Convolutional residual networks have shown great success for classification of polarimetric synthetic aperture radar (PolSAR) images especially when the depth of network is increased and limited training samples are available. But, the ability of convolutional kernels is in extraction of spatial features from neighboring pixels. They may not sufficiently able to extract the radar's physical features from the PolSAR images. To deal with this difficulty, a residual network is proposed in this work. In addition to the feature maps extracted by convolutional kernels of previous layers, the proposed network injects the physical feature maps extracted by the H-A-alpha\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{\alpha }$$\end{document} decomposition to each addition layer. Moreover, the use of a convolutional autoencoder behind the residual blocks is proposed to reduce the speckle noise. The proposed method provides more accurate classification maps compared to conventional residual networks and several state-of-the-art methods. However, it needs more running time for implementation.