Extracting redshifts from 2D slitless spectroscopic images using deep learning for the CSST galaxy survey Facilities: CSST, HSC, DESI
Xingchen Zhou, Yan Gong, Xin Zhang, Xian-Min Meng, Haitao Miao, Run Wen, Nan Li · 2026
Read Extracting redshifts from 2D slitless spectroscopic images using deep learning for the CSST galaxy survey Facilities: CSST, HSC, DESI by Xingchen Zhou, Yan Gong, Xin Zhang, Xian-Min Meng, Haitao Miao, Run Wen, Nan Li (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-2605-16762.txt.
Introduction
Zhou and colleagues extract galaxy redshifts straight from 2D slitless spectral images for the CSST galaxy survey, skipping 1D extraction. Spectra spread onto the detector mix with galaxy shape, making wavelength calibration hard. Their Bayesian neural network, built with Monte Carlo dropout, maps images to redshifts with error bars. For sources with signal-to-noise above 1, scatter reaches 0.0104, dropping to 0.0024 at signal-to-noise above 10, matching needs for baryon acoustic oscillation studies.
Catalog record
| CREATOR | Xingchen Zhou, Yan Gong, Xin Zhang, Xian-Min Meng, Haitao Miao, Run Wen, Nan Li |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 44256 |
| LICENSE | http://creativecommons.org/publicdomain/zero/1.0/ |
| RIGHTS | CC0 |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.16762
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Extracting redshifts from 2D slitless spectroscopic images using deep learning for the CSST galaxy survey Facilities: CSST, HSC, DESI Xingchen Zhou, Yan Gong, Xin Zhang, Xian-Min Meng, Haitao Miao, Run Wen, Nan Li Abstract Wide-field slitless spectroscopic galaxy surveys, such as the one performed by the upcoming Chinese Space Station Survey Telescope (CSST), are crucial for precision cosmology but present formidable data analysis challenges. Because spectra are dispersed directly onto the detector, they are convolved with the 2-dimensional (2D) spatial morphology, which complicates wavelength calibration and consequently degrades the fidelity of subsequent 1-dimensional (1D) spectral extraction. To overcome these limitations, we present a deep learning framework that extracts redshifts directly from 2D slitless spectral images, bypassing 1D extraction entirely. We construct a realistic mock dataset for the CSST G V GV and G I GI band using high-resolution images from HSC-SSP PDR3 and spectral energy distributions (SEDs) from DESI DR1. A Bayesian convolutional neural network implemented by Monte Carlo dropout is employed to map the 2D spectral images to redshift estimations while simultaneously quantifying uncertainties. We find that our model can achieve a precision σ NMAD = 0.0104 \sigma_{\rm NMAD}=0.0104 and mean uncertainty ⟨ E / ( 1 + z true ) ⟩ = 0.0155 \langle E/(1+z_{{\rm true}})\rangle=0.0155 for sources with SNR G I ≥ 1 {\rm SNR}_{GI}\geq 1 . For sources with SNR G I {\rm SNR}_{GI} higher than 3.0, 5.0 and 10.0, σ NMAD \sigma_{\rm NMAD} can achieve 0.0047, 0.0037 and 0.0024 respectively, matching the redshift precision requirements for studies such as BAO using the CSST slitless spectroscopic surveys. Furthermore, by utilizing spatial augmentations, the network demonstrates resilience to wavelength calibration errors. This work provides a novel and robust pathway for data analysis of next-generation slitless spectroscopic galaxy surveys. show][email protected] I Introduction Mapping the large-scale structure (LSS) of the Universe and tracing the evolution of galaxies across cosmic time fundamentally relies on acquiring accurate redshift measurements for vast number of extragalactic sources ( 40 ; 30 ) . While photometric surveys can rapidly derive redshifts (photo- z ) for billions of galaxies, these measurements inherently suffer from broad uncertainties and catastrophic outliers due to the low resolution of broadband filters ( 39 ; 35 ) . In contrast, spectroscopic observations capture the detailed emission and absorption features of galactic spectra, yielding accurate spectroscopic redshifts for high-precision cosmological constraints by Baryon Acoustic Oscillations (BAO) ( 4 ; 15 ) and Redshift-Space Distortions (RSD) ( 22 ) , as well as detailed characterization of galaxy kinematics and stellar populations ( 8 ; 17 ; 7 ) . Despite high spec- z precision, traditional multi-object spectroscopic surveys, which utilize fibers or slit masks, are fundamentally bottlenecked by target pre-selection and the mechanical limitations of placing fibers in dense focal planes ( 5 ; 11 ) . To overcome these challenges, space-base slitless spectroscopy has emerged as a powerful solution. By using slitless instruments, they can perform wide-field and highly multiplexed spectroscopic observations ( 41 ; 6 ; 27 ) . Ongoing and upcoming