Low frequency and radar's physical based features for improvement of convolutional neural networks for PolSAR image classification

Although various deep neural networks such as convolutional neural networks (CNNs) have been suggested for classification of polarimetric synthetic aperture radar (PolSAR) images, but, they have sev-eral deficiencies. CNNs have weakness in producing classification maps with reduced noise and also are disabled in extraction of polarimetric/scattering information to explore the physical characteristics of the radar image. A deep neural network based on convolutional blocks is proposed for PolSAR image classi-fication in this work that deals with the above difficulties. The low frequency components of the PolSAR image are added to the output of convolutional blocks to help the network to learn noise reduction. Moreover, eight fuzzy clustering maps obtained by the polarimetric entropy and averaged alpha angle are extracted as radar's physical feature maps which concatenated with the spatial features extracted by convolutional blocks. So, the proposed network while learns to reduce the speckle noise, learns to simultaneously extract the spatial-physical characteristics of the PolSAR cube. The experiments on two real PolSAR datasets show superior performance of the proposed network compared to CNN, residual net-work and some other well-done networks.(c) 2021 National Authority of Remote Sensing & Space Science. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).