Terrain classification of Pol-SAR based on dilated convolution and polarization decomposition

In order to solve the problems of the lack of labeling data and the failure to make use of the polarization decomposition feature of SAR in Pol-SAR terrain classification methods based on deep learning, a polarimetric SAR terrain classification method based on dilated convolution and polarization decomposition is proposed. The convolutional neural network (CNN) based on pixel classification was trained at low sampling rate, and its convolutional layer parameters were transferred to the expansive convolutional network (DCNN) with the same structure, thus solving the problem of insufficient training data of the expansive convolutional network. Then the polarization decomposition feature reflecting the scattering characteristics of ground objects is combined with the expansive convolution feature map containing the high-dimensional spatial semantic information, and the random forest (RF) is constructed by using the joint feature for classification. Experiments show that the alliance of polarization decomposition features and feature maps can achieve more accurate classification, and the algorithm has high real-time performance due to the dilated convolution and random forest.