Exoplanet detection in direct imaging by statistical learning of the non-stationary patch covariances

Seminar at The European Southern Observatory (Santiago, Chile), The Observatory of Geneva (Geneva, Switzerland) and The Observatory of Lyon (Saint-Genis-Laval, France)


We have recently introduced a new method dedicated to source detection from angular differential imaging (ADI) data: PACO (for PAtch COvariances). Data reduction in ADI is challenging because the faint point sources are hidden in a stronger nonstationary background (speckles) displaying strong spatial correlations. PACO learns locally a statistical model of the background directly from the data. This model captures short-scale spatial correlations up to a separation of ten pixels (i.e., within an image patch). The decision in favor of the presence or the absence of an exoplanet is then performed by a binary hypothesis test.
PACO offers appealing characteristics compared to existing detection approaches. Since no image subtraction is performed, the photometry is preserved. PACO is completely parameter-free, including the computation of a detection map, its thresholding to extract meaningful detections, and the estimation of fluxes of the detected sources. Finally, the resulting detection maps are stationary and statistically well-modeled so that the false alarm rate, the probability of detection, and the contrast can be directly assessed without post-processing and/or Monte-Carlo simulations. We have shown using datasets from the VLT/SPHERE-IRDIS instrument that the proposed method achieves significantly better detection performance than current cutting-edge algorithms such as TLOCI or KLIP. We believe that these significant practical advantages should make PACO a method of choice for the analysis of ADI observations, in particular for large exoplanet surveys.
We have very recently extended this algorithm to the joint processing of the different spectral channels of angular and spectral differential imaging (ASDI) data. The resulting algorithm, named PACO-ASDI, accounts for the spatio-temporo-spectral fluctuations of the data. Our tests conducted on several ASDI datasets from the VLT/SPHERE-IFS instrument show that PACO-ASDI also produces reliable detection maps and unbiased spectral energy distribution of the detected sources (with confidence intervals), outperforming the state-of-the-art exoplanet hunter algorithms.