🔭 astrophysics

Catalog-based detection of unrecognized blends in deep optical ground based imaging

This paper demonstrates that machine learning algorithms applied to catalog-level photometric and morphological data can effectively identify and remove a significant portion of unrecognized blends in deep ground-based imaging, thereby improving sample purity for future cosmological surveys like LSST.

Original authors: Shuang Liang, Prakruth Adari, Anja von der Linden, The LSST Dark Energy Science Collaboration

Published 2026-05-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shuang Liang, Prakruth Adari, Anja von der Linden, The LSST Dark Energy Science 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

The Big Problem: The "Cosmic Crowd"

Imagine you are looking at a massive, crowded concert from a very high balcony. From this distance, you can see thousands of people. But because you are so far away, and the air is a little hazy, some people standing right next to each other look like a single, blurry blob.

In astronomy, this is called a "blend." When telescopes on Earth look deep into space, they often see two or more galaxies that are so close together in the sky that their light merges into one. The telescope's software thinks, "Ah, that's just one big galaxy," and records it as a single object.

The paper calls these "unrecognized blends." They are the "ghosts" in the data—objects that look like one thing but are actually a messy pile-up of two or more. This is a big problem for cosmologists because if you think a galaxy is one thing, but it's actually two, your calculations about the universe's size, shape, and history will be wrong.

The Challenge: How to Spot the Blends?

Usually, to see if a galaxy is actually two, you need a super-powerful telescope in space (like the Hubble Space Telescope) that can see clearly through the "haze" of Earth's atmosphere. But we can't put a Hubble-sized telescope over every single patch of sky we want to study.

The authors asked a simple question: Can we figure out which galaxies are "blends" just by looking at a list of numbers (a catalog) from a ground-based telescope, without needing a space telescope for every single object?

They wanted to know if they could use Machine Learning (computer programs that learn from patterns) to spot these blends based on three simple clues:

Color: What color is the object?
Brightness: How bright is it?
Size: How big does it look?

The Experiment: Training the Detectives

To test their idea, the scientists used a special region of the sky called COSMOS.

The Ground View: They took data from a ground-based telescope (Subaru) that sees the sky a bit blurry.
The Truth: They compared this to data from the Hubble Space Telescope, which sees the same area clearly.

By comparing the two, they could see exactly which ground-based "blobs" were actually multiple galaxies. They found that about 17% of the objects in the ground-based list were actually unrecognized blends.

Then, they fed the "Ground View" data (colors, brightness, size) into four different Machine Learning algorithms to see which one could best guess which objects were blends, using the Hubble data as the answer key.

The Tools: Four Different Detectives

The team tested four different types of "detectives" (algorithms):

The Map Maker (Self-Organizing Map / SOM): Imagine a map where every spot represents a specific combination of galaxy colors and sizes. The computer learns what a "normal" galaxy looks like and maps them onto this grid. If a galaxy lands in a weird, empty spot on the map, or in a spot crowded with known "messy" blends, the computer flags it as suspicious.
The Decision Tree (Random Forest): Imagine a giant flowchart. "Is it red? Yes/No. Is it big? Yes/No." The computer builds thousands of these flowcharts and asks them all, "Is this a blend?" If most of the trees say "Yes," then it's a blend.
The Neighbor Finder (k-Nearest Neighbors): This looks at a galaxy and asks, "Who are your closest neighbors in the data?" If your neighbors are mostly known blends, you are probably a blend too.
The Outlier Hunter (Anomaly Detection): This looks for things that are weird or don't fit in with the crowd. It assumes blends are "weird" because they are a mix of two different things.

The Results: What Worked?

The scientists found some interesting things:

The Winners: The Random Forest and the Map Maker (SOM) were the best detectives. They could correctly identify about 30% to 80% of the hidden blends.
The Cost: To catch these blends, the computer had to be a bit aggressive. To catch 80% of the bad blends, it had to throw away about 50% of all the galaxies (including some good ones). It's like a bouncer at a club who wants to keep out the troublemakers but ends up kicking out half the innocent guests too.
The Size Clue: The most important clue wasn't just color; it was size. Blended galaxies tend to look bigger and "fuzzier" than single galaxies. If you ignore the size, the computer gets much worse at its job.
The Color Surprise: They thought they needed infrared (heat) light to find these blends, but they found that optical light (visible colors) was enough. You don't need the fancy extra sensors to do this job.
The "Outlier" Twist: They also tried to use these tools to find galaxies with the wrong "distance" (redshift). They found that the tools designed to find "weird colors" (Anomaly Hunters) were actually better at finding distance errors than the tools designed specifically to find blends. This suggests that galaxies with the wrong distance look even stranger than blended galaxies do.

The Bottom Line

This paper shows that we don't always need a space telescope to fix our ground-based data. By using smart computer programs and looking at simple things like size and color, we can clean up our lists of galaxies.

While we can't catch every blend, we can catch a significant chunk of them. This helps astronomers get a cleaner, more accurate picture of the universe, which is crucial for understanding things like dark energy and how the universe is expanding. It's like cleaning up a blurry photo by using a smart filter to remove the double-exposed spots, making the final picture much clearer.

