🔭 astrophysics

SAGUI: SED-based Segmentation of Multi-band Galaxy Images -- Application to JADES in GOODS-South

This paper introduces SAGUI, a modular framework that extends spectro-spatial analysis to multi-band galaxy imaging by combining starlet-based decomposition, spectral-similarity partitioning, and copula-based statistical treatment to effectively segment diverse structures—including clumps, bars, and faint low-surface-brightness features—as demonstrated on James Webb Space Telescope data from the GOODS-South field.

Original authors: Rafael S. de Souza (for the COIN collaboration), Andressa Wille (for the COIN collaboration), Shravya Shenoy (for the COIN collaboration), Aarya A. Patil (for the COIN collaboration), Alberto Krone-Ma

Published 2026-04-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Rafael S. de Souza (for the COIN collaboration), Andressa Wille (for the COIN collaboration), Shravya Shenoy (for the COIN collaboration), Aarya A. Patil (for the COIN collaboration), Alberto Krone-Martins (for the COIN collaboration), Ana L. Chies-Santos (for the COIN collaboration), Celine Boehm (for the COIN collaboration), Reinaldo R. Rosa (for the COIN collaboration), Thallis Pessi (for the COIN collaboration), Emille E. O. Ishida (for the COIN collaboration), Kristen C. Dage (for the COIN collaboration), Lilianne Nakazono (for the COIN collaboration), Phelipe Darc (for the COIN collaboration), Rupesh Durgesh (for the COIN collaboration)

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 giant, complex painting of a galaxy. To the naked eye, it looks like a swirling mix of colors and light. But if you want to understand the story behind the painting—where the stars were born, how old they are, and how they move—you need to break the image down into meaningful pieces.

For a long time, astronomers had two main ways to do this:

The "Bucket" Method: They would take the whole galaxy and measure its average light, like pouring the whole painting into a bucket and tasting the soup. You get a general flavor, but you lose all the details.
The "Grid" Method: They would slice the galaxy into a rigid grid (like a checkerboard) or use a Voronoi tessellation (like a honeycomb). This helps with the math, but it cuts right through interesting features. It might slice a spiral arm in half or merge a young star cluster with an old one just because they happened to be next to each other.

Enter SAGUI.

The paper introduces a new tool called SAGUI (SED-based Segmentation of Multi-band Galaxy Images). Think of SAGUI not as a knife that cuts the galaxy, but as a smart, color-sensing detective that groups pixels based on their "personality" rather than just their location.

Here is how it works, using some everyday analogies:

1. The "Starlet" Mask: Finding the Shape in the Noise

Imagine you are trying to find a specific person in a crowded, foggy room. You can't just look at every single speck of dust (noise).

What SAGUI does: It uses a mathematical trick called a "Starlet transform." Imagine taking the galaxy image and looking at it through a series of different zoom lenses.
- At the highest zoom, you just see static and grain (noise).
- At medium zooms, you start to see the swirls of the arms and the clumps of stars.
- At the lowest zoom, you see the smooth, overall shape.
The Result: SAGUI ignores the grainy static and the overly smooth background. It builds a "mask" that highlights only the actual shape of the galaxy, effectively drawing a boundary around the "real" stuff and ignoring the empty space.

2. The "Spectral" Detective: Grouping by Personality

Once SAGUI knows where the galaxy is, it looks at every single pixel (the tiny dots that make up the image).

The Old Way: "You are in this square, so you belong to this group."
The SAGUI Way: "Let's look at your outfit."
- In astronomy, a pixel's "outfit" is its SED (Spectral Energy Distribution). This is like a fingerprint made of light. A pixel with hot, young blue stars has a different light fingerprint than a pixel with cool, old red stars.
The Clustering: SAGUI asks every pixel to introduce itself. "I am a pixel with a blue, hot fingerprint." "I am a pixel with a red, dusty fingerprint." It then groups pixels that have similar fingerprints together, regardless of where they are on the map.
- Analogy: Imagine a party where people are sorted not by which room they are in, but by what music they like. The "Rock" group might be scattered across the room, but SAGUI gathers them all together into one "Rock Zone."

3. The "Copula" Trick: Seeing the Ghosts

Sometimes, galaxies have faint, ghostly tails or bridges connecting them. These are so dim that they look like background noise. Standard tools often delete them.

