pith. machine review for the scientific record. sign in

arxiv: 2604.03182 · v2 · submitted 2026-04-03 · 🌌 astro-ph.GA · astro-ph.IM

Recognition: 2 theorem links

· Lean Theorem

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

Yuanzhe Jiang , Yue Shen , Grant Merz , Shurui Lin , Xin Liu , Zhiwei Pan , Mingyang Zhuang , William Roster
show 3 more authors
Mara Salvato Malgorzata Siudek Grant Stevens
Authors on Pith no claims yet

Pith reviewed 2026-05-13 18:36 UTC · model grok-4.3

classification 🌌 astro-ph.GA astro-ph.IM
keywords Eucliddeep learningphotometric redshiftssource classificationquasarsgalaxiesdeep fieldspixel-level analysis
0
0 comments X

The pith

A pixel-level deep learning framework applied to Euclid Deep Field North nine-band images delivers source detection at 93 percent completeness with 90 percent true purity plus classifications and photometric redshifts for galaxies and quas

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper applies a deep learning framework called DeepDISC to nine-band imaging data from the Euclid Deep Field North. It trains models on existing catalog labels to detect and classify sources while estimating photometric redshifts separately for galaxies and quasars. The results show overall completeness near 93 percent and purity around 80 percent relative to the Euclid catalog, with true purity near 90 percent when checked against JWST data. Classifications recover most spectroscopic types correctly, and photometric redshifts align well with spectroscopic measurements. These outcomes match or exceed other Euclid Quick Data Release products, particularly for quasars, and support releasing a catalog of about 13 million objects.

Core claim

DeepDISC uses a pixel-level deep learning framework trained on Euclid Q1 source catalog and DESI DR1 spectroscopic redshifts to perform source detection and classification on 9-band images, followed by separate photo-z estimation for galaxies and quasars. It achieves ~93% completeness and ~80% purity in detection (with ~90% true purity vs JWST), correctly recovers 99.2% of stars, 99.0% of galaxies, and 84.8% of quasars according to spectroscopic classifications, and produces photo-zs in good agreement with spec-zs. This provides comparable or improved performance over other Euclid Q1 products in detection/deblending, classification, and photo-z, especially for quasars.

What carries the argument

DeepDISC, a deep learning model that ingests 9-band images simultaneously to detect sources, classify them as stars, galaxies or quasars, and estimate photometric redshifts with dedicated galaxy and quasar branches.

If this is right

  • Source detection reaches 93% completeness and 80% purity against the Euclid catalog.
  • True source purity is estimated at 90% using JWST as reference.
  • Classification recovers over 99% of stars and galaxies and 85% of quasars from spectroscopic labels.
  • Photometric redshifts agree well with spectroscopic redshifts for both galaxies and quasars.
  • The released catalog covers 13 million objects in EDF-N with classifications and photo-z probability distributions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If training labels improve in future data releases, the accuracy for faint quasars should rise further.
  • Similar pixel-level methods could be applied to other Euclid deep fields for uniform catalogs.
  • Combining this approach with data from the Roman Space Telescope might enhance photo-z precision through broader wavelength coverage.
  • Traditional catalog-based methods may underperform in deblending compared to this image-level learning in crowded regions.

Load-bearing premise

The training labels drawn from the Euclid Q1 source catalog and DESI DR1 spectroscopic redshifts are assumed to be sufficiently complete and unbiased for all magnitudes, colors, and object types in the field.

What would settle it

A deeper, more complete reference catalog from future JWST or spectroscopic observations revealing a purity significantly below 80% or classification recovery below 85% for quasars would falsify the performance claims.

Figures

Figures reproduced from arXiv: 2604.03182 by Grant Merz, Grant Stevens, Malgorzata Siudek, Mara Salvato, Mingyang Zhuang, Shurui Lin, William Roster, Xin Liu, Yuanzhe Jiang, Yue Shen, Zhiwei Pan.

