A New Statistical Model of Star Speckles for Learning to Detect and Characterize Exoplanets in Direct Imaging Observations

TL;DR

This work tackles the challenge of detecting and characterizing exoplanets in direct imaging by modeling the nuisance speckle field with a multi-scale, convolutional statistical framework that exploits rotational and spectral symmetries. It couples this interpretable statistical model with an end-to-end learnable prior, enabling simultaneous detection and flux/position estimation while providing uncertainty quantification. The method demonstrates improved detection performance, especially in ASDI data, and robust flux/position estimates across varied data quality, validated on SPHERE/VLT datasets and synthetic injections, with calibration to control false-alarm rates. The approach is computationally efficient, scalable to large surveys, and extensible to higher spectral resolution and extended structures like circumstellar disks.

Abstract

The search for exoplanets is an active field in astronomy, with direct imaging as one of the most challenging methods due to faint exoplanet signals buried within stronger residual starlight. Successful detection requires advanced image processing to separate the exoplanet signal from this nuisance component. This paper presents a novel statistical model that captures nuisance fluctuations using a multi-scale approach, leveraging problem symmetries and a joint spectral channel representation grounded in physical principles. Our model integrates into an interpretable, end-to-end learnable framework for simultaneous exoplanet detection and flux estimation. The proposed algorithm is evaluated against the state of the art using datasets from the SPHERE instrument operating at the Very Large Telescope (VLT). It significantly improves the precision-recall trade-off, notably on challenging datasets that are otherwise unusable by astronomers. The proposed approach is computationally efficient, robust to varying data quality, and well suited for large-scale observational surveys.

Figures (11)

• Figure A: Left: typical observations $\boldsymbol{y}$ and PSF $\boldsymbol{h}$ from the SPHERE instrument in ASDI mode. The synthetic exoplanet is very bright for illustration purposes. Right: temporal slice along the vertical line.
• Figure B: Workflow of the proposed method: it exploits both the spectral behavior of speckles and the apparent motion of exoplanets to disentangle the exoplanet signal from the nuisance component in the observations $\boldsymbol{y}$. To achieve this, local patches of the nuisance are modeled as Gaussian distributions, leveraging problem symmetries and incorporating multiple scales. These patches are fed to our convolutional statistical model, and combined to form a detection map. Additionally, a learned object prior, represented by a UNet $f_\nu$, is introduced to denoise this detection map produced by the statistical model. This approach results in an end-to-end learnable architecture.
• Figure C: Proposed convolutional statistical model: spectrally aligned speckles patches, indexed by $j$, with dimensions $Np$ and $CT$ samples, are first linearly projected into a lower-dimensional space of size $m$. In this space, the parameters of the Gaussian distribution $\widehat{\boldsymbol{m}}_j$ and $\widehat{\mathbf{C}}_j$ are estimated and subsequently combined with the PSF $\boldsymbol{h}$ to compute the terms $\boldsymbol{a}^{(j)}$ and $\boldsymbol{b}^{(j)}$. As detailed in Appendix \ref{['sec:conv_ab']}, the efficient computation of $\boldsymbol{a}^{(j)}$ relies on the Cholesky decomposition of the precision matrix $\widehat{\mathbf{C}}^{-1}_j$.
• Figure D: Detection maps on observations of HD 159911 star with synthetic exoplanets. The (calibrated) detection threshold is equivalent for all methods. The proposed approach here detect 1 additional source compared to the second best method (PACO ASDI).
• Figure E: Detection maps on 3 observations (stacked in false RGB colors) of HR 8799 star. The elliptical arcs depict the estimated (projected) orbits of three known exoplanets, with the detection results shown as red, green and blue dots for the corresponding 2016, 2018 and 2021 observations. Squares are for false alarms identified in bodrito2024model, Fig. 17.