The Problem: If you just look at brightness, these ghosts disappear.
The SAGUI Solution: The paper uses a statistical trick called a Copula transform.
- Analogy: Imagine you are trying to hear a whisper in a noisy room. If you just turn up the volume, you hear more noise. But if you listen for a pattern—a specific rhythm that repeats across different microphones (different color filters)—you can isolate the whisper.
- SAGUI looks for pixels that are faint but consistently faint in the same way across all the different color filters. This allows it to "see" the faint tidal bridges and tails that other methods miss.

Why Does This Matter?

The authors tested SAGUI on 11 different galaxies from the James Webb Space Telescope (JWST), including messy mergers, barred spirals, and interacting pairs.

Before SAGUI: You might have a map where a bar and a spiral arm are mixed together, giving you a confusing average age for that area.
With SAGUI: The tool cleanly separates the bar (which is old and dusty) from the spiral arms (which are young and star-forming). It creates a map that respects the galaxy's actual structure.

The Big Picture

The paper argues that to understand how galaxies evolve, we need to stop treating them as blurry blobs or rigid grids. We need to treat them as complex ecosystems where different parts have different histories.

SAGUI is the tool that finally lets us slice the galaxy the way nature intended: by grouping together the parts that are truly related, revealing the hidden story of star birth, death, and cosmic collisions. It's like upgrading from a black-and-white sketch of a galaxy to a high-definition, 3D map where every neighborhood is correctly identified.

1. Problem Statement

The advent of high-quality multi-band imaging from facilities like the James Webb Space Telescope (JWST) and upcoming surveys (LSST, Euclid, CSST) has created vast datasets where each pixel contains a Spectral Energy Distribution (SED). While Integral Field Spectroscopy (IFS) has enabled spatially resolved studies of galaxy evolution, traditional methods for analyzing multi-band imaging data face significant limitations:

Noise Sensitivity: Pixel-level SED fitting is highly sensitive to noise, particularly in low-surface-brightness (LSB) regions.
Inadequate Binning: Common strategies like Voronoi tessellation improve signal-to-noise (S/N) ratios by grouping pixels based on geometric proximity or flux thresholds. However, these methods often fragment physically coherent structures or merge spectrally distinct populations, sacrificing spectral homogeneity for statistical robustness.
LSB Detection: Faint, diffuse structures (e.g., tidal tails, bridges) are often lost in standard noise filtering or background subtraction because they rely on absolute flux contrast rather than spectral coherence.

There is a need for a segmentation framework that partitions galaxy light into regions that are spectrally coherent (similar SED shapes) while preserving morphological structure, without relying solely on S/N-driven binning.

2. Methodology: The SAGUI Framework

The authors introduce SAGUI (SED-based Segmentation of Multi-band Galaxy Images), a modular framework extending the spectro-spatial paradigm of their previous IFS tool, capivara, to broadband imaging. The method operates in two primary stages, followed by a specialized extension for LSB features.

A. Morphology-Aware Spatial Masking (Starlet Decomposition)

Before clustering, the framework defines a foreground support to isolate the galaxy from the background while preserving multi-scale structures.

White-Light Construction: A white-light image is created by summing all available bands.
Starlet Transform: An isotropic undecimated wavelet transform (starlet transform) is applied. This decomposes the image into a hierarchy of detail maps ( $w_j$ ) at different scales and a coarse residual ( $c_J$ ).
Masking: The finest scale ( $j=1$ , dominated by noise) and the coarsest scale ( $c_J$ , dominated by background variations) are excluded. The intermediate scales are reconstructed to create a support image ( $R$ ).
Binarization: A binary mask ( $M$ ) is generated where $R > 0$ , effectively isolating the galaxy's structural components while suppressing background-dominated pixels.

B. Spectral Segmentation (Hierarchical Clustering)

Once the spatial mask is defined, the algorithm partitions the pixels based on their SED similarity rather than geometric proximity.

SED Representation: Each pixel is represented by a normalized flux vector across all available bands.
Normalization: Fluxes are centered by subtracting the median flux per pixel to make the clustering sensitive to SED shape rather than absolute surface brightness.
Distance Metric: Pairwise dissimilarity is calculated using the Euclidean distance ( $\ell_2$ norm) between normalized SED vectors.
Clustering: A hierarchical agglomerative clustering algorithm using Ward's criterion is applied. This minimizes the increase in within-cluster variance, merging pixels into groups that share similar photometric behavior.
Output: The dendrogram is cut at a desired number of clusters ( $K$ ), assigning a label to each pixel.