Figure 1
Figure 1. Figure 1: — Magnitude (total flux) distributions in Euclid bands (IE, YE, JE, and HE) for spectroscopically confirmed stars, galaxies, and QSOs, matched between the Euclid MER catalog in EDF-N and DESI DR1. Total fluxes in each band are computed from the 2FWHM aperture flux in the Euclid MER catalog, following the Photometry Cookbook (Euclid Collaboration: Aussel et al. 2025), and converted to magnitudes for display… view at source ↗
Figure 2
Figure 2. Figure 2: — Redshift distributions of galaxies and QSOs matched between the Euclid MER catalog in EDF-N and DESI DR1. (Euclid Collaboration: Guglielmo et al. 2020; Gruen et al. 2023), a fully joint training regime will become feasible and is planned for future work. As illustrated in Figure 3b, the current staged strategy isolates the photo-z ROI heads from the backbone during Stage 2 training precisely to protect t… view at source ↗
Figure 3
Figure 3. Figure 3: — (a) The DeepDISC three-model pipeline applied to the Euclid Deep Field North. Nine-band Euclid+UNIONS images are processed by Model 1 for simultaneous detection, instance segmentation, and star/galaxy/QSO classification. Models 2 and 3 append galaxy and QSO photometric redshift PDFs respectively to produce the final merged catalog of ∼13 million sources. Dashed lines indicate training label inputs. (b) S… view at source ↗
Figure 4
Figure 4. Figure 4: — Detection recall (completeness) and precision (purity) of Model 1 as functions of IE on the test set. Blue circles denote recall, and red triangles denote precision. Poisson uncertainties are shown. case, the Euclid MER catalog): r = ED ED + E , (1) p = ED ED + D , (2) where we use three detection counts ED, E, and D, cor￾responding to sources detected by both Euclid and Deep￾DISC, those only detected by… view at source ↗
Figure 5
Figure 5. Figure 5: — Example comparisons between Euclid (left) and Model 1 (right) segmentation maps. The Euclid segmentation maps are shown in the left panels, while the Model 1–predicted segmentation maps are shown in the right panels. The top row displays cutouts in the IE band, and the bottom row shows cutouts in the YE band. Red solid contours indicate objects detected by both Euclid and Model 1 (ED); blue dashed contou… view at source ↗
Figure 6
Figure 6. Figure 6: — DeepDISC classification results on the test set. The left panel shows a confusion matrix comparing DeepDISC predictions with DESI spectroscopic classifications. Each cell reports both the number and fraction of DESI objects assigned to each category by DeepDISC. Color intensity indicates the fraction in each cell, and the dark diagonal signals demonstrate that most DeepDISC classifications are consistent… view at source ↗
Figure 7
Figure 7. Figure 7: — The 2D histograms of DeepDISC photo-z, zphot, versus DESI spec-z, zspec, for objects identified as galaxies or quasars by both DESI and the models in the test set. The left panel shows Model 2 predicted zphot, trained only with DESI galaxy zspec, which shows good agreement with DESI results. The right panel displays the comparison between DESI quasar zspec and zphot predicted by Model 3, trained only wit… view at source ↗
Figure 8
Figure 8. Figure 8: — Comparison of source classification performance among matched sources from DeepDISC Model 1, the Euclid PHZ catalog (MT25), and DESI DR1. The analysis is further restricted to sources with a single classification in the Euclid PHZ catalog, corresponding to classifications with dominant probabilities in Euclid. The upper and lower panels show the results for DeepDISC and the Euclid PHZ catalog, respective… view at source ↗
Figure 9
Figure 9. Figure 9: — Same as [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: — Comparison of galaxy photo-z estimates from DeepDISC with those from the Euclid PHZ catalog (upper MT25) and the AstroPT catalog (lower MS25). The left column shows the DeepDISC results, while the right column shows the Euclid PHZ (upper panel) and AstroPT (lower panel) results. Sources included in each comparison are required to have DESI spec-zs as well as photo-zs from DeepDISC and Euclid PHZ/AstroPT… view at source ↗
Figure 11
Figure 11. Figure 11: — Comparison of quasar photo-z estimates from DeepDISC (left) with those from the Euclid PHZ catalog (middle MT25) and the X-ray CTP catalog (right WR25). Sources included in each comparison are required to have DESI spectroscopic counterparts. The median bias relative to zspec, scatter σIQR, and outlier fraction η (defined in Section 3.3) for each sample are shown in each panel. The 1:1 relation and the … view at source ↗
Figure 12
Figure 12. Figure 12: — Redshift distributions of galaxies (left) and quasars (right) in the final DeepDISC catalog (Section 3.4) compared with those of the training set. Blue solid lines denote the DeepDISC distributions, while orange dashed lines denote the DESI training-set distributions [PITH_FULL_IMAGE:figures/full_fig_p015_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: — Three examples of mismatches between Euclid (top panels) and DeepDISC (bottom panels) detections for bright stars or extended galaxies. The left panels show a star for which Euclid and DeepDISC predict similar centers, yet the offset still exceeds the cross-matching threshold. In both cases, the positions are defined as the centers of the bounding boxes, which can be biased by diffraction spikes. The mi… view at source ↗
Figure 14
Figure 14. Figure 14: — Three examples of image artifacts affecting the detection. The top and bottom panels are for Euclid and DeepDISC detections, respectively. The left panels show suspicious Euclid detections along diffraction spikes of bright stars that are not detected by DeepDISC, which are usually very faint in IE band (IE > 24). The middle panels illustrate cases where Euclid detections include parts of the diffractio… view at source ↗
Figure 15
Figure 15. Figure 15: — Three examples of deblending by Euclid (top panels) and DeepDISC (bottom panels). The left panels illustrate DeepDISC separating sources embedded in a stellar diffraction spike. The middle panels present DeepDISC deblending of multiple closely spaced objects, while the right panels show the deblending of a complex system. These examples suggest that DeepDISC can more effectively separates close companio… view at source ↗
Figure 16
Figure 16. Figure 16: — Three examples illustrating how multi-band information improves DeepDISC detection and deblending. The top panels show a case where objects that are faint in the optical bands but bright in the near-infrared are missed by Euclid yet detected by DeepDISC. The middle and bottom panels present examples where DeepDISC successfully separates sources that are blended in Euclid MER detection by leveraging mult… view at source ↗
Figure 17
Figure 17. Figure 17: — Two examples of objects with multiple detections by DeepDISC at nearly identical positions, resulting in DeepDISC-only (D) detections. Both examples are faint in IE band, with a magnitude larger than 24. In top panels, two detections appear similar and faint in all bands and are likely produced by noise fluctuations. In bottom panels, the objects exhibit different morphologies and magnitudes in optical … view at source ↗
read the original 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) for these sources are crucial to maximizing the value of these data. In this work, we perform source classification and photo-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, (2) galaxy photo-z, and (3) quasar photo-z. The Euclid Q1 input source catalog, and classifications and spectroscopic redshifts (spec-z) from the Dark Energy Spectroscopic Instrument Data Release 1 are adopted as our training data. DeepDISC source detection achieves overall completeness of ~93% and purity of ~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% for DeepDISC sources. About 99.2%, 99.0%, and 84.8% of stars, galaxies, and quasars, respectively, are correctly recovered with their spectroscopic classifications. The DeepDISC photo-zs 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, 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 million objects) for EDF-N with classifications and photo-zs, including photo-z probability distributions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript presents DeepDISC, a pixel-level deep learning framework applied to 9-band Euclid Q1 imaging in the Euclid Deep Field North (EDF-N). Three models are trained on labels from the Euclid Q1 source catalog and DESI DR1 spectroscopic redshifts to perform source detection/classification (stars/galaxies/quasars), galaxy photo-z estimation, and quasar photo-z estimation. Reported metrics include ~93% completeness and ~80% purity relative to the Euclid catalog (with ~90% true purity vs. a JWST reference), recovery rates of 99.2% (stars), 99.0% (galaxies), and 84.8% (quasars), and good photo-z agreement with spec-z; the work claims comparable or improved performance over other Euclid Q1 products and releases a catalog of ~13 million objects with classifications and photo-z PDFs.

