MODEL&CO: Exoplanet detection in angular differential imaging by learning across multiple observations

Abstract

Direct imaging of exoplanets is particularly challenging due to the high contrast between the planet and the star luminosities, and their small angular separation. In addition to tailored instrumental facilities implementing adaptive optics and coronagraphy, post-processing methods combining several images recorded in pupil tracking mode are needed to attenuate the nuisances corrupting the signals of interest. Most of these post-processing methods build a model of the nuisances from the target observations themselves, resulting in strongly limited detection sensitivity at short angular separations due to the lack of angular diversity. To address this issue, we propose to build the nuisance model from an archive of multiple observations by leveraging supervised deep learning techniques. The proposed approach casts the detection problem as a reconstruction task and captures the structure of the nuisance from two complementary representations of the data. Unlike methods inspired by reference differential imaging, the proposed model is highly non-linear and does not resort to explicit image-to-image similarity measurements and subtractions. The proposed approach also encompasses statistical modeling of learnable spatial features. The latter is beneficial to improve both the detection sensitivity and the robustness against heterogeneous data. We apply the proposed algorithm to several datasets from the VLT/SPHERE instrument, and demonstrate a superior precision-recall trade-off compared to the PACO algorithm. Interestingly, the gain is especially important when the diversity induced by ADI is the most limited, thus supporting the ability of the proposed approach to learn information across multiple observations.

Figures (18)

• Figure 1: Overview of the proposed MODEL&CO algorithm. Left: illustration of the database $\mathcal{D}$ containing multiple speckles realizations $\boldsymbol{s}$ (shown in logarithmic scale), associated binary masks $\boldsymbol{m}$ identifying known real sources, and vectors $\phi$ of parallactic angles. Center: construction of the training samples from observations drawn from $\mathcal{D}$. For a selected data cube, data augmentation is first applied, then synthetic sources are injected in $\boldsymbol{s}$ through direct model (\ref{['eq:data_model']}). The synthetic signals are weighted by the variability of the speckles along their trajectories to form a target signal $\boldsymbol{x}$ to be reconstructed. Right: reconstruction of the target signal $\boldsymbol{x}$ by supervised learning from inputs $\boldsymbol{y}$ and $\phi$. The weights $\boldsymbol{\theta}$ of the learnable module $\mathbf{F}$ are optimized by minimizing MSE between $\widehat{\boldsymbol{x}}$ and $\boldsymbol{x}$. See text for details.
• Figure 2: Main statistics for the F150 database of observations of the VLT/SPHERE SHINE survey. (b): DB_H23 is for observations conducted in the H2-H3 dual band ($\lambda_0 = 1.59 \, \micro \meter$, $\lambda_1 = 1.67 \, \micro \meter$), DB_K12 stands for K1-K2 dual band ($\lambda_0 = 2.11 \, \micro \meter$, $\lambda_1 = 2.25 \, \micro \meter$), and BB_H is for observations in broadband H ($\lambda \in \left[ 1.48, 1.77 \right] \, \micro \meter$). (d): The classification procedure between bad, average, and good observing conditions based on seeing and coherence time $\tau_0$ is detailed in Sect. \ref{['subsec:obs_cond']}.
• Figure 3: Schematic representation of the masked temporal aggregation operator (MTA) applied for pre-conditioning of the training samples. This procedure corresponds to the grey box in the architecture of the proposed method given in Fig. \ref{['fig:architecture']}.
• Figure 4: Schematic representation of the architecture of MODEL&CO. The first processing stage of the proposed approach, working on speckles-aligned images, is depicted in the blue box. The sub-blocks associated with the whitening and normalization of the features through patch covariances are in orange (the central part illustrates the effect of the whitening on two arbitrary features $(f_i, f_j)$), sub-blocks containing learnable parameters (projections and 2-D spatial convolutions) are highlighted in green, and other operations (non-linear activation, patch extraction, normalization and aggregation are in grey). The output of the first stage is derotated and temporally aggregated by the MTA block. For illustration purposes, the rotation vector and the binary mask are omitted in this figure. A more detailed view of the MTA module is represented in Fig. \ref{['fig:mta']}. The object-aligned features are then filtered by a to produce the final reconstruction $\widehat{\boldsymbol{x}}$. For each stage, examples of some intermediate quantities and their associated shapes are given.
• Figure 5: Calibration mapping $c$ linking the natural output values $\widehat{\boldsymbol{x}}$ of MODEL&CO to a detection score interpretable as a S/N of detection. The experiments are conducted with different numbers $T_{\text{calib}}$ of frames within the calibration datasets in the absence of sources of interest.