C. Extension for Low-Surface-Brightness (LSB) Features: Copula Transform

To detect faint, diffuse structures (like tidal bridges) that are spectrally coherent but faint in absolute flux, the authors introduce a Copula-based transform:

Scale Invariance: Instead of using raw flux, the method uses the probability integral transform to map pixel intensities in each band to a uniform distribution, then to a standard normal distribution (Gaussian margins).
Dependence Modeling: This removes marginal distributions, focusing purely on the dependence structure (correlation) between bands.
Energy Map: A scalar "energy" map is constructed by summing the squares of positive deviations in the transformed space. This amplifies correlated low-level emission across bands while suppressing uncorrelated noise, making faint LSB features detectable before the starlet masking step.

D. Physical Property Inference

Following segmentation, the framework performs Spatially Resolved SED Fitting:

Integration: Fluxes from all pixels within a segmented region are summed to create an integrated SED for that region.
Fitting: The Prospector code (using FSPS models) fits these integrated SEDs to derive physical parameters: stellar mass, star formation rate (SFR), metallicity, dust attenuation ( $A_V$ ), and mass-weighted age.
Mapping: These derived parameters are mapped back to the spatial grid, creating 2D physical property maps that respect the morphological boundaries defined by the segmentation.

3. Key Contributions

Novel Framework: Introduction of sagui, the first modular framework specifically designed for SED-based segmentation of multi-band imaging data, bridging the gap between IFS analysis and broadband photometry.
Spectral Coherence over Geometry: Demonstrates that clustering in SED space yields regions that are more physically meaningful than geometric binning (Voronoi), preserving distinct stellar populations even when they are spatially adjacent.
LSB Recovery via Copulas: Proposes a novel application of Copula transforms in astronomy to detect faint, spectrally coherent structures (tidal bridges) that are invisible to standard flux-thresholding methods.
Open Source: The code is released under an MIT license on GitHub, promoting reproducibility and community adoption.

4. Results and Case Studies

The method was applied to 11 morphologically diverse galaxies from the JADES survey in the GOODS-South field (redshifts $z \approx 0.36 - 1.09$ ).

Morphological Diversity: The framework successfully segmented complex structures including:
- Bars and Spirals: In Sagui-10 (barred spiral), sagui cleanly separated the bar, spiral arms, and central bulge, revealing that the bar hosts older stellar populations with lower SFR compared to the arms.
- Interacting Systems: In Sagui-7 (merging pair) and Sagui-4 (post-merger), the algorithm disentangled overlapping components and identified star-forming clumps triggered by the interaction.
- Crowded Fields: In the Sagui-1/2/3 triplet, the method successfully separated three distinct galaxies at different redshifts that appeared projected close together, avoiding artificial merging of their components.
Performance Metrics:
- Compared to Voronoi binning and piXedfit-like baselines, sagui produced regions with lower spectral dissimilarity (higher internal spectral homogeneity) while maintaining comparable or higher mean cluster S/N.
- The segmentation naturally followed physical structures (e.g., spiral arms, tidal tails) rather than arbitrary geometric shapes.
LSB Case Study (Sagui-11):
- A pair of interacting galaxies connected by a faint tidal bridge.
- Standard segmentation missed the bridge.
- Applying the Copula transform prior to segmentation successfully recovered the bridge, revealing it as a coherent structure with low SFR and low metallicity, distinct from the main galaxies.

5. Significance

Bridging the Gap: SAGUI provides a critical link between modern deep multi-band surveys and the physical interpretation of galaxy evolution. It allows astronomers to move beyond "integrated light" measurements to spatially resolved studies without the need for expensive spectroscopy for every pixel.
Physical Insight: By ensuring that segmented regions are spectrally homogeneous, the method reduces biases in derived physical parameters (e.g., preventing the "outshining" of old populations by young clusters within the same bin).
Future-Proofing: The framework is designed to be scalable and adaptable, making it suitable for the massive datasets expected from the LSST, Euclid, and Roman Space Telescope. It offers a pathway to study the internal physics of galaxies (assembly, merger history, secular evolution) with unprecedented spatial resolution in the photometric regime.

In conclusion, SAGUI represents a significant methodological advance in astronomical data analysis, shifting the paradigm from geometric binning to physics-driven segmentation, thereby enabling more robust and detailed studies of galaxy structure and evolution.