Significance. If the performance holds under more independent validation, the work demonstrates the practical utility of simultaneous multi-band pixel-level deep learning for source deblending, classification, and photo-z in large surveys such as Euclid. The public release of the full EDF-N catalog with probability distributions is a concrete strength that supports community reuse and further testing.

major comments (2)
  1. [§4] §4 (performance metrics): Completeness (~93%) and purity (~80%) are evaluated against the same Euclid Q1 source catalog used to supply training labels. This setup creates circularity that risks overestimating detection performance, as the model is optimized to reproduce the catalog's detections; the limited JWST overlap provides only a partial external anchor and does not fully mitigate the concern for the full field.
  2. [§4.3] §4.3 (quasar recovery): The 84.8% quasar recovery rate assumes DESI DR1 spec-z provide an unbiased sample across magnitude, color, and redshift; DESI fiber targeting and selection functions are known to miss subpopulations, and this potential bias is not quantified or tested against an alternative quasar reference.
minor comments (2)
  1. [§3] The description of the training/validation split and any steps taken to prevent data leakage between the Euclid Q1 labels and the evaluation should be expanded for clarity.
  2. [Figure 5] Figure captions and axis labels for the photo-z comparison plots would benefit from explicit mention of the outlier fraction definition used.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments and the recommendation for minor revision. We address each major comment point by point below.

read point-by-point responses

  1. Referee: §4 (performance metrics): Completeness (~93%) and purity (~80%) are evaluated against the same Euclid Q1 source catalog used to supply training labels. This setup creates circularity that risks overestimating detection performance, as the model is optimized to reproduce the catalog's detections; the limited JWST overlap provides only a partial external anchor and does not fully mitigate the concern for the full field.

