🔭 astrophysics

Photometric classification of quasars from DES and photo- $z$ estimation with Machine Learning

This paper presents a machine learning-based study that utilizes Dark Energy Survey DR2 data cross-matched with SDSS DR16 to achieve high-precision quasar classification and robust photometric redshift estimation, resulting in a reliable catalog of over 675,000 objects suitable for large-scale structure and cosmological investigations up to redshifts of $z \approx 4$ .

Original authors: Pablo Motta, Filipe B. Abdalla, Elcio Abdalla, Gabriel S. Costa, Camila Cardoso

Published 2026-05-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Pablo Motta, Filipe B. Abdalla, Elcio Abdalla, Gabriel S. Costa, Camila Cardoso

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 giant, crowded party. In this party, there are three main types of guests: Stars (like bright, steady lighthouses), Galaxies (like sprawling, fuzzy cities), and Quasars (like incredibly bright, distant spotlights that look exactly like stars from far away).

The problem for astronomers is that from Earth, the "spotlights" (quasars) and the "lighthouses" (stars) look almost identical. They are both just tiny dots of light. If you try to study the structure of the party (the universe) but accidentally count the lighthouses as spotlights, your map of the party will be wrong.

This paper is about building a better guest list for the Dark Energy Survey (DES), a massive telescope project that took pictures of the southern sky. The authors used Machine Learning (computer programs that learn from examples) to solve two big problems:

Sorting the guests: Figuring out which dots are real quasars and which are just stars pretending to be quasars.
Measuring the distance: Guessing how far away each quasar is, just by looking at its color, without needing to take a slow, expensive "spectral fingerprint" of every single one.

Here is how they did it, explained simply:

1. The Training Camp (The Data)

To teach the computer, the authors needed a "ground truth" list. They took a huge list of objects from the DES (the camera) and matched them with a list from the Sloan Digital Sky Survey (SDSS).

Think of SDSS as a group of experts who had already walked up to the guests and asked, "What are you?" and "How far away are you?"
They matched about 168,000 objects where both surveys agreed. This became the "training school" for their computer models.

2. The Sorting Machine (Classification)

The first job was to separate the "real" quasars from the "impostor" stars.

The Tool: They used a method called K-Nearest Neighbors (KNN). Imagine you are at a party and you see a new person. You look at the 11 people standing closest to them. If 10 of those neighbors are "Quasars," you assume the new person is a Quasar too.
The Trick: They realized that just looking at the shape of the dot wasn't enough (because they already filtered out fuzzy galaxies). Instead, they looked at the colors of the light in different filters (like looking at the guest through red, green, and blue glasses).
The Result: They tuned the computer to be very strict. They said, "We don't want any stars in our Quasar list." They set the computer to be 99% sure before calling something a Quasar.
- The Trade-off: Because they were so strict, they missed about 23% of the real quasars (they only caught 77%). But the ones they caught were almost certainly real quasars, not stars. This is crucial because a few fake stars can ruin the whole study.

3. The Distance Guessers (Redshift Estimation)

Once they had a clean list of quasars, they needed to know how far away they were. In astronomy, "redshift" is a measure of distance (and time).

The Challenge: Usually, you need a detailed spectrum (a rainbow of light) to know the exact distance. But the DES only has broad colors.
The Solution: They built a Hybrid Machine Learning team.
- One team member was a Boosted Decision Tree (a smart tree that asks a series of "Yes/No" questions about the colors).
- The other was a Decision Tree Regressor (another type of smart tree).
- They combined these two to guess the distance.
The "Outlier" Problem: Sometimes, the computer gets confused and guesses a distance that is wildly wrong (like thinking a guest is in the next room when they are actually on the moon). These are called "catastrophic outliers."
The Fix: They built a second layer of security, a "Stacked Outlier Classifier." Think of this as a bouncer who checks the computer's guess. If the computer says, "This quasar is 1 billion light-years away," but the colors look weird, the bouncer says, "Nope, that's a mistake," and throws the guess out.

4. The Final Guest List (The Catalog)

After all the sorting and distance-guessing, they created a massive catalog of 872,373 potential quasars.

The "Clean" Version: After the bouncer (outlier classifier) kicked out the bad guesses, they were left with 675,683 highly reliable quasars. This list is safe to use for studying the universe between redshift 0.5 and 3 (a vast range of cosmic history).
The High-Redshift Bonus: They found a special group of quasars that are very far away (around redshift 4). Interestingly, the computer didn't make many mistakes with these distant ones, so even the "messier" list is safe to use for this specific group.

Why Does This Matter?

The authors say this catalog is like a high-quality map for future explorers.

Because they successfully separated the "stars" from the "quasars," scientists can now use these quasars to measure the large-scale structure of the universe (how matter is spread out).
They can also study the intergalactic medium (the gas between galaxies) by looking at how the light from these distant quasars gets absorbed.
This work complements a previous study that mapped galaxies. Now, scientists have a map for the stars/quasars too. When you combine the two maps, you get a much clearer picture of the universe's expansion and the mysterious "Dark Energy" driving it.

In summary: The authors taught a computer to spot the difference between a star and a distant quasar using color clues, and then taught it to guess their distance. They added a "safety check" to remove the computer's worst mistakes. The result is a massive, reliable list of cosmic lighthouses that helps us understand how the universe is built.

