Skip to main content
SearchLoginLogin or Signup

Improving exoplanet detectability by combining high contrast imaging, high resolution spectroscopy, and machine learning techniques

Presentation #102.421 in the session Poster Session.

Published onJun 20, 2022
Improving exoplanet detectability by combining high contrast imaging, high resolution spectroscopy, and machine learning techniques

Advances in observing methods, and the advent of high-contrast imaging spectrographs, allow to simultaneously produce high contrast images (HCI) and extract high resolution spectra (HRS) of exoplanet targets. This has produced multispectral images of the targets making it possible to simultaneously use the image and spectral dimensions of such data on extrasolar planetary systems such as HR8799 (Konopacky et al. 2013, Ruffio et al. 2019), PDS 70 (Haffert et al 2020., Christiaens et al. 2020), etc. Advanced data processing techniques have been proposed to improve the current detection limit taking advantage of the large feature set provided by multispectral imaging. Machine learning (ML) has had a particularly high success in the imaging domain (e.g Gomez Gonzalez et al 2018). However ML has proven ineffective when using spectra alone (e.g Fisher et al 2020) owing to the a large number of spectral channels that do not contribute discriminatory features in a star-dominated spectrum. Therefore, dimensionality reduction have been suggested in order to effectively harness HRS data. Consequently, this project investigates whether, after reducing the dimensionality of HRS, the spatial diversity provided by the HCI can be used to improve the detection limit.

We use SDI cubes from SINFONI. This consist of the HD142527 data cube and an empty data cube with injected companions. We implement dimensionality reduction by replacing the spectral dimension with a relative velocity dimension and the pixel values with cross correlation (CCF) values This produces a spatial CCF map, swhere pixels containing spectra closer to the template that they are correlated with have a higher value. However, this map is still contaminated by noise and field rotation.

It has been proven that the application of derotation and STIM algorithms with appropriate thresholding (Pairet et al 2019) to a standard ADI cube produces a reliable detection map. In our case we replace the ADI cube with a CCF cube and apply derotation+STIM. The choice of an appropriate threshold converts this STIM map into a detection map The thresholding in the STIM map has been shown to be somewhat noise dependent. In order to now harness the power of ML to our project we will replace the STIM+thresholding with an appropriate noise independent ML algorithm and summarize the improvement in detectability.

Comments
0
comment
No comments here