Unwrapping SAR interferograms with localized subsidence signal using deep neural network

Phase unwrapping is an indispensable processing step of InSAR. However, conventional methods often underestimate the deformation due to severe noise and/or dense fringes. Here, we develop a new deep neural network to unwrap interferograms with localized subsidence signal. We train the network using synthetic interferograms with two-dimensional Gaussian shape subsidence and complex Gaussian noises, and apply the network to real interferograms with localized mining subsidence. The proposed method outperforms the standard methods by 76.3% on synthetic interferograms and similar to 50-times faster on real interferograms. The promising result shows the potential for rapid monitoring and quantification local deformation distributed in large area.