Research On Fully Polarized SAR Image Classification Method Based On Convolutional Neural Network

Polarimetric synthetic aperture radar(PolSAR)is an active imaging technology.It obtains the characterization of the observed targets by transmitting electromagnetic waves and receiving the electromagnetic echo reflection,and is widely used in vegetation observation,environmental protection,and urban planning,etc.As one of the most concerned issues in the interpretation of the PolSAR image,the PolSAR image classification has received critical attention.In recent years,the classification methods based on the convolutional neural networks has become a research hotspot in the field of PolSAR image classification due to the powerful feature extraction ability.Based on the research of the convolution neural network,we introduced it to the PolSAR image classification,and the main works of this paper are as follows:1.We studied the PolSAR image classification methods based on the 3D convolutional neural networks,and proposed a lightweight 3D convolutional neural network that has fewer parameters and easier to train.The proposed method adopts three lightweight strategies,including spatial scattering separated convolution,global average pooling,and bottleneck structure,which reduces the number of parameters from convolution layers,full connection layers,and the network architecture respectively.Therefore,compared to the 3D convolutional neural network,the proposed method reduces the dependence on the training samples and is more suitable for the PolSAR image classification.The experimental results demonstrate that the proposed method can significantly reduce the parameters and improve the classification accuracy.2.Based on the complex-valued coherent matrix of the PolSAR image,we proposed a novel PolSAR image classification method based on the complex light-weight 3D convolutional neural network,which can fully exploit the polarimetric information of the complex data.The proposed method takes the complex-valued PolSAR data as the input matrix and extracts the polarimetric characteristics by complex-valued convolution layers and complex-valued nonlinear layers.Since there is no complex-valued input processing ability,the traditional convolution neural networks treat the PolSAR data as the real-valued data,which leads to the loss of the correlation information between the amplitude and the phase.The proposed method maintains the corresponding between the amplitude and phase of PolSAR data in the forward propagation,so as to reduce the loss of polarimetric information and improve the classification performance.3.A PolSAR image classification method based on the fully separated convolutions has been proposed.Firstly,in order to reduce the parameters,we introduced the fully separated convolution which replaces the 3D convolution with three 1D convolutions.Then,considering the 1D convolutions lack of feature extraction ability than 3D convolutions which affects the classification performance of the model,we adopt the convolution joint strategy to enhance the computational complexity of the convolution layer.In the traditional convolutional layer,the output feature map of each convolution kernel is independent of the other convolution kernels.The convolution joint strategy associates each output feature map with all the convolution kernels in the same layer,and eliminating the independence of the convolution kernels,thus increasing the complexity of the convolutional operation.For the proposed method,the network has become more light-weight,and also enhance the ability of feature extraction in the convolutional layers,which finally improves the classification performance.