🔭 astrophysics

Beyond Colors: Probing Redshifts from Galaxy Morphology in Single-band Images with ViT-MDNz

This paper introduces ViT-MDNz, a deep learning model that leverages a Vision Transformer and Mixture Density Network to accurately estimate galaxy redshifts and their uncertainties from single-band images by exploiting the correlation between morphology and redshift, thereby offering a practical solution for surveys with limited photometric coverage.

Original authors: Zhijian Luo, Yangyang Li, Jianzhen Chen, Qishen Cao, Duo Cao, Shaohua Zhang, Hubing Xiao, Chenggang Shu

Published 2026-02-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhijian Luo, Yangyang Li, Jianzhen Chen, Qishen Cao, Duo Cao, Shaohua Zhang, Hubing Xiao, Chenggang Shu

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 you are looking at a vast, crowded city from a helicopter, but you can only see the buildings in black and white. You can't see the colors of the cars, the paint on the houses, or the signs on the shops. Usually, to guess how far away a building is, you'd look at its color (because distant objects often look redder due to the expansion of the universe). But here, you have no color. You only have the shape and structure of the buildings.

This is the challenge astronomers face when they look at distant galaxies with only a single "color" (or band) of light. Usually, they need a rainbow of colors to guess how far away a galaxy is. If they only have one, the guess is usually terrible.

This paper introduces a new, super-smart AI detective named ViT-MDNz that solves this problem by looking at shapes instead of colors.

Here is the story of how it works, broken down into simple parts:

1. The Problem: The "Colorless" Mystery

In astronomy, to measure how far away a galaxy is (its "redshift"), scientists usually need a full spectrum of light (like a rainbow). It's like trying to identify a fruit by taste; if you can taste the sweet, sour, and bitter notes, you know exactly what it is.

The Issue: Sometimes, telescopes only capture one "slice" of that rainbow (a single black-and-white image). Without the other colors, the "taste" is missing, and traditional methods fail.
The Old Way: If data is missing, scientists often just throw those galaxies away. This is like throwing away a whole basket of apples just because you can't taste the skin. It wastes valuable data.

2. The Solution: The "Shape Detective"

The authors realized that even without color, the shape of a galaxy tells a story.

The Analogy: Think of a galaxy like a person. As people age, they change. A young person might be energetic and spread out, while an older person might be more compact or have different posture.
The Discovery: In the universe, galaxies change shape as they get farther away (older in terms of light travel time). Distant galaxies tend to look smaller, fuzzier, and more irregular. Nearby galaxies look bigger and more structured (like perfect spirals or round blobs).
The AI's Job: The new AI, ViT-MDNz, is trained to be a master of body language. It ignores the missing colors and focuses entirely on the "posture" and "structure" of the galaxy to guess its distance.

3. How the AI Works (The Two-Part Brain)

The model has two special parts working together, like a detective team:

Part A: The "Eagle Eye" (Vision Transformer or ViT)
Imagine a detective looking at a crime scene photo. Instead of looking at just one spot, this detective looks at the entire photo at once, understanding how the left side relates to the right side.
- In the paper, this is the Vision Transformer. It scans the single black-and-white galaxy image, breaking it into tiny puzzle pieces, and figures out the global shape and structure. It learns: "Ah, this galaxy is very small and clumpy; that usually means it's far away."
Part B: The "Gambler" (Mixture Density Network or MDN)
Traditional AI gives you one single answer: "This galaxy is 1 billion light-years away." But what if the AI is wrong?
- This part is the Mixture Density Network. Instead of giving one number, it acts like a wise gambler who says: "I think it's 1 billion, but there's a 70% chance it's between 0.9 and 1.1 billion, and a 30% chance it's actually 0.8 billion."
- It gives a probability map. It doesn't just guess a distance; it tells you how confident it is in that guess.

4. The Results: A Miracle with Limited Data

The team tested this AI on 300,000 galaxies from the DESI survey (a massive map of the universe).

The Test: They fed the AI only single-band (black and white) images.
The Outcome: Even without color, the AI guessed the distances with surprising accuracy.
- It was accurate enough to be useful for major scientific studies.
- It was so good that it could even handle cases where the data was missing or corrupted, saving galaxies that would have otherwise been thrown in the trash.
The Bonus: When they did give the AI color information (multi-band), it got even better, proving it can use both shape and color when available.

5. Why This Matters

Imagine you are building a 3D map of the universe. If you throw away every galaxy that doesn't have a full color spectrum, your map has huge holes in it.

The Impact: This new method allows astronomers to fill in those holes. It means we can use every single image we take, even the "incomplete" ones.
The Future: As we launch bigger telescopes (like the LSST or the Chinese Space Station Telescope) that will take billions of pictures, many of those pictures might only have one color band. This AI ensures we won't lose that data. It turns "bad" data into "good" science.

In a Nutshell

This paper is about teaching a computer to guess how far away a galaxy is just by looking at its shape, even when it can't see its color. It's like being able to guess someone's age just by looking at their silhouette in the dark. It's a clever, statistical trick that saves valuable data and helps us build a more complete picture of our universe.

1. Problem Statement

Traditional photometric redshift ( $z_{phot}$ ) estimation relies heavily on multi-band photometric data (typically $\ge 4$ bands) to capture spectral energy distribution (SED) features like the 4000Å break or Lyman break. However, in many observational scenarios (e.g., data gaps, foreground contamination, or specific survey constraints), only single-band images are available.

