Hyperspectral Remote Sensing Image Classification Based On Complex Networks And BOVW

Hyperspectral remote sensing technology used the hundreds of electromagnetic wave bands to image the ground objects.The obtained Hyperspectral remote sensing images presented extremely abundant material information.The images fully reflected the physical or chemical composition,so as to determine the types of ground objects in a more refined manner.At the same time,due to the large amount of data acquired by hyperspectral remote sensing images,there were also many problems in hyperspectral remote sensing images classification,such as limited data available for training,the stability of the diversity data of hyperspectral features of ground objects in a large space is affected by illumination,atmosphere,and the dimension disaster caused by high dimension.Therefore,it was an urgent problem to be solved how to ensure a high accuracy of classification in the relatively limited data while reduce the calculation pressure caused by high-dimensional data.According to the problems,in order to solve the problems of low classification accuracy and low algorithm performance caused by large amount of data in the classification technology of hyperspectral remote sensing images,we proposed the following three measures to explore a more effective method for hyperspectral remote sensing images classification.(1)We proposed a hyperspectral remote sensing image feature extraction and classification method based on complex network and bag of visual words.Firstly,the high-dimensional spectral vectors were clustered by K-means method,and the codebook was formed by clustering center.The spectral vector to be classified was compared with the cluster codebook vector to obtain the word frequency characteristics of the sample points.Second,The spectral vector to be measured is transformed from 1×N structure to matrix structure.Then complex networks to be created which used of pixel values and position matrix as a reference.After the network dynamic evolution,we calculated the subnetwork topology parameters and combined them as spectral characteristic vectors.It is concluded that the inner link between sample points vector wave bands.Finally,the two are fused for classification.(2)In view of the spatial characteristics of hyperspectral remote sensing images,this paper combines the complex network method with the local binary mode(LBP)algorithm and Gabor algorithm respectively,which was CN-LBP algorithm and CN-Gabor algorithm.Among them,CN-LBP algorithm extracted features from two aspects of hyperspectral remote sensing images respectively.Complex networks method was used to describe the spectral curve structure features of a single pixel,and LBP algorithm was used to represent the correlation features between each sample point and the surrounding sample points.CN-Gabor algorithm also used complex networks for spectral structure analysis.In terms of spatial features,three-dimensional Gabor filter banks were used to filter hyperspectral remote sensing images.Voting mechanism was used to conduct statistics on classification of multiple Gabor filters and obtain quantized histograms,which were used as texture features of images.The algorithm used the feature parallel fusion method,which took spectral features and spatial features as the real and imaginary parts of the complex vectors,and reduced the dimension through PCA algorithm.(3)The classification performance of CN-BOVW algorithm is proved by experiments on open Salinas and KSC data sets.It was proved from the overall classification accuracy(OA),average classification accuracy(AA)and Kappa coefficient.The classification accuracy of the proposed CN-BOVW algorithm could reach 98.97%.At the same time,the fusion algorithm of complex networks and spatial features was verified by Indian Pines and KSC datasets,and the overall classification accuracy of CN-LBP and CN-Gabor algorithms could also reach 97.25%and 97.21%.Through the above three works,it could be proved that the method of complex networks fusion bag of visual words and the methods of complex networks fusion spatial features proposed in this paper had better classification accuracy than the existing algorithms,and the algorithm also had a certain application value.