Polarimetric SAR Image Classification Based On Dictionary Learning And Softmax Deep Stacking Network

Polarized Synthetic Aperture Radar(Pol SAR)is an active microwave sensor which is carried on a satellite or aircraft.It transmits and receives electromagnetic waves of different polarization modes through two channels,and acquires polarization data of four channels,which can fully reflect the polarization character of the targets.As an important component of Pol SAR image interpretation,it has significant research and application value in the academia and military fields.A Pol SAR image classification method based on dictionary learning and Softmax Deep Stacking Network(SDSN)is proposed.Firstly,SDSN is proposed based on the Deep Stacking Network(DSN).Then,a dictionary is extracted by analysis KSVD algorithm and is used to initialize the lower-weight matrix of the lowest module in SDSN.Finally,the corresponding acceleration algorithm for SDSN is proposed and SDSN is trained for Pol SAR image classification.The main work of this paper is as follows:Pol SAR image classification based DSN is utilized.The original features extracted from the polarization coherence matrix are directly input into the DSN to extract deeper feature and classify Pol SAR image.The powerful feature extraction ability of deep learning is used to overcomes the problems of existing Pol SAR feature extraction methods which lack of autonomy and adaptability.Its a simple idea,and the obtained classification results have good regional consistency,and the classification accuracy is also improved significantly.The SDSN and corresponding acceleration algorithm are proposed and applied to Pol SAR image classification.Firstly,the linear output layer in each module of the DSN is replaced by a nonlinear output layer with Softmax activation function,and changed the objective function to propose the SDSN.At the same time,in order to accelerate the training of the SDSN,a corresponding acceleration algorithm is proposed.The approximate optimal solution of upper-layer matrix is calculated by L-BFGS algorithm before update lower-layer matrix by back propagation in each iteration.SDSN overcomes the shortcoming of DSN which are not suitable for multi-classification problems,and improves the accuracy of Pol SAR image classification.The corresponding acceleration algorithm converges faster than non-accelerated algorithms.A Pol SAR image classification method based on dictionary learning and SDSN is proposed.Firstly,an analysis dictionary is extracted by analyisis KSVD algorithm from the original data features and then the analytical dictionary is used as the initial lower-weight matrix of the lowest module of SDSN.Finally,the SDSN is trained for Pol SAR image classification.This method combines dictionary learning and SDSN naturally.It can not only solve the network initialization problem with a simple method,but also use the shallow features extracted from the data as the initialization point,which is more suitable to the data.In this method,the model converges fast and the classification accuracy is high.