Nils FOIX-COLONIER – Deep learning-based compression and despeckling of SAR images

Abstract: The usefulness of spaceborne synthetic aperture radar (SAR) systems, with its wide range of valuable applications, is well established. It can be used anytime and under any weather to monitor forest evolution, to secure maritime areas, to measure geophysical parameters remotely, etc. This is why data downlink and its interpretability are key issues. The aim of the present work is to maximize on-board compression to increase the amount of data transmitted and reduce the inherent costs in using such infrastructures. It should be noted that an unavoidable multiplicative noise, speckle, heavily disrupts SAR images. Since speckle can be mostly considered as a high-frequency unwanted information, it seems natural to introduce on-board denoising of SAR images before compression and transmission. Doing so will improve compression without recurring to the usual trade-off of sacrificing image quality, but rather by estimating the underlying SAR reflectivity before downlink. Therefore, combining despeckling and compression tasks is advantageous since decreasing the amount of information to encode will result in a more efficient data downlink. The work presented introduces a self-supervised solution to perform joint compression and despeckling of SAR images, with an estimation of the reflectivity based on an original adaptation of the latest machine learning-based advances in the fields of image compression and SAR images despeckling with artificial neural networks. The proposed solution was successfully implemented and tested on real-world data from TerraSAR-X, showing great potential for achieving state-of-the-art despeckling under the constraints of end-to-end optimized compression with variational autoencoders.