Technical Summary: Photometric Classification of Quasars from DES and Photo-z Estimation with Machine Learning

Problem Statement
Quasars serve as critical probes for studying cosmic history, large-scale structure, and the intergalactic medium. However, their photometric identification is complicated by their morphological similarity to stars (both appear as point sources) and color overlaps with specific stellar types (e.g., hot blue stars or cool red dwarfs). Traditional selection methods, such as fixed color cuts or morphological analysis, often fail to resolve these ambiguities, particularly at high redshifts where UV features shift into optical bands. Misclassification introduces systematic errors that propagate into photometric redshift (photo-z) estimates, compromising cosmological measurements like Baryon Acoustic Oscillations (BAO). Furthermore, while existing catalogs often focus on extended sources (galaxies), point-like objects are frequently excluded, leaving a gap in the characterization of quasar populations for cosmological analyses.

Methodology
The authors constructed a comprehensive pipeline to classify point-like objects and estimate their photometric redshifts using data from the Dark Energy Survey (DES) and the Sloan Digital Sky Survey (SDSS).

Data Preparation:
- Photometry: DES Data Release 2 (DR2) provided photometric data for point-like sources in the $g, r, i, z, Y$ bands.
- Spectroscopy: SDSS Data Release 16 (DR16) provided spectroscopic classifications (Quasar, Star, Galaxy) and redshifts.
- Cross-Match: A positional cross-match (tolerance of 1 arcsecond) between DES DR2 and SDSS DR16 yielded an initial sample of 168,738 objects. This sample was split into 75% for training and 25% for testing classification models.
Quasar Classification (Star vs. Quasar/Galaxy):
- Algorithm: A K-Nearest Neighbors (KNN) classifier was employed.
- Feature Selection: The authors tested various features (magnitudes, fluxes, spread model parameters). They determined that PSF magnitudes in the $g, r, i, z$ bands provided the optimal separation between stars and non-stars (quasars/galaxies). Morphological spread parameters were found less effective because the initial DES selection had already filtered for point-like sources.
- Optimization: The model was tuned to maximize recall while maintaining a strict precision threshold of 0.99 to minimize stellar contamination. The optimal hyperparameter was found to be $k=11$ neighbors with a probability threshold of 0.8.
Photometric Redshift Estimation:
- Hybrid Approach: A hybrid machine learning model was developed combining a Boosted Decision Tree (BDT) from the ANNz framework and a Decision Tree Regressor (DTR) from scikit-learn.
- Features: The regression utilized color indices ( $g-r, r-i, i-z, z-y$ ), dereddened $r$ -band magnitude, and the Kron radius.
- Outlier Mitigation: To address catastrophic outliers (defined as $|z_{photo} - z_{spec}| \geq 0.4$ ), a stacked ensemble classifier was developed. This ensemble combined k-Nearest Neighbors (kNN), Gradient Boosting (GB), and Random Forest (RF) classifiers to identify and filter outliers.
Catalog Construction:
- The pipeline was applied to the full DES DR2 point-like sample (~50 million objects).
- Selection cuts were applied to restrict the data to the feature space covered by the training set (magnitudes between specific percentiles and Kron radius constraints).
- Masks were applied to exclude regions around bright local galaxies and dwarf systems to prevent contamination.

Key Results

Classification Performance: The optimized KNN model achieved a recall of 0.77 at a precision of 0.99. The Area Under the Precision-Recall Curve (AUC-PR) was 0.975.
Redshift Estimation: The hybrid BDT+DTR model produced a main cluster of objects with small errors in the range $0.5 < z < 3$ . A distinct population was successfully recovered at $z \approx 4$ .
Outlier Reduction: The stacked outlier classifier significantly reduced the fraction of catastrophic outliers. The final classification threshold was set to achieve a precision of 0.9 and recall of 0.73 for outlier identification.
Catalog Statistics:
- Full Sample: The initial photometric quasar catalog contains 872,373 objects.
- Cleaned Sample: After outlier removal, the catalog contains 675,683 objects.
- Redshift Range: The catalog spans $z \approx 0.5$ to $z > 3$ , with a peak at $z \sim 1.5$ .
- High-Redshift Subset: A specific subset of 10,756 objects with $z \geq 3.5$ was identified. The authors note that for this high-redshift range, the larger catalog (without outlier removal) remains reliable because catastrophic outliers do not extend beyond $z \approx 3.5$ .

Significance and Claims
The paper claims to provide a robust, large-scale catalog of point-like objects (quasars/galaxies) derived from DES DR2, specifically designed to complement existing galaxy catalogs (such as those in reference [58]) which focus on extended sources.

Complementarity: The work addresses a gap in cosmological datasets by providing a dedicated resource for point-like sources, enabling joint analyses with galaxy catalogs to improve constraints on dark energy, structure growth, and primordial non-Gaussianity.
Reliability: The authors assert that the cleaned catalog (675,683 objects) is suitable for large-scale structure studies in the range $0 < z < 3$ .
High-Redshift Utility: The catalog is highlighted as a valuable resource for probing the early universe ( $z \approx 4$ ), a regime where the high-redshift population is reliably identified despite the small sample size.
Methodological Contribution: The study demonstrates that a hybrid machine learning approach, combined with a stacked outlier classifier, effectively mitigates catastrophic redshift errors while maintaining high sample completeness, offering a template for future photometric surveys like LSST and Euclid.

The authors position this work as a foundational step for future cosmological investigations, particularly in synergy with radio surveys (e.g., BINGO, SKA) that utilize cross-correlations with optical catalogs to map the large-scale structure of the universe.

1. The Training Camp (The Data)

2. The Sorting Machine (Classification)

3. The Distance Guessers (Redshift Estimation)

4. The Final Guest List (The Catalog)

Why Does This Matter?

More like this