Optimizing Φ-Net Using TanDEM-X Bistatic Data for High-Resolution DEM Generation

Abstract

Phi-Net is a deep residual learning architecture and is currently one of the most accurate state-of-the-art approaches for Synthetic Aperture Radar interferometric (InSAR) phase filtering and coherence estimation. The network has proven to outperform classical denoising strategies concerning the estimation of InSAR parameters and therefore is very suitable for the generation of high-resolution Digital Elevation Models (DEMs). However, some limitations are shown over critical areas characterized by low signal-to-noise ratio (SNR) or by geometric distortions, i.e. shadow and layover. There, Phi-Net wrongly reconstructs the InSAR phase and inserts artefacts, thus leading to an unreliable input for typical phase unwrapping algorithms. Moreover, being trained with synthetic data only, there is still potential for an optimization which is closely related to the patterns that can be found in real InSAR data. In this paper we propose a preliminary analysis which exploits TanDEM-X bistatic data in order to optimize and fine tune Phi-Net. In particular, we address the problem of removing artefacts in the filtered InSAR phase and improving the estimation performance, which would allow for the generation of a more accurate and reliable unwrapped phase and higher-precision DEMs.