2024 workshop program
Modeling Image Systematics Using Conditional Diffusion Models
The work
Abstract
Scattered light from the Earth and Moon can significantly impact the background levels in TESS full frame images (FFIs), hindering the search for transiting exoplanets and other astronomical phenomena. While scattered light is often corrected at the light curve level, we present a novel approach to model and remove scattered light at the image level using deep learning. We have developed a conditional diffusion model that accurately captures the scattered light patterns in FFIs probabilistically, using only the angles and distances of the Earth and Moon with respect to the TESS cameras. The model learns the complex, dynamic patterns of scattered light and produces corrected FFIs along with uncertainties. By removing scattered light at the image level, we can enable improved photometry and planet searches in scattered light-affected regions of the FFIs. We demonstrate the performance of our model on all TESS sectors. This deep learning approach has the potential to revolutionize scattered light and systematic corrections in present and future space-based telescopes.