Spectral unmixing for exoplanet direct detection in hyperspectral data

The direct detection of exoplanets with high-contrast instruments can be boosted high spectral resolution. For integral field spectrographs yielding hyperspectral data, this means that the view consists diffracted starlight spectra and a spatially localized planet. Analysis usually relies on cross-correlation theoretical spectra. In purely blind-search context, supervised strategy biased model mismatch and/or computationally inefficient. Using an approach is inspired by remote-sensing community, we aim to propose alternative fully data-driven, which decomposes data into set individual their corresponding spatial distributions. This called unmixing. We used orthogonal subspace projection identify most distinct in view. Their distribution maps were then obtained inverting data. These break original images via non-negative least squares. performance our method was evaluated compared using simulated medium resolution from ELT/HARMONI spectrograph. show unmixing effectively leads planet solely based dissimilarities at significantly reduced computational cost. extracted spectrum holds significant signatures while being not perfectly separated residual starlight. sensitivity three four times higher than unsupervised unmixing, gap toward former because injected correlated match perfectly. algorithm furthermore vetted real VLT/SINFONI beta Pictoris system.