Exploring Compressive Sensing for Earth Observation

Abstract

This master thesis explores the application of compressive sensing in satellite Earth observation instruments. Firstly, a general state of the art of compressive sensing is made by introducing the mathematical concepts and describing some existing designs that implement the method. The essence of compressive sensing consists in reconstructing images with fewer measurements than in classical imaging. The method can bring drastic reduction of data quantity requirements and detector sizes as well as an increase of spatial resolution. These advantages are particularly interesting in Earth observation instruments considering the vast amount of data that they generate and the size limitations of satellites. This is even more considerable in the infrared spectrum where detectors are typically large. A deep learning compressive sensing reconstruction algorithm dubbed ISTA-Net+ is tested an proved to work on satellite multispectral data during simulations. Finally, a complete compressive sensing experimental chain has been implemented within laboratory environment. For the reconstruction, the hardware-compressed data could not be passed to the ISTA-Net+ algorithm, thus a simpler algorithm applying an inpainting using iterative hard thresholding is applied. The experiment is satisfactory and the method is proven to work. Nonetheless, the optical system has to be optimized and a more efficient algorithm must be implemented. Therefore, this work opens the way to further improvements and investigations.