Autoencoder for On-Board InSAR Phase Denoising and Compression

Abstract

In the last decades, SAR missions have been proposed in the framework of interplanetary exploration, featuring the on-board generation of higher-level products, such as SAR interferograms. However, the limited down-link capacity of the satellites has posed significant challenges, thus motivating the development of efficient on-board compression strategies. In this work, we present a novel deep-learning approach, through the implementation of a Convolutional AutoEncoder (CAE), which allows for the joint denoising and compression of the interferometric (InSAR) phase. The proposed network is trained and tested using synthetic datasets, derived starting from real TanDEM-X observations and assuming corresponding InSAR acquisition geometries and underlying topography. Results are assessed against a combination of boxcar filtering and JPEG 2000 compression, which reflects one of the possible strategies reported in the literature. In particular, we focus on three different performance metrics, i.e. denoising capability, data volume reduction and preservation of high-resolution details, thus showing the enhanced flexibility of the proposed methodology with respect to a state-of-the-art baseline method.