Research On Dimensionality Reduction And Classification Of Hyperspectral Image Based On Manifold Learning

Hyperspectral remote sensing is a comprehensive discipline based onelectrical-magnetic spectraltheory, geoscience and computer science. With the rapiddevelopment of remote sensing technology, the capability of obtaining data of remotesensing image increases constantly. Because of the special high resolution ofhyperspectral remote sensing image, it has been playing an important role in theanalysis of geological hazards, military target recognition, ecological environmentalmonitoring and agricultural production. Many methods designed for multispectralimage tend to perfect, but the high dimensionalities and large data of hyperspectralimage bring some new problems to traditional methods of multispectral, so it is veryimportant to explore new methods of hyperspectral image. This paper starts with thecharacteristics of hyperspectral image and existing algorithms, and then dimensionalityreduction and classification methods are further researched. The main research contentsare as follows:①Dimensionality reduction methods of hyperspectral image are deeplyresearched in this paper, and they are divided into three types: band selection, featureextraction and data fusion, and feature extraction is commonly used. On the basis offeature extraction methods, a new semi-supervised manifold learning method isproposed─semi-supervised marginal Fisher analysis(SSMFA). SSMFA learns theintrinsic structures of hyperspectral data using the local geometry information of thelabeled samples and unlabeled samples, and then it extracts discriminant features fromthe data. It can not only reduce data dimensionality and redundancy, but also providemore available information for classification. SSMFA is semi-supervised andself-learning method, so it can resolve the problem of lack of labeled hyperspectral datasamples to some extent.②Classification methods of hyperspectral image are researched in this paper, anda few common supervised methods are described in detail. The criteria for assessing theclassification accuracy derived from different classification methods are described.kNNS is applied to the classification of hyperspectral image based on SSMFA. Thelabel of new sample is determined by the distance between it and its neighbours fromdifferent classes. The experimental results on the hyperspectral image prove that it isvery necessary to reduce the dimensionality of hyperspectral data. The separability of data processed by SSMFA is better, which makes the classification accuracy of kNNShigher, and the computational complexity is decreased greatly.③Most of the existing feature extraction methods of hyperspectral image assumethat hyperspectral data may reside on one single manifold, and the data are projectedonto a uniform manifold space, but actually data from different classes may reside ondifferent manifolds. In this paper, a multi-manifold learning algorithm based on localand global preserving embedding (LLGPE) is proposed. First, the manifolds of differentclasses are learned by LLGPE for each class separately, and the data are projected ontolow-dimensional space. Then, the optimal dimensionality of each class is founded bygenetic algorithm (GA) from the viewpoint of classification. At last, classification isperformed under a minimum reconstruction error based classifier. It is very convenientto research the difference of manifold structure of different classes. The experimentalresults on the hyperspectral data show the effectiveness of M-LLGPE algorithm.Inclusion, dimensionality reduction and classification methods of hyperspectralimage are researched in this paper, and new algorithms are proposed. The experimentalresults prove that the algorithms can achieve good results.