Study Of Hyperspectral Remote Sensing Imagery Classification Based On Semi-supervised Locality Preserving Projections

Hyperspectral remote sensing has some characteristics such as detailed spectralinformation and strong correlation between the large amount of data. It is easy to fallinto “dimension disaster” if use the traditional classification algorithm. Therefore, itbecomes particularly important to process dimension reduction of high-dimensionaldata. There are some disadvantages in many dimensionality reduction algorithms, suchas the Principal Component Analysis (PCA), the Linear Discriminant Analysis (LDA)and so on. They can’t effectively use the category information of data or be strict to it.To solve these problems, this paper introduces a dimensionality reduction method,called the Semi-supervised Locality Preserving Projections (SSLPP).First of all, hyperspectral image and its characteristics were briefly introduced inthis paper, combined with supervised learning and unsupervised learning it make asummary of the feature extraction methods for high-dimensional data, then the SSLPPalgorithm is proposed; Secondly the principles and process of the SSLPP algorithm iselaborated; Subsequently, compared with several mainstream feature extractionalgorithms such as the Principal Component Analysis (PCA), the Locality PreservingProjections (SLPP), and the Supervised Locality Preserving Projections (SLPP), doescomparative experiments to validate the effectiveness of the SSLPP algorithm. Theclassification experiment is based on the real hyperspectral remote sensing imagery. Thefirst using a variety of algorithms to process the original data set, then to use the KNearest Neighbor (KNN) classifier to classify the processed low dimensional data, andcalculate the overall classification accuracy of each algorithm. the experimental resultswould validate the effectiveness of the SSLPP algorithm; Finally to investigate whetherthe SSLPP algorithm has better cooperation with different classifiers, using threedifferent classifier combined with the SSLPP algorithm to classify four realhyperspectral remote sensing image data, and get the related recognition rates.According to experimental result, there are following advantages of the SSLLPPalgorithm compared with the other feature extraction algorithms:①Relative tounsupervised dimensionality reduction algorithm, it makes full use of the categoryinformation of data, which making the high-dimensional data getting a better separablein low-dimension mapping;②By comparison with the supervision dimensionreduction algorithm, it is not only using the labelled samples but also making full use of a large number of unlabelled samples at the same time, which preferably maintain theintegrity of the original data in low-dimension mapping;③In classifying processing forhyperspectral data, it gets a high accuracy and avoids calibrating all categories oforiginal data, which improving the efficiency of data processing.Above all, this paper mainly studies the hyperspectral remote sensing image basedon a semi-supervised learning of feature extraction and classification. A semi-supervisedfeature extraction algorithm is proposed in this paper, and it is proved the validityaccording to several classification experiment of hyperspectral remote sensing image.