Research On Hyperspectral Imagery Classification Based On Support Vector Machine And Neighboring Representation

As a three-dimensional image,hyperspectral remote sensing image integrates the spatial and spectral information of the image.With the rapid development of hyperspectral remote sensing technology,hyperspectral images have been widely used in image fusion,spectral separation,classification,agricultural monitoring,target detection and fast calculation in the past decades.Due to the characteristics of high spectral resolution of hyperspectral images,hyperspectral image classification has gradually become an important research field.It can not only detect and distinguish small differences between land cover categories with high accuracy,but also play an indispensable role in the economic and military fields.However,the characteristics of hyperspectral data also make hyperspectral classification face many challenges.Therefore,this paper focuses on how to fully extract and combine the spectral and spatial features of hyperspectral images to improve the classification accuracy.The main contents of this paper are as follows:1.Hyperspectral images have tens or even hundreds of bands,but the number of training samples is limited.This fact leads to the dimensionality disaster called Hughes phenomenon.To overcome the shortage of training samples in traditional SVM-based hyperspectral image classification algorithm,a semi-supervised classification algorithm based on weighted spectral and spatial information is proposed.Firstly,the hyperspectral data is updated by the weight between adjacent samples,which makes the hyperspectral data discriminant.At the same time,semi-supervised classification is introduced to alleviate the shortage of hyperspectral training samples.Finally,unlike the usual combination method based on SVM and logistic regression,the spatial neighborhood information extracted from SVM by rectangular window is sent to logistic regression for classification,so as to make full use of the classification information of data.Experiments on three sets of real hyperspectral datasets show that the proposed algorithm can effectively improve the classification accuracy when the number of labeled samples is limited.2.In hyperspectral image classification,in order to better characterize the variability of spatial features at different scales and the details and edges of hyperspectral images,a new framework of multi-scale spatial information fusion is proposed.The new framework includes five parts: spectral gradient,SVM,multi-scale spatial information extraction,information fusion and spatial random forest.First,spectral gradient technique is used to process the original hyperspectral data to obtain more intrinsic and comprehensive information.Then,the updated data is sent to SVM to obtain probabilistic output,and the spatial context information with different scales is further extracted.Finally,the multi-scale spatial features are fused with the corresponding weights,and then fed to the random forest classifier as input.In the experimental results,a large number of experiments on three sets of real HSI datasets including AVIRIS Indian Pines,Salinas and ROSIS Pavia University verify the effectiveness of the proposed method.3.In order to better fit the complex non-linear characteristics of hyperspectral images,a spectral and spatial classification algorithm based on multi-scale adaptive bilateral filter is proposed.Firstly,the algorithm uses spectral-based SVM to process hyperspectral images to obtain details and edge information.Next,the S-function is used to normalize and transform the output of the first step to further enhance the non-linear structure of hyperspectral images.Then,we use smoothing filters to extract the spatial neighborhood information of each pixel in hyperspectral images,which is expected to eliminate the inherent changes in small neighborhoods.Finally,we adopt two strategies to obtain the final classification results: one is to perform symbolic operations directly on the obtained features;the other is to send the result features back to the SVM classifier to obtain the final classification results.The experimental results show that the proposed method can not only effectively improve the classification accuracy of hyperspectral images,but also achieve fast classification.