Technical Summary: Catalog-Based Detection of Unrecognized Blends in Deep Optical Ground-Based Imaging

Problem Statement
Deep ground-based optical surveys, such as the Legacy Survey of Space and Time (LSST), face a critical challenge: "unrecognized blends." As source density increases, galaxies overlap in projection. While source detection algorithms can often deblend objects with distinct peaks, a significant fraction of overlapping sources are detected as single objects. These "unrecognized blends" (or ambiguous blends) are indistinguishable from isolated galaxies in ground-based imaging but can be identified via high-resolution space-based data (e.g., HST).

These blends pose severe risks to weak lensing cosmology. They corrupt two essential measurements:

Galaxy Shapes: Blending introduces multiplicative shear bias.
Photometric Redshifts (photo-z): Blended objects exhibit unphysical colors due to the mixture of components, leading to catastrophic outliers in redshift estimation.
Previous studies suggest that unrecognized blends constitute 14–30% of detections in deep surveys and can induce shifts in cosmological parameters (e.g., $S_8$ ) comparable to current tensions between cosmic shear and CMB results.

Methodology
The authors propose a catalog-level detection method using Machine Learning (ML) to identify unrecognized blends using only photometric colors, magnitudes, and size information, without requiring pixel-level image analysis.

Data: The study utilizes the COSMOS 2015 catalog (ground-based, 9-band photometry: $uBVri^+z^++YJH$ , plus size) matched to HST/ACS imaging (F814W filter) which serves as the "truth" for blends.
Ground-Truth Definition: A two-step matching algorithm is employed to distinguish "pure" sources from "unrecognized blends."
- Pure: A ground detection matches exactly one HST source.
- Unrecognized Blend: A ground detection matches one primary HST source and at least one additional HST source within 2 arcseconds and within 2 magnitudes ( $\Delta i \le 2$ ).
- The sample is divided into "weak blends" ( $1 \le \Delta i \le 2$ ) and "strong blends" ( $\Delta i < 1$ ).
Machine Learning Algorithms: Four classes of algorithms are tested:
1. Self-Organizing Maps (SOM): Used in two ways:
  - Blending Cell Approach: An unsupervised neural network maps objects to a grid. Cells with a high ratio of blends to pure objects (identified via a training set) are flagged as "blending cells."
  - Distance Cut Approach: Objects are flagged if their $\chi^2$ distance to their best-match cell (trained on pure galaxies) is large, indicating unphysical colors.
2. Random Forest (RF): A supervised classifier trained to distinguish pure galaxies from blends.
3. k-Nearest Neighbors (k-NN): A non-parametric supervised method.
4. Anomaly Detection: Unsupervised methods (LOF, Isolation Forest, etc.) trained on pure galaxies to find outliers.
Evaluation Metrics: Performance is measured by Recall (fraction of true blends identified) and Cost (fraction of the total sample removed/flagged as blends).

Key Results

Blend Fraction: In the ground-based COSMOS catalog ( $i^+ < 24.5$ ), approximately 17% of objects are identified as unrecognized blends.
Algorithm Performance:
- Random Forest (RF) and SOM (Blending Cell) perform best. They can identify 30% to 80% of unrecognized blends while rejecting 10% to 50% of all detected galaxies.
- k-NN performs similarly to SOM.
- Anomaly Detection: Most unsupervised anomaly detectors perform only marginally better than random selection. However, Local Outlier Factor (LOF) shows promise, particularly for identifying photo-z outliers.
- SOM Distance Cut: This method is less effective at finding blends than the Blending Cell approach but is highly effective at finding photo-z outliers.
Feature Importance:
- Size (flux radius) and $i$ -band magnitude are the most critical features.
- Near-Infrared (NIR) bands ($YJH$) have a negligible impact on performance; optical bands alone yield comparable results.
- Removing size information significantly degrades performance.
Photo-z Outliers: Algorithms specifically targeting "feature outliers" (SOM Distance Cut and LOF) outperform those targeting blends (RF, SOM Blending Cell) in identifying photo-z outliers. This suggests that photo-z outliers possess more extreme/exotic features than unrecognized blends.

Significance and Claims
The paper claims that its methods offer a practical mitigation strategy for unrecognized blends in future surveys like LSST, where space-based imaging (e.g., Euclid, Roman) will not be available for the entire survey area.

Sample Purity: The methods can produce a "cleaner" galaxy sample by removing a significant fraction of blends at a moderate cost to statistical power (sample size).
Synergy with LSST: Since SOM is already proposed for LSST photo-z estimation, the "blending cell" approach can be integrated with minimal extra effort. A small overlapping region with space-based data can train the SOM to identify "blending cells" for the entire survey.
Cosmological Impact: By reducing the contamination of unrecognized blends, the methods can improve the accuracy of cosmological parameter constraints (e.g., $S_8$ ) and reduce systematic uncertainties in weak lensing measurements.
Limitations: The authors note that while they can identify and remove blends, this introduces a selection bias (removing fainter, larger objects). They acknowledge that shape-based selection (e.g., ellipticity) is risky for weak lensing due to shear-dependent bias, though size information appears manageable.

The work does not claim to eliminate unrecognized blends entirely but demonstrates that catalog-level ML can effectively flag a substantial portion of them, thereby improving the purity of samples used for cosmological analysis.