DeepRed: an architecture for redshift estimation

This paper introduces DeepRed, a deep learning pipeline that leverages diverse modern computer vision architectures to achieve state-of-the-art redshift estimation across galaxies, gravitational lenses, and supernovae, significantly outperforming existing methods on both simulated and real astronomical datasets while demonstrating robust generalization and interpretability.

The Gist

Imagine the universe is a giant, cosmic library. In this library, every book (a star, a galaxy, or a black hole) has a "page number" that tells us how far away it is and how fast it's moving away from us. Astronomers call this redshift.

The problem? Reading these page numbers is incredibly hard, expensive, and slow. It's like trying to read a book by holding it up to a massive, high-powered microscope one letter at a time. We need a faster way to read the whole book at a glance.

Enter DeepRed, a new "super-reader" built by scientists Alessandro Meroni and his team. Here is how it works, explained simply:

1. The Challenge: The "Blurry" Universe

When light travels from a distant galaxy to Earth, it gets stretched out, turning redder. This is the redshift. Sometimes, massive objects (like giant galaxies) act like a cosmic magnifying glass, bending the light of objects behind them. This creates weird shapes called Einstein Rings or Lenses.

The scientists wanted to build an AI that could look at a picture of these cosmic objects and instantly guess their "page number" (redshift) just by looking at the image, without needing the slow, expensive microscope (spectroscopy).

2. The Solution: A "Taste-Test" Panel (DeepRed)

Instead of building just one AI brain, the team built a panel of experts. Think of it like a cooking competition where you have four different chefs, each with a different style:

• The Classic Chef (ResNet): Good at recognizing standard patterns.
• The Efficient Chef (EfficientNet): Fast and great at spotting details without wasting energy.
• The Transformer Chef (SwinT): Good at seeing how different parts of the image relate to each other, like connecting the dots.
• The Mixer Chef (MLP-Mixer): A new style that mixes information in a unique way.

DeepRed is the "Head Judge." It lets all four chefs taste the image (analyze the galaxy) and give their own guess. Then, it takes their answers and blends them together into one final, super-accurate prediction. This "ensemble" method is like asking a group of experts instead of just one person; the group is almost always smarter than the individual.

3. The Training: From Simulations to Reality

To teach these chefs, the scientists used two types of ingredients:

• Simulated Ingredients (DeepGraviLens): They created millions of fake cosmic images on computers. These were perfect, clean images of Einstein rings and lensed supernovae (exploding stars).
• Real Ingredients (KiDS & SDSS): They used real photos taken by giant telescopes. These are "noisier" and messier, like trying to read a book in a windy, rainy night.

The AI learned on the fake images first, then was tested on the real ones to see if it could handle the chaos of the real universe.

4. The Results: A New Record

The results were amazing.

• Speed and Accuracy: DeepRed was significantly better than previous methods. On the simulated data, it improved accuracy by up to 55%. On real data, it improved by 16% to 27%.
• The "Why" Factor (Explainability): One of the biggest fears with AI is that it's a "black box"—it gives an answer, but we don't know why. The team used a tool called SHAP (think of it as a "highlighter pen").
  • When the AI looked at a galaxy, the highlighter showed exactly which pixels it was focusing on.
  • The Result: In 95% of cases, the AI was looking exactly at the galaxy or the lens, not at the empty space around it. This proved the AI wasn't cheating or guessing; it was actually learning the physics of the universe.

5. Why This Matters

Future telescopes (like the LSST) are going to take petabytes of data—millions of galaxies every night. Humans can't possibly look at them all.

DeepRed is the scalable, reliable robot librarian that can:

1. Read the "page numbers" of millions of galaxies in seconds.
2. Work on different types of cosmic objects (lenses, supernovae, normal galaxies).
3. Tell us why it made its guess, so astronomers can trust it.

In short, DeepRed turns the impossible task of mapping the entire universe into a manageable job, helping us understand how the universe expands and evolves, one pixel at a time.

Technical Summary

1. Problem Statement

Estimating redshift (the shift of spectral lines toward longer wavelengths) is fundamental in astrophysics for determining distances, mapping the 3D structure of the universe, and measuring expansion rates. However, obtaining spectroscopic redshifts is costly and time-consuming. While photometric redshift estimation (using multi-band images) is faster, current image-based methods face several challenges:

• Generalization: Existing models are often validated on homogeneous datasets and struggle to generalize across different morphologies (e.g., standard galaxies vs. complex gravitational lenses) and observational conditions (e.g., varying noise levels and Point Spread Functions).
• Transients: There is a lack of robust methods for estimating redshifts for rare, transient phenomena like gravitationally-lensed supernovae.
• Interpretability: Many deep learning models act as "black boxes," making it difficult to verify if they are focusing on relevant astronomical features or background noise.

2. Methodology: The DeepRed Pipeline

The authors propose DeepRed, a deep learning pipeline designed to estimate redshift directly from multi-band astronomical images. Rather than relying on a single custom architecture, DeepRed leverages a diverse ensemble of modern computer vision (CV) backbones.

