Liu 1,2,3,∗, Zhiwei Pan 1, Mingyang Zhuang 1, William Roster 4, Mara Salvato 4, Malgorzata Siudek 5, Grant Stevens 6
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE79006
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.03182

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

DeepDISC-Euclid: Source Classification and Photometric Redshifts in Euclid Deep Field North With a Pixel-Level Deep Learning Approach ∗ Corresponding Authors: Yuanzhe Jiang ( [email protected] ), Yue Shen ( [email protected] ), Xin Liu ( [email protected] ). Yuanzhe Jiang 1,∗, Yue Shen 1,2,∗, Grant Merz 1, Shurui Lin 1, Xin Liu 1,2,3,∗, Zhiwei Pan 1, Mingyang Zhuang 1, William Roster 4, Mara Salvato 4, Malgorzata Siudek 5, Grant Stevens 6 Abstract The first Euclid Quick Data Release (Q1) provides extensive imaging and spectroscopic data for hundreds of millions of photometric objects across several deep fields. Accurate classifications and photometric redshifts (photo- z z ) for these sources are crucial to maximizing the value of these data. In this work, we perform source classification and photo- z z estimation for the Euclid Deep Field North (EDF-N) around the North Ecliptic Pole, using a deep learning framework (DeepDISC) that learns and infers using 9-band images simultaneously. We train three dedicated models for (1) source detection and classification (Model 1), (2) galaxy photo- z z (Model 2), and (3) quasar photo- z z (Model 3). The Euclid Q1 input source catalog, and classifications and spectroscopic redshifts (spec- z z ) from the Dark Energy Spectroscopic Instrument Data Release 1 are adopted as our training data. DeepDISC source detection achieves overall completeness of ∼ 93 % {\sim}93\% and purity of ∼ 80 % {\sim}80\% if using the Euclid source catalog as the ground truth. Using a JWST source catalog within EDF-N as the reference, we estimate a true purity of ∼ 90 % \sim 90\% for DeepDISC sources. About 99.2 % 99.2\% , 99.0 % 99.0\% , and 84.8 % 84.8\% of stars, galaxies, and quasars, respectively, are correctly recovered with their spectroscopic classifications. The DeepDISC photo- z z s show good agreement with spectroscopic redshifts, for both galaxies and quasars. Comparisons with other Euclid Q1 products demonstrate that DeepDISC provides comparable or improved performance in source detection/deblending, classification and photo- z z , especially for quasars. These results demonstrate the potential of pixel-level deep learning approaches for large-scale sky surveys such as Euclid and Roman, which will continue to improve with better training labels. We release the full DeepDISC source catalog ( ∼ 13 \sim 13 million objects) for EDF-N with classifications and photo- z z s, including photo- z z probability distributions. \defcitealias Stevens25GS25 \defcitealias Tucci25MT25 \defcitealias Siudek2025MS25 \defcitealias Roster2025WR25 keywords techniques: image processing – methods: data analysis – catalogs – surveys 1 Introduction Space-based wide-field imaging and spectroscopic surveys such as Euclid ( 15 ) are starting to revolutionize the field of astronomy. The first quick data release (Q1) from Euclid ( 12 ) provided an extensive set of multi-band (ground-based optical and Euclid) images, grism spectra and catalogs for the Euclid Deep Fields. These early data products demonstrate the capabilities of Euclid and the significant improvements over ground-based wide-area surveys. To facilitate the primary science goals of Euclid, the Euclid mission has developed the PHZ processing function to perform source classification, photometric redshift estimation ( 43 ; 38 ) , and measurements of galaxy physical properties ( 20 ; TM25 hereafter). The PHZ pipeline provides incredibly rich and valuable data products, which form the foundation for a myriad of science applications. However, there are several places where this process can be improved to enhance the scientific value of these data. The PHZ function classifies sources into stars, galaxies, and quasars (QSOs). It delivers a high-purity star sample and a highly