Deep Co-training Densely Connected Network for PolSAR Image Classiffcation via Clustering-Based Sample Selected Strategy

The Polarimetric Synthetic Aperture Radar (PolSAR) image classification methods based on the deep learning have attracted more and more people's attention. However, the high classification accuracy at the cost of a large number of samples does not meet the characteristics of the PolSAR data. Therefore, the need for higher classification accuracy for the PolSAR data under the condition of small samples is imminent. In this article, we propose a deep co-training densely connected network (DCT-DN), which is a semi-supervised classification method to combine the collaborative training and the deep learning. Firstly, in order to classify the PolSAR images from two perspectives of pixels and regions, a deep co-training neural network based on the 1D convolution neural network (1D-CNN) and the 3D densely connected network (3D-DN) is constructed. The input data of the 1D-CNN is decomposed by the Freeman-Durden and Cloude, and the input data of the 3D-DN is the coherence matrix data. Secondly, in order to improve the confidence of the selected unlabeled samples, an unlabeled sample selection strategy combined with clustering method is proposed. Experimental results on real PolSAR data prove that the classification performance of proposed DCT-DN is higher than other comparative classification methods