🔭 astrophysics

Photometric Redshifts in JWST Deep Fields: A Pixel-Based Alternative with DeepDISC

This paper demonstrates that DeepDISC, a deep learning algorithm utilizing raw NIRCam images from the JADES program, produces reliable photometric redshifts with accuracy comparable to or better than traditional template fitting methods while offering significantly faster processing speeds and eliminating the need for prior photometric measurements.

Original authors: Grant Merz, Ming-Yang Zhuang, Junyao Li, Qian Yang, Yue Shen, Xin Liu, John Franklin Crenshaw

Published 2026-02-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Grant Merz, Ming-Yang Zhuang, Junyao Li, Qian Yang, Yue Shen, Xin Liu, John Franklin Crenshaw

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 massive, dark library filled with billions of books (galaxies). To understand these books, astronomers need to know how far away they are. In astronomy, this distance is measured by something called "redshift." The further away a galaxy is, the more its light stretches out, shifting toward the red end of the spectrum.

For a long time, the only way to get a precise distance was to take a "spectrum" of the galaxy's light—a detailed chemical analysis. But this is like trying to read every single page of every book in the library by hand; it takes forever and you can only do it for the brightest, easiest-to-reach books.

To solve this, astronomers developed a shortcut called "photometric redshift" (photo-z). Instead of reading every page, they just look at the color of the book's cover. If it looks redder, it's probably further away. Traditionally, this has been done by measuring the brightness of the galaxy through specific colored filters (like looking at a book through red, blue, and green glasses) and comparing those numbers to a library of known galaxy templates.

The New Approach: DeepDISC

This paper introduces a new, faster, and more direct way to do this using a computer program called DeepDISC. Think of DeepDISC not as a calculator that crunches numbers, but as a super-smart security camera that looks at the actual picture of the galaxy.

Here is how it works, using simple analogies:

Looking at the Whole Picture, Not Just the Numbers:
Traditional methods take a photo, cut it up, measure the brightness of the galaxy in different colors, and then feed those numbers into a calculator. DeepDISC skips the measuring step. It looks at the raw pixels of the image itself. It's like the difference between describing a painting by listing the amount of blue and yellow paint used, versus an AI that looks at the actual painting to understand its style and age. DeepDISC sees the shape, the fuzziness, and the color gradients directly.
The "De-blending" Magic:
In deep space images, galaxies often overlap, like two people standing so close in a crowd that they look like one blob. Traditional methods struggle to separate them. DeepDISC is trained to act like a skilled editor who can look at a crowded photo, identify where one person ends and another begins, and assign a distance to each person individually.
Learning from Experience (Pre-training):
The researchers faced a problem: they didn't have enough "answer keys" (spectroscopic redshifts) to teach the AI how to guess distances for the new James Webb Space Telescope (JWST) images.
- The Analogy: Imagine trying to teach a student to identify rare birds in a new forest, but you only have 300 photos of those birds.
- The Solution: They first taught the AI on a massive dataset of other galaxies (from a different telescope) to learn what galaxies generally look like. Then, they fine-tuned it with the few JWST examples they had. They found that using a specific type of AI architecture (called ResNet) that had been "pre-trained" on galaxy images worked much better than trying to teach it from scratch or using a different type of AI (Transformers) that usually needs millions of examples.

What Did They Find?

The team tested DeepDISC on images from the JWST's "JADES" survey (a deep look into the early universe).

Speed: DeepDISC is incredibly fast. It can process 94,000 galaxies in just 4 minutes on a single computer chip. Traditional methods would take much longer because they have to measure every single galaxy individually first.
Accuracy: When they compared DeepDISC to the traditional "template" method (EAZY), DeepDISC performed just as well, and sometimes even better, especially when they used the same set of color filters for both.
The Catch (The "Training" Limit): The AI is only as good as the examples it was taught. If a galaxy looks very different from the ones in its training data (for example, a very faint or unusual galaxy), the AI might get confused. The researchers created a "quality flag" system to warn users: "This guess is based on a galaxy that looks like the ones we studied," or "This guess is a bit of a wild guess because we haven't seen many galaxies like this before."

The Bottom Line

This paper proves that you don't need to measure the brightness of every galaxy to know how far away it is. You can just show a smart computer the picture, and it can guess the distance by recognizing patterns in the pixels.

While the traditional method is still very good, DeepDISC offers a faster, more efficient alternative that is particularly good at handling crowded images and works well even when you don't have a huge library of "answer keys" to learn from. It's a promising new tool for mapping the vast, dark library of the universe.

