PolSAR image despeckling using trained models on single channel SAR images

Abstract

The use of polarimetric synthetic aperture radar (PolSAR) images requires efficient PolSAR speckle filtering algorithms. Many traditional and classical PolSAR filters have been proposed during the last decade. But none of them are using deep learning to reduce the speckle of PolSAR images. In this paper, we employed the single channel SAR images to train the denoising convolutional neural network (DnCNN) model. To extend these models to PolSAR speckle reduction problem, the MuLoG (MUlti-channel LOgarithm with Gaussian denoising) framework is introduced. The experimental results show that this method is effective and efficient.