🔭 astrophysics

Coherence differential imaging using gradient-boosted decision trees for the direct detection of exoplanets

This paper introduces EPICX, an enhanced Coherence Differential Imaging method that utilizes gradient-boosted decision trees to model the differential signal as a high-dimensional regression problem, thereby improving the direct detection of exoplanets by effectively distinguishing incoherent planet signals from coherent speckle noise in coronagraphic imaging.

Original authors: Johan Mazoyer, María J. Mellado-Tenorioa, Axel Potier, Christian Wilkinson, Lukas Delaye, Raphaël Galicher, Iva Laginja

Published 2026-07-20

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Johan Mazoyer, María J. Mellado-Tenorioa, Axel Potier, Christian Wilkinson, Lukas Delaye, Raphaël Galicher, Iva Laginja

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 night sky as a giant, glittering stage. On this stage, the stars are the main actors, shining so brightly that they drown out the tiny, faint whispers of the planets orbiting them. Trying to see an exoplanet—a world circling another star—is like trying to spot a firefly hovering next to a blinding spotlight from miles away. The light from the star doesn't just shine; it scatters, creating a fuzzy, shimmering haze called "speckles." To the naked eye, or even a standard camera, these speckles look exactly like tiny planets. They are the ultimate cosmic impostors.

For decades, astronomers have tried to filter out this noise. Some wait for the Earth to rotate, hoping the star's glare shifts while the planet stays put. Others use reference stars to subtract the noise, but these methods are slow and often leave behind a messy residue. A newer, cleverer trick involves "Coherence Differential Imaging" (CDI). This technique plays a game of "spot the difference" by wiggling the telescope's mirror slightly. Because the star's light is coherent (like a laser beam) and the planet's light is incoherent (like a flashlight), they react differently to the wiggle. By comparing the before-and-after pictures, scientists can theoretically cancel out the star's glare and reveal the planet. However, this method usually relies on complex math models that can get things wrong if the telescope isn't perfect.

This is where a new study steps in, introducing a digital detective named EPICX. Instead of relying on a rigid mathematical formula to guess which speckle is which, the researchers trained a machine learning algorithm to learn the difference on its own. Think of it as teaching a computer to recognize the "personality" of a speckle versus a planet by showing it thousands of examples. The paper, titled "Coherence differential imaging using gradient-boosted decision trees for the direct detection of exoplanets," presents a method that uses these smart decision trees to separate the real planets from the fake ones much faster and more accurately than before.

The Digital Detective and the Cosmic Masquerade

The core problem the authors tackle is the "cosmic masquerade." In the images taken by powerful telescopes like the VLT/SPHERE or the upcoming Roman Space Telescope, the star's leftover light (speckles) looks exactly like a planet. Traditional methods try to subtract the star's light using a pre-made map of what the star should look like. But if the telescope has even a tiny, unknown flaw, that map is wrong, and the subtraction fails.

The authors propose a different approach using EPICX (Exoplanet Patch-based Image Classification with XGBoost). Instead of trying to build a perfect model of the star, they treat the problem like a game of "Where's Waldo?" but with a super-smart AI.

Here is how the process works, step-by-step:

The Wiggle (Pair Wise Probing): The telescope's mirror is nudged in specific patterns (called "probes"). This is like wiggling a flashlight to see how the shadows move. The star's light (the speckles) moves and changes brightness because it is coherent. The planet's light, however, is incoherent and stays stubbornly still.
The Snapshot: The telescope takes a series of pictures: one without the wiggle, and several with different wiggles. This creates a "PWP scene" (Pair Wise Probing scene).
The Hunt: The computer scans these images to find every tiny dot that looks like a point of light. It doesn't care if it's a planet or a speckle yet; it just marks them all as "suspects."
The Interrogation (Machine Learning): This is where EPICX shines. The algorithm looks at each suspect and asks a series of questions based on how that dot reacted to the wiggles.
- Did it flicker? (Speckles flicker; planets don't).
- How did it change when the mirror moved?
- What does its neighborhood look like?
  The algorithm uses Gradient-Boosted Decision Trees, which is a fancy way of saying it builds a team of many small, simple decision-makers. Each one makes a guess, and if it's wrong, the next one tries to fix the mistake. Together, they form a super-accurate judge.
The Verdict: The computer assigns a probability score to every suspect. If the score is high enough, it's a planet. If it's low, it's just a speckle.

