Research On PolSAR Crop Classification Method Based On Tensor Affine Transformation Neural Network

Crop classification is an important basis for agricultural monitoring and grain yield estimation.In recent years,with the increasing number of spaceborne polarized synthetic aperture radar(Pol SAR)systems,multi-temporal Pol SAR images provide abundant hydrogeological information for agricultural production and effectively improve crop classification accuracy.The traditional neural network needs a large number of parameters when processing multi-temporal data,which often leads to the network can not complete the training,and usually needs to use the dimensionality reduction method to compress the data.However,this makes the dimensionality reduction network and the classification network in crop classification disjoint and finally reduces the classification accuracy.In view of the above problems,this paper tries to propose an integrated solution of dimension reduction and classification from the perspective of tensor analysis,and the main research results are as follows:(1)In this paper,tensor affine transformation is applied to the forward calculation of BP neural network,and a Tensor Affine Transformation Network(TATN)is proposed.In the proposed network model,the crop feature data are input into the network in the form of original high-order tensors.In the hidden layer,tensor affine transformation is used to extract the features,which significantly reduces the number of parameters and computation of the hidden layer.The output layer is classified by tensor weighting and softmax function.This paper deduces the parameter training method based on error back propagation and completes the network training.TATN can directly process multi-temporal data,avoiding the use of dimensionality reduction and forming a single network model for crop classification.In the process of multi-temporal data,the structure information of tensor affine transformation can be preserved to achieve accurate classification of crops.The experimental results show that compared with SAE-SVM and SAE-CNN,TATN network has higher classification accuracy,which is improved by 23.7% compared with SAE-SVM and 4.7% compared with SAE-CNN.(2)In order to further improve the classification performance of TATN network,the paper proposes Local Tensor Affine Transformation Network(LTATN)inspired by the local perception characteristics of CNN.The proposed network model completes local affine transformation of tensors with smaller transformation matrix,which accelerates the training speed of the network and improves the classification performance.In terms of crop classification of multi-temporal data,the classification accuracy of LTATN can reach 99.36%,which is superior to TATN network and CNN network.