space-based missions, e.g. Euclid ( 3 ) , Nancy Grace Roman Space Telescope ( 33 ) , Chinese Space Station Survey Telescope (CSST, 19 ; 20 ; 9 ), utilize slitless spectroscopic surveys to capture the spectra of the luminous sources within their expansive fields of view, completely free from atmospheric interference. However, the rich data product of slitless spectroscopy introduces great challenges in data analysis. Because there is no slit to disperse light with high spectral resolution, the spectrum recorded on the detectors is 2D morphological shape convolved with point spread function (PSF). In densely populated survey fields, this leads to severe spectral overlapping and cross-contamination among neighboring objects. Furthermore, the slitless spectroscopy introduces severe complications for accurate wavelength calibration. Without a fixed spatial reference, the zero-point of the wavelength for each dispersed spectrum depends fundamentally on the exact spatial location and morphological centroid of the source. Any asymmetry in the galaxy or uncertainty in its position directly propagates into systematic wavelength shifts. Traditional data reduction pipelines attempt to calibrate these shifts, deblend overlaps, and extract 1D spectra from the 2D detector images, but this extraction process is highly complex, prone to compounding systematic errors, and inevitably results in a significant loss of spectral information ( 6 ; 26 ; 32 ) . For CSST, these challenges are more severe. Unlike other slitless spectroscopic surveys that typically acquire direct, undispersed images of the target field alongside the dispersed observations to pinpoint source coordinates, the specific focal plane design of CSST excludes simultaneous direct imaging during survey ( 9 ) . Consequently, establishing the astrometric centroid of a source and the zero-point required to anchor the wavelength, necessitates cross-matching the dispersed spectra with separately and independently acquired photometric catalogs. Any inherent astrometric uncertainties, such as focal plane distortions and alignment errors between these distinct observations propagate directly into systematic shifts in the assigned wavelengths, severely affecting the 1D extraction pipelines. On the other hand, the issue of spectral overlapping can be safely neglected, since the estimated source density is only \sim 2-3 galaxies per \rm arcmin^{2} ( 19 ) . Deep learning has rapidly emerged as a powerful tool to overcome these data analysis challenges. In particular, convolutional neural networks (CNNs) are exceptionally well-suited for astronomical imaging due to their ability to learn hierarchical and non-linear feature representations directly from complex and high-dimensional data ( 28 ; 13 ; 25 ) . Unlike traditional parametric modeling, which relies on rigid assumptions and hand-crafted features, deep learning models can automatically identify optimal mappings between noisy observational inputs and physical parameters, making them highly resilient to complex systematics and survey-specific artifacts. Leveraging deep learning allows us to fundamentally rethink how slitless spectra are analyzed: by operating directly on the 2D spectral images. This approach bypasses the complicated and error-prone 1D extraction step entirely. By treating the dispersed 2D image as the primary data product, CNNs can jointly utilize both the spectral dispersion features and the spatial morphological information, maximizing the outcomes from low signal-to-noise ratio (SNR) observations. In this work, we present a deep learning framework to estimate redshifts directly from 2D slitless spectral images, tailored specifically for the upcoming CSST spectroscopic survey. Due to the focal plane design of CSST, we concentrate on wavelength calibration. To train and validate our models, we construct a highly realistic mock dataset of 2D spectral images, combining high-resolution imaging from the Hyper Suprime-Cam (HSC, 2 ; 1 ) with precise spectral templates from the Dark Energy Spectroscopic Instrument (DESI, 10 ; 12 ). Recognizing the critical importance of error characterization in precision cosmology, we employ the Bayesian neural network (BNN, 34 ; 18 ). This network allows our model not only predict accurate point estimates for the spec- z s, but also provide robust and statistically reasonable epistemic and aleatoric uncertainties for every measurement. This paper is organized as follows: Section II describes the generation of the mock slitless spectroscopic dataset. In Section III , we discuss the details of the neural network architectures employed in this study. We present the performance of our framework in Section IV , followed by several discussions in Section V . Finally, we