A Phase Filtering Method based on Deep Learning Network.

Phase unwrapping is a key step of interferometric synthetic aperture radar (InSAR), which transforms wrapped phase into the absolute phase. The accurate phase unwrapping requires high signal-to-noise ratio (SNR) value. Thus phase filtering is necessary to filter out the noise. This paper proposes an improved U-Net neural network for phase filtering. The images of noisy and noiseless interferometric phase are fed into the network for training. With the trained neural network, the noisy input interferometric phase can be directly smoothed to obtain high SNR phase image. Experiments show that the network can filter out phase noise effectively in the phase contiguous region, meanwhile it keeps the structure information of phase ambiguous edge region.