    Authors: We agree that evaluating detection performance against the training catalog introduces circularity, as the model is trained to reproduce the Euclid Q1 detections. This is an inherent limitation of supervised approaches when no fully independent large-scale reference exists. The JWST overlap provides an external check yielding ~90% true purity, which we already report. We will revise §4 to explicitly discuss this circularity, clarify that the metrics measure fidelity to the input catalog, and emphasize the independent JWST validation as the primary external anchor. The reported numbers themselves will remain unchanged. revision: partial

  2. Referee: §4.3 (quasar recovery): The 84.8% quasar recovery rate assumes DESI DR1 spec-z provide an unbiased sample across magnitude, color, and redshift; DESI fiber targeting and selection functions are known to miss subpopulations, and this potential bias is not quantified or tested against an alternative quasar reference.

    Authors: We acknowledge that DESI DR1 quasar samples are affected by fiber-targeting selection biases. The 84.8% recovery rate is computed only for objects with existing DESI spectroscopic classifications and redshifts; it does not claim to represent the full underlying quasar population. No alternative large-area quasar reference catalog is available for the entire EDF-N that would allow a direct bias quantification. We will add a dedicated paragraph in §4.3 noting this limitation and stating that the recovery statistics are relative to the available spectroscopic subsample. revision: partial

Circularity Check

0 steps flagged

No significant circularity: performance metrics are standard supervised evaluations against external labels with independent JWST cross-check

full rationale

The paper trains DeepDISC models on labels from the Euclid Q1 source catalog and DESI DR1 spectroscopic redshifts, then reports completeness (~93%), purity (~80%), classification recovery rates, and photo-z agreement using those catalogs as ground truth while additionally validating true purity (~90%) against a JWST source catalog. This follows standard supervised learning practice of measuring model output against provided training/evaluation labels, with an external benchmark for one metric. No equations or steps reduce by construction to the inputs (no self-definitional definitions, no fitted parameters renamed as independent predictions). No load-bearing self-citations or ansatzes are present in the derivation. The claims are empirically grounded in comparisons to independent external references rather than internal tautology.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on standard supervised deep-learning assumptions plus the quality of the external training labels; no new physical entities or ad-hoc constants are introduced.

axioms (1)
  • domain assumption The Euclid Q1 source catalog and DESI DR1 spectroscopic redshifts provide sufficiently complete and unbiased training labels across the magnitude and color range of EDF-N.
    Invoked when training the detection, classification, and photo-z models; performance is reported relative to these labels.

pith-pipeline@v0.9.0 · 5695 in / 1280 out tokens · 28201 ms · 2026-05-13T18:36:18.036966+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. A systematic evaluation of vision-language models for observational astronomical reasoning tasks

    cs.AI 2026-04 accept novelty 7.0

    Vision-language models underperform specialized astronomical methods on real observational data, with accuracy improving when physical explanations are provided in prompts and when raw numerical measurements replace r...

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages · cited by 1 Pith paper · 1 internal anchor

  1. [1]

    Bishop, C. M. 1994, Technical Report NCRG/94/004 Brammer, G. B., van Dokkum, P. G., & Coppi, P. 2008, ApJ, 686, 1503, doi: 10.1086/591786 Burke, C. J., Aleo, P. D., Chen, Y.-C., et al. 2019, MNRAS, 490, 3952, doi: 10.1093/mnras/stz2845 Cai, Z., & Vasconcelos, N. 2018, in Proceedings of the IEEE conference on computer vision and pattern recognition, 6154–6...

    work page Pith review doi:10.1086/591786 1994

  2. [2]

    Data Release 1 of the Dark Energy Spectroscopic Instrument

    https://arxiv.org/abs/1910.01259 Deng, J., Dong, W., Socher, R., et al. 2009, in 2009 IEEE conference on computer vision and pattern recognition, IEEE, 248–255 18 DESI Collaboration, Karim, M. A., Adame, A. G., et al. 2025, arXiv e-prints, arXiv:2503.14745, doi: 10.48550/arXiv.2503.14745 Dey, A., Schlegel, D. J., Lang, D., et al. 2019, AJ, 157, 168, doi: ...

    work page internal anchor Pith review doi:10.48550/arxiv.2503.14745 1910