Optimizing PHI-NET for High-Resolution InSAR Parameters Estimation

Abstract

Φ-Net is a novel deep learning architecture developed for the estimation of interferometric SAR parameters, i.e. the interferometric phase and coherence. The network, trained on a data set of synthetic images only, has shown outstanding denoising capabilities, surpassing state-of-the-art algorithms when tested on both synthetic and real InSAR data. We aim to further improve its performance by investigating new loss functions and optimizing the network structure. A preliminary analysis shows auspicious results which will be detailed in the final paper.