Technical Summary: Photometric Redshifts in JWST Deep Fields: A Pixel-Based Alternative with DeepDISC

Problem Statement
Photometric redshift (photo-z) estimation is essential for large-scale astronomical surveys where spectroscopic follow-up is infeasible due to time and cost constraints. While template Spectral Energy Distribution (SED) fitting (e.g., EAZY) is the standard for high-redshift near-infrared data, it relies on pre-computed aperture photometry and is susceptible to degeneracies (e.g., Lyman vs. Balmer breaks) and systematic errors from source extraction. Machine learning (ML) approaches have shown promise in low-redshift optical surveys, often utilizing raw pixel information to capture morphological features. However, the efficacy of image-based algorithms on high-redshift, near-infrared data from the James Webb Space Telescope (JWST) remains underexplored. Key challenges include the scarcity of spectroscopic training samples, the risk of overfitting on small datasets, and the difficulty of handling blended sources or out-of-distribution objects in color-magnitude space.

Methodology
The authors apply DeepDISC (Detection, Instance Segmentation and Classification with Deep Learning), an image-based deep learning framework, to estimate photometric redshifts directly from multi-band JWST NIRCam images. The approach bypasses traditional source extraction and aperture photometry, operating instead at the pixel level to simultaneously detect, deblend, and classify sources while predicting redshift probability density functions (PDFs).

Data: The study utilizes JWST NIRCam imaging (9 filters: F090W–F444W) from the JADES DR2 GOODS-S field. Spectroscopic redshifts for training and testing are compiled from JADES DR3, ASTRODEEP-GS43, and MUSE-HUDFS, resulting in a reliable sample of 4,341 objects. The dataset is partitioned into sub-images (8.3" × 7.5") to ensure complete filter coverage.
Architecture: The model is built on a two-stage instance segmentation framework (based on Mask R-CNN) comprising a backbone, a Region Proposal Network (RPN), and Region of Interest (ROI) heads. The redshift head utilizes a Mixture Density Network (MDN) to output a Gaussian mixture model representing the redshift PDF.
Training Strategy: The authors investigate the impact of backbone architecture and pretraining:
- Backbones: They compare a ResNet50 (CNN) and a Multi-scale Vision Transformer (MViTv2).
- Pretraining: Models are tested with weights pretrained on ImageNet (terrestrial images) and the Galaxies ML (GML) dataset (HSC optical galaxy images).
- Simulation: JAGUAR simulated images are used to test scaling laws with larger datasets, though domain gaps between simulation and real data were noted.
Evaluation: Performance is assessed using bias, scatter ( $\sigma_{IQR}$ ), and outlier fraction ( $\eta$ ) against a held-out spectroscopic test set. Results are compared against EAZY template fitting using both the full JADES DR2 catalog (including HST data) and a restricted 9-filter subset (EAZY9) to ensure a fair comparison.

Key Contributions and Results

Pixel-Based Performance: DeepDISC successfully produces reliable photo-z estimates directly from JWST images up to $z \sim 8$ . The model generates a full redshift PDF for each source, capturing uncertainties and degeneracies.
Architecture and Pretraining:
- ResNet vs. Transformer: On the limited JWST training sample (~2,000 objects), the ResNet50 backbone outperformed the MViTv2 transformer. The transformer, while powerful, proved "data-hungry" and less robust to small dataset sizes.
- In-Domain Pretraining: Pretraining the ResNet50 on galaxy images (GML dataset) significantly improved performance compared to ImageNet pretraining or training from scratch. This suggests that learning galaxy-specific features (e.g., bulge/disk structures, color-redshift relations) is crucial for adapting CNNs to astronomical data.
Comparison with Template Fitting:
- When compared to EAZY using the same 9 NIRCam filters (EAZY9), DeepDISC outperformed the template-based method, achieving lower scatter and fewer outliers.
- When compared to the full JADES DR2 EAZY catalog (which includes HST data), DeepDISC showed slightly higher scatter but a comparable or better outlier fraction, demonstrating that pixel-level information can compensate for the lack of broader wavelength coverage.
Filter Robustness: DeepDISC demonstrated greater robustness to reduced filter sets (e.g., 4-filter COSMOS-Web or 6-filter NEXUS configurations) compared to EAZY, maintaining better point estimates with fewer filters.
Scalability and Limitations: The model scales with training set size, but performance is currently limited by the small volume of spectroscopic training data. The authors identified that objects in color-magnitude space underrepresented in the training set (identified via a Self-Organizing Map) yield less trustworthy estimates.

Significance and Claims
The paper claims that DeepDISC represents a viable, high-speed alternative to traditional template fitting for JWST data. A primary advantage is computational efficiency: the model can generate a catalog of 94,000 photo-zs in 4 minutes on a single GPU without the intermediate step of source extraction. The authors emphasize that while current spectroscopic samples are small and incomplete, the method demonstrates the potential of image-based approaches for future, larger JWST surveys and upcoming missions like Roman.

The study concludes that while deep learning photo-zs are competitive and often superior to template fitting when filter sets are matched, their applicability is currently constrained by the "out-of-distribution" problem inherent to ML methods. The authors advocate for the development of astronomy-specific foundation models and the use of generative data augmentation to bridge the gap between limited training data and the vast parameter space of the high-redshift universe. The work provides a catalog of photo-zs for all JADES DR2 GOODS-S sources, accompanied by quality flags indicating the reliability of the training data representation.

More like this