Polarimetric SAR Classification Method Based On Feature Dimension Reduction And Classifier Fusion

Polarimetric Aperture Radar(SAR)is an advanced active microwave remote sensing method which can obtain rich information of objects.It has the multi-channel polarized image to obtain the target object and enhances the understanding of the target scattering mechanism through the polarization information processing of the object,and improves the detection,recognition and classification ability of the object.In recent years,the development and application of polarimetric SAR system at home and abroad has promoted the rapid development of polarimetric SAR theory and classification method.Polarization SAR target classification and image interpretation can achieve fine classification,identification and description of different objects,such as geology and geomorphology exploration,vegetation coverage survey,crop growth assessment,marine environmental monitoring and battlefield investigation and monitoring,etc.Military and civilian areas have important application prospects.Therefore,the study of polarimetric SAR classification is a hotspot in the field of radar signal processing and remote sensing information processing,which has important theoretical significance and application value.The feature extraction algorithm and the classifier design of the object are the key links to realize accurate polarimetric SAR image classification and image interpretation.Therefore,in order to improve the accuracy of feature recognition in polarimetric SAR images,the algorithm of polarized SAR feature extraction algorithm and different classifier fusion algorithm is studied emphatically.A variety of classification algorithms are developed to integrate polarized SAR Geographic classification software,obtained a higher classification rate of polarimetric SAR simulation and measured data processing results.The specific research contents are as follows:In order to solve the problem of feature extraction in polarized SAR image interpretation,T-LDA and T-LPP feature extraction based on tensor representation are proposed along the combination of tensor algebra theory and traditional feature dimension reduction algorithm.algorithm.By using the neighborhood sliding window method to construct the local tensor feature data of the pixel,and the matrix form feature extraction algorithm is extended by tensor expansion.The results of simulation and measured data show that the characteristic data of polarized terrain can be used to extract the feature quantity with high degree of discrimination,and the redundancy between features can be removed,which can help to get the classification result The In addition,the influence of the key parameters such as training sample size,neighborhood window length and dimensionality dimension after dimension reduction on the accuracy of classification is also discussed.In this paper,we support the SVM,K Nearest Neighbor algorithm and H/A/Wishart three classifier and decision fusion algorithm for the design problem of polarized object classifier.An improved method based on weighted voting algorithm is proposed to improve the correct rate of decision making.In this paper,we consider that different classifiers have different classification effects on different objects,and there are complementary information between them.We introduce the weighted voting method to combine the SVM,K Nearest Neighbor algorithm and H/A/Wishart.On the basis of the traditional weighted voting method,considering the difference of the classification effect of different classifiers on different polarized objects,the classifier correctness matrix is introduced,and each sample is obtained in each classifier The classification of the situation,the dynamic calculation of each weight.Through the classification experiment of the measured data,it is proved that the fusion classifier has good effect in reliability and stability.In this paper,based on the theoretical study of feature extraction algorithm and classification algorithm,this paper designs and composes a set of polarized SAR classification software,integrates polarimetric SAR image reading,feature extraction algorithm,data classification algorithm,image display,image preservation and other functions.The software can meet the needs of practical application analysis by including a variety of data feature extraction and classification algorithms.