Research On Classification Method Of Hyperspectral Remote Sensing Image Under Random Projection Framework

With the continuous improvement of sensor technology,the spectral resolution of remote sensing images is getting higher and higher,usually including dozens to hundreds of bands.While providing rich spectral information for image classification,it also seriously increases the computational cost required for hyperspectral remote sensing image classification,so dimensionality reduction is usually required before classification.For large-scale hyperspectral remote sensing images,the traditional dimensionality reduction methods have some problems,such as high hardware requirements and long running time.Random Projection(RP)has the characteristics of being independent of high dimensional data,simple in calculation,etc.It is a dimensionality reduction method with little information loss.To solve these problems,under the framework of RP,this thesis proposes a fast dimensionality reduction method suitable for largescale hyperspectral remote sensing images and RP oriented hyperspectral remote sensing image classification method.The main contents are as follows.(1)Aiming at the problem that the very wide lower bound of the projection dimensionality of RP may lead to inability to reduce the dimensionality,a Tighter RP(TRP)algorithm is proposed,and it is theoretically proved that the algorithm satisfies the relative invariance of distance.The proposed TRP algorithm defines a tighter bound of the projection dimensionality by constraining the probability bound of all vector pairs to satisfy the relative invariance of distance after projection,thereby projecting hyperspectral remote sensing images into a lower dimensional space.In addition,for hyperspectral remote sensing images whose scale is much larger than the number of bands,it shows the maximum scale of the image that can be applied by the TRP algorithm when dimensionality reduction can be guaranteed.To verify the feasibility of the TRP algorithm,the lower bounds of projection dimensionalities of five real hyperspectral remote sensing images were analyzed.The results show that the lower bound of the projection dimensionality of the TRP algorithm is smaller than the lower bound of the projection dimensionality of the RP algorithm,and can be smaller than the number of bands,that is,the dimensionality reduction of hyperspectral remote sensing images can be realized.(2)Aiming at the problem that the lower bound of projection dimensionality of RP is limited by the amount of high dimensional data,a Partitioned RP(PRP)method is proposed,and it is theoretically proved that the dimensionality reduction method by dividing large-scale hyperspectral remote sensing images does not destroy the relative invariance of distance.The proposed PRP algorithm divides the hyperspectral remote sensing image into multiple subimages evenly,and each small-scaled sub-image uses the same projection matrix to obtain multiple low dimensional sub-images,thereby generating a low dimensional image by combining all low dimensional sub-images.In addition,for hyperspectral remote sensing images whose scale is much larger than the number of bands,it shows the maximum scale of sub-images that can be applied by the PRP algorithm when dimensionality reduction can be guaranteed.To verify the feasibility of the PRP algorithm,the lower bounds of projection dimensionalities of five real hyperspectral remote sensing images were analyzed.The results show that the lower bounds of the projection dimensionality of the PRP algorithm are smaller than the number of bands of the corresponding images,that is,the dimensionality reduction of large-scale hyperspectral remote sensing images can be achieved.(3)To fully consider the separable information of each class,a projection matrix selection strategy based on the separable information of a single class is proposed to improve the class separability of low dimensional images,thereby generating low dimensional images suitable for subsequent hyperspectral remote sensing image classification.The strategy uses the intra-class variance and inter-class distance of a certain class as measures to weigh the degree of separability of the class.Each element in the projection matrix is multiple samplings,and the projection matrix element with the strong separability of the class is selected in turn,thereby selecting the projection matrix that optimizes the separability of a certain class in the low dimensional image.To verify the effectiveness of the projection matrix selection strategy,the nonlinear projection method is used to visually analyze the degree of separability.The visualization results show that the projection matrix selection strategy based on single class separable information can effectively increase the class separability between classes.(4)A RP oriented hyperspectral remote sensing image classification method is proposed.First,on the basis of TRP and PRP algorithm,the projection matrix selection strategy based on the separable information of a single class is used to project the features of a certain class of objects,and the separability degree as the measure is used to weigh the quality of the projection results,thereby obtaining the low dimensional image with the best separability of the class.The Minimum Distance(MD)classifier is used to calculate the distance matrix,in which the low dimensional samples are also projected through the same projection matrix.Then,the above steps are repeated for all classes.Finally,the distance matrix of all classes is integrated as the final similarity measure matrix.The similarity measure matrix is used to obtain the membership class of the spectral vector by minimizing the similarity measure matrix,thereby realizing the classification of hyperspectral remote sensing images.(5)Verify the validity and feasibility of the proposed classification method.Taking largescale hyperspectral remote sensing images as experimental data,the MD classification method based on projection matrix selection of all classes of separable information,the MD classification method based on Compressive-Projection Principal Component Analysis(CPPCA)and Cumulative Agreement Fuzzy C-Means(CAFCM)methods are compared and analyzed.At the same time,the classification experimental results are analyzed from two perspectives,qualitative and quantitative.Among them,the qualitative evaluation is to analyze the experimental results by visual interpretation.And the quantitative evaluation is to evaluate the experimental results through the classification accuracy index calculated by the confusion matrix and the running time of the algorithm.The experimental results show that the proposed classification method has the ability to reduce the number of large-scale hyperspectral remote sensing image bands,and has excellent classification accuracy and running time.There are 45 figures,45 tables and 138 references.