A. Architecture Families

The pipeline trains and compares several distinct architectural families, adapting them from classification tasks to regression:

1. Convolutional Neural Networks (CNNs): ResNet, EfficientNet (B1–B3), and domain-specific baselines (A1, A3, NetZ, PhotoZ).
2. Transformers: Swin Transformer (SwinT), utilizing hierarchical self-attention to capture both local and global dependencies.
3. MLP-Mixer (MLPm): A non-convolutional architecture that processes spatial and channel information using Multi-Layer Perceptrons.
4. Traditional ML: A baseline using Histogram of Oriented Gradients (HOG) features combined with a Support Vector Regressor (SVR).

B. Ensemble Strategy (Linear Regression Ensemble - LRE)

The core innovation is the ensemble approach. Instead of selecting a single best model, DeepRed trains multiple architectures independently. Their output latent vectors are then combined using a Linear Regression Ensemble (LRE). The LRE learns optimal weights to combine the predictions of the base models, minimizing the Mean Squared Error (MSE).

C. Datasets

The pipeline was evaluated on four simulated datasets and two real-world datasets:

• Simulated (DeepGraviLens - DGL): Four subsets (LSST-wide, DESI-DOT, DES-deep, DES-wide) containing ~15,000 images each of galaxy-galaxy lenses, Type Ia lensed supernovae, and core-collapse lensed supernovae.
• Real (KiDS): ~4,500 candidate gravitational lenses (low and high probability) with photometric redshift ground truth.
• Real (SDSS): ~517,000 non-lensed galaxies with spectroscopic redshift ground truth.

D. Explainability (XAI)

To ensure reliability, the authors integrated SHapley Additive exPlanations (SHAP). They quantitatively evaluated the models by:

1. Generating SHAP heatmaps to identify the most influential pixels.
2. Calculating Localization Accuracy: The percentage of times the most influential pixel falls within the ground-truth bounding box of the astronomical object.
3. Measuring the distance between the peak SHAP value and the object center.

3. Key Contributions

1. Systematic Adaptation of Generic CV Architectures: DeepRed is the first framework to systematically apply generic, state-of-the-art CV architectures (EfficientNet, SwinT, MLP-Mixer) to redshift estimation, demonstrating they outperform domain-specific custom networks (like A1/A3) without specialized hand-crafted designs.
2. Heterogeneous Generalization: The study validates performance across a wide spectrum of morphologies, including standard galaxies, Einstein rings, and lensed supernovae, proving the pipeline's suitability for future large-scale surveys (e.g., LSST).
3. Quantitative Explainability: Unlike previous works that used XAI qualitatively, DeepRed introduces quantitative metrics (localization accuracy >95% on clean data) to validate that models focus on the correct astronomical objects rather than background noise.
4. Robust Ensemble Performance: The LRE strategy consistently yields state-of-the-art results, often surpassing the best individual architecture.

4. Key Results

The DeepRed pipeline achieved State-of-the-Art (SOTA) results across all datasets, significantly outperforming baselines (HOG+SVR, A1, A3, NetZ, PhotoZ).

• Simulated Datasets (DGL):

  • NMAD Improvement: Achieved improvements of 46% to 55% over the best baseline (PhotoZ).
  • Best Performers: EfficientNet excelled on the noisy DES-deep dataset (55% improvement), while the Ensemble performed best on DES-wide, DESI-DOT, and LSST-wide.
  • Outlier Reduction: Reduced catastrophic outlier rates by up to 64% on challenging datasets.

• Real Datasets (KiDS):

  • NMAD Improvement: Outperformed the best baseline (NetZ) by 16% on the general test set and 27% on high-probability lenses.
  • Robustness: Successfully generalized from low-probability candidates to high-probability lenses despite the "label noise" inherent in photometric redshift ground truths.

• Real Datasets (SDSS):

  • MLP-Mixer Dominance: The MLP-Mixer architecture achieved a 5% improvement in MAE and NMAD over the best baselines (A3, NetZ) for non-lensed galaxies.
  • Efficiency: Demonstrated that complex architectures can generalize well even when trained on different data distributions (KiDS to SDSS).

• Explainability Metrics:

  • Localization Accuracy: Models achieved >95% accuracy in focusing on the target object for high-quality images (DES-wide, DESI-DOT, LSST-wide, KiDS, SDSS).
  • Noise Sensitivity: Accuracy dropped to ~38% on the noisy DES-deep dataset, correctly reflecting the difficulty of the task and the spread of the lens across the image.

5. Significance and Conclusion

DeepRed represents a paradigm shift in astronomical redshift estimation by moving away from custom, task-specific neural networks toward modular, generic computer vision architectures combined with ensemble learning.

• Scalability: The pipeline is robust enough to handle the heterogeneous data expected from next-generation surveys like the Large Synoptic Survey Telescope (LSST), which will generate petabytes of data containing rare transients and complex lenses.
• Trustworthiness: By integrating quantitative explainability, DeepRed ensures that predictions are derived from relevant physical features, addressing the "black box" concern in astrophysics.
• Future Impact: The work establishes a foundation for applying interpretable deep learning to multimodal astronomical data, paving the way for automated discovery and characterization of cosmological phenomena in the era of big data astronomy.