Abstract
Exoplanets are very dim. An Electron Multiplying Charged Coupled Device (EMCCD) operating in photon counting (PC) mode reduces the noise level and enables their detection. Typically, PC images are added together as a co-added image before processing. We present here a signal detection and estimation technique working directly with the PC images. The method is based on the generalized maximum likelihood ratio test (GLRT) and uses a binomial distribution between PC images. It can be applied online, so it is possible to stop observations once we have reached a threshold confidence for the existence (or absence) of planets, and thus the observation time is efficiently used. It directly uses the detector model, which more accurately represents its noise characteristics. We show that it outperforms our previous GLRT method using co-added images under Gaussian noise assumption.
1. Introduction: The Challenge of Exoplanet Imaging
Directly imaging an Earth-like exoplanet is an immense technical challenge. The flux ratio between a planet and its host star is approximately $10^{-10}$, comparable to spotting a firefly next to a powerful searchlight from miles away. Even with starshades or coronagraphs to suppress starlight, the remaining signal is often buried beneath the "read noise" of traditional cameras.
To overcome this, astronomers utilize Electron Multiplying CCDs (EMCCDs) in Photon Counting (PC) mode. In this regime, each pixel reports a binary value: "1" if a photon is detected and "0" otherwise. The conventional approach involves "co-adding" (averaging) thousands of these binary frames into a single image, then applying detection algorithms. However, this averaging assumes Gaussian noise, which is mathematically inaccurate at the extremely low light levels encountered in exoplanet research. This mismatch leads to a loss of valuable statistical information.
2. Methodology: Sequential GLRT and Binomial Modeling
The researchers propose a Sequential Generalized Likelihood Ratio Test (GLRT) that processes raw, binary PC images directly, preserving the underlying statistical structure.
2.1 Stochastic Modeling
Instead of the Gaussian approximation, the method employs a detailed physical model of the EMCCD. Each pixel in a single frame is treated as a Bernoulli trial. Over a sequence of $N$ images, the total count for a pixel follows a Binomial distribution. This model more faithfully represents the detector's physics at low photon rates.
2.2 Estimation and Sequential Testing
The algorithm uses Maximum Likelihood Estimation (MLE) via gradient descent to simultaneously estimate the planet's intensity ($\alpha$) and the background noise ($\beta$). Crucially, the test is "online": it updates its confidence level after every new frame. If the likelihood of a planet's existence reaches a predefined upper threshold, the observation stops successfully. Conversely, if it hits a lower threshold, the search concludes no planet is present, allowing the telescope to move to the next target immediately.
3. Simulation and Experimental Setup
The study validated this framework using high-fidelity simulations based on the Nancy Grace Roman Space Telescope (formerly WFIRST) specifications. The researchers modeled a solar system viewed from 10 parsecs, with Venus and Earth as the primary targets. The simulations included realistic noise sources such as dark current and clock-induced charge across thousands of 1-second integration frames.
4. Key Results and Performance
The Binomial GLRT demonstrated significant advantages over traditional methods:
- Enhanced Sensitivity: Receiver Operating Characteristic (ROC) curves showed that the Binomial method consistently outperformed Gaussian-based co-adding, particularly when data was sparse (e.g., only 20 frames).
- Precision in Estimation: The algorithm accurately estimated simulated planet intensities within a 5% error margin after several thousand frames.
- Improved Signal Clarity: Log-likelihood ratio maps produced sharper, more distinct peaks for targets like Venus and Earth compared to standard co-added images, facilitating easier signal separation from background noise.
5. Conclusion and Future Directions
By avoiding premature averaging, the Sequential GLRT extracts maximum information from every detected photon. This efficiency is vital for space-based missions where telescope time is a premium resource.
While the method is computationally more intensive than processing a single co-added image, it is well within the capabilities of modern hardware. Future work will focus on increasing robustness to detector parameter drift (e.g., gain or dark current changes) and uncertainties in the Point Spread Function (PSF) caused by telescope jitter.
Summary updated