Research On Crop Classification Based On Multi Temporal PolSAR Data Polarization Feature Dimension Reduction

Agricultural crop monitoring and classification is an important application of polarimetric synthetic aperture radar（POLSAR）imagery.Polarimetric SAR system is an active remote sensing system,which can obtain the high-resolution remote sensing information of observation scene,including the polarization intensity and phase of the target.In crop classification,on the one hand,because the same crop has different external characteristics in the growth process,it is difficult to provide enough information for single time POLSAR image.In addition,with more and more spaceborne SAR systems entering the earth orbit,a large number of POLSAR data have been obtained,which provides a practical operation basis for multi temporal data analysis.On the other hand,with the continuous development and breakthrough of deep learning technology,the combination of deep learning technology and remote sensing interpretation has achieved great success.Combining these two major trends,this article has conducted an in-depth discussion on the crop classification problem of PolSAR data,Specifically,from the two aspects of single-phase and multi-temporal aspects,In the single-phase classification,a lightweight and efficient Tiny_CNN convolutional classification network is constructed.In the task of multi temporal classification.The Stack Sparse Auto-Encoder（S-SAE）is introduced to reduce the dimension of high-dimensional polarization features generated by multiple polarization decomposition methods,and combine with Tiny_CNN to complete the classification.Finally,this article quantitatively compares,discusses and analyzes the influence of S-SAE parameters on the classification performance of the classifier.The research content and conclusions of this article mainly include:（1）First of all,this paper first combines deep learning technology.In the single-phase PolSAR crop classification process,based on the classic convolutional neural network LeNet₅,combined with the characteristics of the large scale of remote sensing images and limited training samples,a light and efficient convolutional neural network classification method Tiny_CNN.Through experimental comparison with different convolutional integration methods on different data sets,under all control methods and different feature inputs Compared with other CNN networks such as LeNet₅,the classification accuracy is improved by at least 5%.（2）Aiming at the single-phase PolSAR remote sensing image classification problem,the built Tiny_CNN classification network is used to explore the classification performance under different inputs.Through the AirSAR standard data set and the Indian Head sixth-phase data experiment,under different classification models,the classification accuracy of polarization decomposition features is at least 3%higher than that of the original data.In addition,compared with non-dimension reduction features,the classification accuracy of polarization decomposition features is improved by at least 1.5%under the same classification method.（3）Aiming at the classification problem of multi-temporal data,based on the discussion of single-temporal classification,this paper proposes a classification method based on dimensionality reduction features of polarization decomposition.This paper uses the most commonly used multiple incoherent decomposition methods to fully extract Characteristic information of multi-temporal PolSAR data.In order to avoid the problem of "dimensionality disaster" when directly using multi-temporal data,the specific performance of the classification method based on time series vector and the classification method based on feature dimensionality reduction were compared and verified.The experimental results show that the multi temporal dimension reduction classification method based on Principal Component Analysis（PCA）can make more efficient use of the classification information in multi temporal data,and the classification accuracy is improved by 15%compared with the multi temporal classification method LSTM based on time series vector.Compared with the 1-dimensional 1D_TCNN constructed in this paper,the classification accuracy is improved by about 3%.（4）Finally,in the discussion of multi-temporal classification,this paper proposes a crop classification method based on stack sparse autoencoder（S-SAE）dimensionality reduction.The features are efficiently processed by dimensionality reduction to fuse multi-temporal features,so as to achieve the purpose of improving classification accuracy.Based on the analysis of the influence of S-SAE network parameters and network structure on the dimensionality reduction effect,a three-hidden-layer S-SAE network is established to efficiently reduce the dimensionality of the acquired high-dimensional and multi-temporal features,and finally use different reductions.Dimensional methods such as PCA,Locally Linear Embedded（LLE）,and different classification complex Wishart,Convolutional Neural Network（CNN）and Support Vector Machine（SVM）classification method compares and verifies the methods proposed in this paper.The experimental results show that:in terms of training rate,increasing the training rate of unsupervised training samples can improve the performance of S-SAE.However,with the increase of pre-training samples,the improvement of classification performance becomes weaker.Regarding the depth of the network and the size of the hidden layer,as the network deepens,the classification accuracy is getting higher and higher.However,when a certain depth is reached,the network dimensionality reduction effect will no longer improve.In addition,by reducing the step size of the search grid,this article also has new discoveries.If the number of neurons in the first hidden layer is close to the number of input features,and the number of neurons in the second hidden layer keeps a certain gap with the two layers before and after,better data dimensionality reduction performance can be obtained.In addition,experiments show that sparse regularization termsβ,regularization term weights λ,and sparse term parameters ρ have uncertain effects on the performance of S-SAE.If the sparse self-encoding in a single layer can properly select the value parameters,the OA of the classification result will be increased by 5%,which is even higher than the classification accuracy of the optimized S-SAE.The experimental results show that with the control methods,the optimized parameters of the S-SAE+Tiny_CNN method has obvious classification accuracy advantages on three different experimental data.Compared with traditional dimension reduction classification method,the classification accuracy is at least increased by 3.5%.