Semi-supervised Neighborhood Preserving Embedding Used In Hyperspectral Image Classification

Hyperspectral remote sensing image is spectral space formed by spectralcharacteristics figures of ground objects detected by satellites. It contains abundant ofspectral, space and radiation information. The classification process of hyperspectralremote sensing images is the category of images’ sample data, the purpose is to dividethe image space into some regional spaces of surface features through classification ofimage space. Currently, hyperspectral remote sensing images are more and more usedin fields like ocean remote sensing. However, because of the adjacent bands ofhyperspectral date have strong relevance and the spectral has high dimensions, thetraditional classification methods are used to be unable to get the desired results due tothe large amount of operation data and the poor classified identification, and get intothe ‘Huges’ phenomenon. Thus, how to get the information which is useful for sampleclassification from large amount of hyperspectral remote sensing data and improve thesample classification accuracy of images is becoming one of the main researchingobjects of hypersectral image classification. We do some research work as followsbased on the features of hyperspectral images and existing dimension reduction andclassification methods:①We do a lot of work in dimension reduction problem. This paper reviews thedevelopment process of dimension reduction algorisms, we select the frequently usedneighborhood preserving embedding(NPE) algorism from the existing dimensionreduction manifold algorisms, which is to preserve the local manifold structure of thesamples while dimension reducing, and we propose semi-supervised neighborhoodpreserving embedding(SSNPE) algorism based on NPE. SSNPE algorism use both thelabeled samples and the unlabeled samples on the basis of preserving the manifoldstructure of the samples, which is to use both the labeled samples from the same classand the unlabeled samples from neighborhood, and then we improve the identificationfeature of dimension reduction data through improving the weight of neighborhoodlabeled samples. SSNPE algorism is a semi-supervised algorism, it solves the sampleinadequate problems of hyperspectral remote sensing images to some extent andimprove the identification feature of the data samples.②We do a lot of work in classification methods of hyperspectral remote sensingimages. This paper introduce the frequently used classification methods of hyperspectral remote sensing images, we select the K-nearest neighborhood(KNN)method which is to classify samples through comparing the distance between thesamples and other samples of different labels to classify the dimension reduction databased on SSNPE dimension reduction. We prove the necessity of SSNPE algorism’sdimension reduction process which is to reduce the amount of calculation of data andto improve the identification feature of date through classification using frequentlyused hyperspectral image classification methods which are BP neutral network, SVM,KNN methods to classify the sample data.③The existing classification methods are largely supervised and unsupervisedmethods, this paper proposes a semi-supervised K-nearest neighborhood(SSKNN)algorism based on KNN classification method and realize the SSNPE algorism’s actualapplication of hyperspectral image classification.To sum up, this paper proposes a new semi-supervised dimension reductionmethod and classification method based on NPE and KNN algorism, we reduce theamount of calculation and improve the identification feature of data. The experimentalresults on hyperspectral data set show that the algorism’s classification feature hasimproved clearly.