What the Simulations Showed

The researchers didn't just dream this up; they tested it in a virtual universe. They used a Python tool called Asterix to simulate high-contrast images for three different types of telescopes: a "Perfect" telescope, one with a FQPM coronagraph, and the Roman HLC (Hybrid Lyot Coronagraph) designed for the upcoming Roman Space Telescope.

They created 2,250 simulated scenes. In each scene, they hid 0, 1, or 2 fake planets among hundreds of fake speckles. They then fed this data to the EPICX algorithm to see if it could find the planets.

The results were promising:

Spotting the Fakes: The algorithm was incredibly good at ignoring the speckles. In the simulations, the "False Positive Rate" (mistaking a speckle for a planet) was kept very low. This is crucial because, in real life, astronomers have to spend expensive telescope time re-checking every "planet" they find. They don't want to waste time on fake ones.
The Trade-off: The algorithm did miss some planets. In the test set, about 20% of the injected planets were missed (False Negatives). The authors note that this is partly because they didn't include the "grainy" noise of real photons in these initial simulations, and they injected planets at the same brightness as the speckles, making them very hard to spot.
The "Big Wiggle" Surprise: One of the most interesting findings was about how much they wiggled the mirror. Traditional methods require tiny, precise wiggles (about 36 nm or λ/16) to keep the math simple and linear. However, EPICX worked just as well even when they wiggled the mirror much harder (192 nm or λ/3). This suggests the AI can handle "non-linear" chaos that would confuse older, math-heavy methods. This could mean faster observations in the future, as bigger wiggles might reveal the truth more quickly.
Fewer Photos: They also tested using only 2 probes instead of the standard 3. The performance dropped slightly, but the authors suggest this could still be useful if you only need to check a small part of the sky, saving valuable telescope time.

The Bottom Line

The paper concludes that EPICX is a powerful new tool for the future of exoplanet hunting. By using machine learning to analyze how light reacts to mirror wiggles, it can separate real planets from star-glow noise without needing a perfect, error-free model of the telescope.

The authors are careful to state that these results are based on simulations. They haven't yet tested this on a real telescope or with real starlight. The next steps involve adding realistic noise (like photon noise) to the simulations and testing the algorithm on a physical testbed called THD2. If it holds up, this method could revolutionize how we look for new worlds, allowing telescopes like Roman and SPHERE+ to find planets faster and with fewer false alarms than ever before.

In short, while the universe is full of cosmic tricksters, this new digital detective is learning to see through the disguise.

Technical Summary: Coherence Differential Imaging using Gradient-Boosted Decision Trees for the Direct Detection of Exoplanets

Problem Statement
Direct imaging of exoplanets is fundamentally limited by residual stellar speckles, which mimic planetary signals and obscure faint astrophysical targets. While advanced post-processing techniques like Angular Differential Imaging (ADI) and Reference Differential Imaging (RDI) are widely used, they suffer from significant drawbacks: ADI requires long observation sequences and struggles at small separations, while RDI is time-consuming and sensitive to speckle evolution during target switching. Coherence Differential Imaging (CDI) offers a faster alternative by exploiting the incoherence between planetary light and coherent speckles. However, traditional CDI methods rely heavily on accurate, linearized instrumental models (e.g., Pair Wise Probing or PWP) to reconstruct and subtract the speckle field. These model-based approaches are limited by non-linearities and unknown errors in the optical system, restricting their performance and detection capabilities, particularly close to the star.

Methodology
The authors introduce EPICX (Exoplanet Patch-based Image Classification with XGBoost), a machine learning-enhanced CDI method that replaces the linear model-based reconstruction of speckles with a high-dimensional regression and classification approach.