The Challenge: Single-band images lack color information, making it impossible to break the degeneracy between redshift, galaxy type, and luminosity using traditional SED-fitting or standard machine learning methods.
Current Limitations: Existing mitigation strategies (e.g., Bayesian imputation) introduce systematic errors, while discarding incomplete data leads to sample loss and selection bias.
The Goal: To determine if statistically reliable redshift estimates can be derived solely from morphological features (shape, size, structure) present in single-band images, leveraging statistical priors learned from large-scale data.

2. Methodology: ViT-MDNz

The authors propose a novel hybrid deep learning architecture named ViT-MDNz, which integrates a Vision Transformer (ViT) for feature extraction and a Mixture Density Network (MDN) for probabilistic regression.

A. Data Source and Preprocessing

Dataset: ~300,000 galaxies from the DESI Legacy Imaging Surveys (DESI-LS) DR9, cross-matched with spectroscopic redshifts from SDSS DR16.
Sample Characteristics: The sample is dominated by low-redshift galaxies ( $z \lesssim 1$ ), with a bimodal distribution of Main Galaxy Samples (MGS) and Luminous Red Galaxies (LRGs).
Preprocessing:
- Images are center-cropped from $256 \times 256$ to $192 \times 192$ pixels.
- Pixel values are normalized to the range $[-1, 1]$ .
- Data is split into 80% training and 20% testing.

B. Architecture Design

ViT Encoder (Feature Extraction):
- Input: Single-band galaxy images ( $192 \times 192 \times 1$ ).
- Mechanism: The image is split into $16 \times 16$ patches (yielding 576 patches). Unlike standard ViTs, the authors omit the classification token and instead use global average pooling over the patch sequence after the Transformer encoder layers.
- Configuration: 8 Transformer encoder layers, 4 attention heads, hidden dimension $D=128$ .
- Advantage: The self-attention mechanism captures global dependencies and long-range morphological correlations (e.g., between bulge and spiral arms) more effectively than Convolutional Neural Networks (CNNs).
Mixture Density Network (MDN) (Probabilistic Output):
- Input: The 128-dimensional global feature vector from the ViT.
- Output: Instead of a single point estimate, the MDN predicts the full conditional probability density function (PDF) of the redshift, $p(z|x)$ .
- Model: A Gaussian Mixture Model (GMM) with $K=15$ components.
- Parameters: The network outputs mixture weights ( $\pi_k$ ), means ( $\mu_k$ ), and variances ( $\sigma^2_k$ ) for each component.
- Loss Function: Negative Log-Likelihood (NLL) is minimized to train the model to concentrate the predicted distribution around the true spectroscopic redshift.

3. Key Contributions

First ViT Application in Photometric Redshift: This is the first study to introduce the Vision Transformer architecture into the field of photometric redshift estimation, demonstrating its superiority in extracting deep morphological features from single-band images.
Morphology-Redshift Correlation: The study validates that morphological features (angular size, concentration, surface brightness profiles) contain statistically significant information about redshift, even without color data.
Probabilistic Framework: By using an MDN, the model provides not just a point estimate but a full uncertainty quantification (PDF), which is crucial for cosmological statistical analyses.
Flexibility: The architecture is designed to handle both single-band (data-limited) and multi-band (data-rich) scenarios seamlessly.

4. Experimental Results

The model was evaluated on the DESI-LS $r$ -band (and $g, z$ bands) for redshifts $z \lesssim 1$ .

A. Single-Band Performance

Accuracy: In the $r$ -band, the model achieved a Normalized Median Absolute Deviation ( $\sigma_{NMAD}$ ) of 0.034 and a catastrophic outlier fraction ( $f_{out}$ ) of 2.6%.
Comparison: While a standard 5-band color-based method (SDSS) achieves $\sigma_{NMAD} \approx 0.022$ , the ViT-MDNz single-band result is remarkably competitive given the lack of color information.
Bias: The bias is negligible ( $\approx -0.0012$ ).
Calibration:
- PIT (Probability Integral Transform): The distribution is nearly uniform, indicating well-calibrated uncertainties.
- CRPS (Continuous Ranked Probability Score): Mean CRPS of 0.0377, indicating the predicted PDFs are tightly concentrated around the true values.
- Multimodality: Only ~0.7% of samples showed significant multimodal distributions, suggesting the morphological features effectively resolve redshift degeneracies for most galaxies.

B. Multi-Band Extension

When extended to multi-band inputs (concatenating $g, r, z$ channels):

Performance Gain: The $\sigma_{NMAD}$ improved to 0.020, and the outlier fraction dropped to 1.5%.
CRPS: Reduced to 0.0244.
Conclusion: The model effectively fuses color information with morphological features, proving its versatility.

5. Significance and Future Outlook

Scientific Impact: This work challenges the conventional paradigm that photometric redshifts must rely on multi-band colors. It provides a viable "rescue" method for galaxy samples with incomplete band coverage, enhancing the completeness of redshift catalogs for future surveys like LSST, Euclid, and the Chinese Space Station Telescope (CSST).
Methodological Shift: It demonstrates that deep learning can learn complex, non-linear mappings between galaxy morphology and redshift directly from raw pixels, bypassing the need for hand-crafted features.
Limitations:
- Accuracy is still lower than high-quality multi-band methods due to the fundamental information limit of single-band images.
- The "black-box" nature of the model makes physical interpretability difficult (i.e., which specific morphological markers drive the prediction is unclear).
- Generalization to very high redshifts ( $z > 1$ ) or peculiar morphologies requires further training with more diverse datasets.
Future Directions: The authors suggest incorporating Explainable AI (XAI) to visualize decision-making, using physics-guided constraints, and developing lightweight architectures for real-time application.

In summary, ViT-MDNz represents a significant step forward in astrostatistics, proving that morphological features alone can yield statistically robust redshift estimates, thereby maximizing the scientific return from data-limited astronomical surveys.