complete galaxy sample. However, quasar identification remains challenging, as QSOs are difficult to distinguish from galaxies, and the classification pipeline can be further refined. Additionally, the PHZ photometric redshifts are primarily designed to support Euclid’s core science goals, which focus on galaxies. Consequently, PHZ classification and photometric redshifts are not optimal for QSOs. Pixel-level deep learning algorithm is a powerful method to conduct source detection, deblending, classification, and photo- z s ( 50 ; 48 ; 32 ) . Such approaches have been increasingly adopted in large astronomical surveys and have demonstrated strong performance ( 10 ; 3 ; 39 ; 22 ; 34 ; 33 ; 41 ; 17 ; 18 ; 19 , e.g.,) . For example, DeepDISC ( 34 ; 33 ) is a pixel-level deep learning method, which can provide source detection, deblending ( 31 , i.e., instance segmentation;) , classification, and photo- z estimation with full posterior probability density functions (PDFs) simultaneously. Using simulated Rubin Observatory Legacy Survey of Space and Time ( 29 , LSST;) data, 33 validates DeepDISC’s performance on source detection, deblending, classification, and photo- z estimation and demonstrates that DeepDISC photo- z can outperform traditional catalog-based (both template-fitting and ML-based) methods for both point estimates and full PDFs. Most recently, 35 extend DeepDISC to JWST Near Infrared Camera (NIRCam) imaging, demonstrating reliable probabilistic photo- z estimation to z{\sim}8 on JADES ( 11 ) data and showing that DeepDISC matches or outperforms EAZY ( 2 ) template fitting when the same photometric filters are available. In this work, we train DeepDISC models on real EDF-N images to produce an enhanced catalog with improved source classification and photometric redshifts. This paper is organized as follows. In Section 2 , we describe the data products used for training and the corresponding training strategies. In Section 3 , we evaluate the performance of the model predictions and present the final enhanced catalog for EDF-N. In Section 4 , we compare the DeepDISC classification and photometric redshifts with other Euclid results. Finally, we summarize this work in Section 5 . 2 Methodology 2.1 Data The Euclid Deep Field North (EDF-N) is a circular field covering approximately 20\,\mathrm{deg}^{2} around the ecliptic north pole, reaching a depth comparable to that of a single visit of the Euclid Wide Survey in the Q1 data release ( 12 ) . The EDF-N Q1 data release provides multi-wavelength images and photometry by combining Euclid the VISible instrument (VIS) and the Near-Infrared Spectrometer and Photometer (NISP) observations with external ground-based data from the Ultraviolet Near Infrared Optical Northern Survey ( 5 , UNIONS;) . Specifically, the Euclid observations include I_{\mathrm{E}} , Y_{\mathrm{E}} , J_{\mathrm{E}} , and H_{\mathrm{E}} bands, while UNIONS contributes Canada-France-Hawaii Telescope (CFHT) u , r , Hyper Suprime-Cam (HSC) g , z , and Pan-STARRS i imaging. The Euclid MERge Processing Function ( 16 , MER PF;) calibrates the multi-wavelength data in units of predefined regions of the sky, referred to as “tiles.” It produces a suite of data products, including the MERge catalog (MER) with calibrated photometry for all detected sources, background-subtracted mosaicked images in the nine bands listed above, and the corresponding segmentation maps. Within each tile, the multi-band images are uniformly rescaled to a common pixel scale by the MER PF. In this work, we utilize the 9-band background-subtracted mosaicked images in EDF-N from the Euclid Q1 data release. For GPU processing, the EDF-N multi-band images in each tile are divided into 800\times 800 pixel patches for training and applications. DeepDISC is a supervised deep-learning method and therefore requires ground-truth labels for source detection, classification, and photo- z estimation. For detection training, we adopt the segmentation maps provided by the Euclid MER PF as the ground truth. We apply

Read the complete text