PolSAR Image Classification Based On Fully Convolutional Neural Network

Polarimetric synthetic aperture radar(PolSAR)adopts multi-frequency and multi-channel imaging mode,which can monitor the earth day and night.Besides,PolSAR has the advantages of strong penetration and high resolution.As one of the key techinologies of PolSAR image interpretation,PolSAR image classification has been payed great attention at home and abroad.PolSAR image classification is a pixel-wise prediction task.The convolutional neural network(CNN)has achived outstanding accomplishments for PolSAR image classification recently,but it exists the disadvantages of repeated calculation and memory occupation.The fully convolutional neural network(FCN)is a pixel-to-pixel,end-to-end dense classification network,which is greatly promising for the PolSAR image classification.Compared with CNN,FCN has obvious merits for PolSAR image classification: 1)take arbitrary-size images as input;2)enable to maintain the 2-D structure of the input;3)settle PolSAR image classification tasks with efficient dense learning,which avoids repeated calculation and memory occupation.In order to fully exploit the characteristic of PolSAR image,this thesis proposes three methods for PolSAR image classification based on the complex characteristic of PolSAR image and FCN model.The specific research work is summarized as follows:(1)Based on the defect of insufficient labeled samples of PolSAR image and excessive loss of detail information in FCN model,this thesis proposes a new parallel dual-channel dilated fully convolutional network(DCDFCN)for PolSAR image classification.First,to settle lacking sufficient labeled samples,the semi-supervised fuzzy c-means clustering algorithm(SSFCM)algorithm combined with Wishart distribution is exploited to preprocess the PolSAR images,aiming to obtain the pseudo labels of unlabeled pixels and enlarge the labeled samples.Then,to reduce the detailed loss caused by successive pooling layers and improve the density of output feature maps,we design a new FCN variant named dilated FCN(DFCN)by introducing a dilated convolution block to FCN.Finally,the thesis uses two similar DFCN frameworks in parallel with different convolutional kernels to design a new DCDFCN model and completes the classification task of PolSAR image.Experiments on different PolSAR images prove that the property of proposed method is superior to several other methods.(2)Based on the significance of phase information for PolSAR image and the disadvantage of insuffient labeled samples of PolSAR image,this thesis proposes a multi-scale FCN model for PolSAR image classification by adopting complex-valued stacked dilated convolution(CVSDC)in the structure of FCN,which is short for CVSDFCN.Firstly,a stacked dilated convolution layer with different dilation rates is constructed to capture multi-scale features of PolSAR image.Meanwhile,sharing weight is employed to reduce the parameter redundancy and calculation burden.Then,due to limited available labeled samples for PolSAR image,the encoder-decoder structure of original FCN is reconstructed.Finally,to take fully use of the phase information of PolSAR image,the proposed model is trained in complex-valued domain rather than real-valued domain.Experiments on different PolSAR images prove that the property of proposed method is superior to several other methods.(3)Inspired by the idea of "ensemble learning",this thesis combines three independent network models: slide window fully convolutional neural network(SFCN),complex value fully convolutional neural network(CVFCN)and CSDFCN to form a new network framework with multi models’ joint learning for PolSAR image classification,which enables the different models’ respective advantages complementary to each other.Experimental results on different PolSAR images show that this method can further improve the classification accuracy of PolSAR images.