Data Generation: The study utilizes the Asterix simulation package to generate a dataset of PWP sequences. These sequences include unprobed focal-plane images and three pairs of probe images (push/pull) generated by applying phase perturbations on a deformable mirror. Simulations cover three coronagraph types: a Perfect coronagraph, a Four Quadrant Phase Mask (FQPM), and the Roman Space Telescope's Hybrid Lyot Coronagraph (HLC). The dataset includes 2,250 scenes with injected planets (0, 1, or 2) and associated ground truth maps.
Pipeline Architecture:
1. Source Identification: An iterative DAOStarFinder algorithm identifies all PSF-shaped structures (both speckles and planets) in the unprobed image.
2. Feature Engineering: Instead of using raw pixel intensities, the algorithm extracts physically meaningful features from the entire scene (unprobed + 6 probed images). A sliding window computes local statistics, emphasizing temporal fluctuations caused by probe modulation. Coherent speckles fluctuate significantly under probing, while incoherent planet signals remain stable. An Iterative Feature Elimination (IFE) process reduced the initial feature set to 12 critical features.
3. Classification: The extracted features are fed into an XGBoost (Extreme Gradient Boosting) model. The algorithm treats the problem as a binary classification task (Planet vs. Speckle) for each detected patch.
4. Post-processing: Detections are consolidated by merging overlapping patches based on probability scores, prioritizing the minimization of False Positives (FPs) to ensure only high-confidence candidates are flagged for follow-up.
Training Strategy: The model is trained on 1,200 scenes, validated on 300, and tested on a strictly independent set of 750 scenes. The decision threshold is dynamically tuned to maximize the $F_{0.5}$ score, which weights precision (low FP rate) twice as heavily as recall, reflecting the high cost of false positives in astronomical follow-up.

Key Contributions

Model-Free CDI: EPICX decouples speckle discrimination from the need for a precise, linearized instrumental model. By learning the modulation patterns directly from data, it avoids the limitations imposed by non-linearities and model errors that constrain classical PWP-CDI.
Gradient-Boosted Decision Trees: The application of XGBoost to high-contrast imaging post-processing allows for robust handling of the extreme class imbalance (speckles vastly outnumber planets) and effective discrimination based on coherence properties.
Robustness to Large Probes: The study demonstrates that the algorithm remains effective even when probes are pushed to large amplitudes ( $\lambda/3$ ), well beyond the linear regime ( $\lambda/16$ ) required by traditional PWP. This suggests potential for faster speckle calibration with shorter exposure times.

Results

Performance Metrics: Confusion matrices across Perfect, FQPM, and Roman HLC coronagraphs show that EPICX successfully minimizes False Positives, achieving the primary goal of identifying only the most likely planetary candidates.
Detection Rates: While the algorithm achieves high precision, it misses approximately 20% of injected planets (False Negatives) in the initial tests. The authors note this is partly due to the lack of photon noise in the simulations and the broad injection levels relative to speckle brightness.
Probe Configuration: The method performs well with 3 probes. A reduction to 2 probes results in a slight performance degradation, consistent with the reduced modulation of certain speckle regions.
Non-Linearity: The algorithm successfully classifies speckles and planets even with large probe amplitudes ( $\lambda/3$ ), where the linear PWP model would fail, indicating the method's ability to capture non-linear modulation effects.

Significance and Claims
The paper claims that EPICX represents a significant step toward faster and more robust speckle calibration for current and future high-contrast instruments (e.g., VLT/SPHERE+, Roman/CGI). By leveraging machine learning to model the differential signal, the method offers a faster alternative to ADI and RDI, potentially reducing the need for long observation sequences. The authors emphasize that EPICX has the potential to minimize model errors and escape the linear constraints of classical CDI.

The authors maintain a modest tone regarding future implications, stating that the current results are a proof-of-concept using simulated data without photon noise. They outline the next steps as developing more complex simulation models (incorporating noise and realistic instrument simulators) and validating the method on the THD2 experimental testbed before proceeding to on-sky validation. The work does not claim immediate superiority over all existing methods but highlights its potential to improve detection efficiency and reduce reliance on complex physical models.

The Digital Detective and the Cosmic Masquerade

What the Simulations Showed

The Bottom Line

More like this