Dimension Reduction Of Hyperspectral Remote Sensing Image Based On Manifold Learning Method Research

Recently, hyperspectral remote sensing has become one of the important directionof remote sensing since. In hyperspectral remote sensing, its resolution is higher and itimprove the recognition ability of covered objects and also makes the quantitative orsemi-quantitative recognition of the terrain factor possible. It has been widely used inthe precision agriculture, water quality parameter inversion, mineral mapping and otherfields. There are rich in space, radiation and spectral information in hyperspectralimages but there are a lot of bands, large amounts of data, information redundancy,and the dimension disaster, so it is difficult to effectively use these information.Therefor, how to effectively deal with high-dimensional data to reduce the amountof data, identify its inherent laws, and obtain valid information, is one of hot researchtopic in the field of Pattern recognition and Machine learning. In this paper, the LocalPreserving Projection (LPP), the linear discriminant(LDA), and the method of acombination of linear discriminant analysis and Local Preserving Projection (LDA-PP)are introduced to the dimensionality reduction of hyperspectral data, and this methodmaintains the information of hyperspectral image and effectively reduces thehyperspectral image data. Then Support Vector Machine (SVM) is used to verify theeffectiveness of dimensionality reduction in classification. The main work of this paperis as follows:(1) The principles of the traditional dimension reduction metrods are analysised,including the globle preserving dimensionality reduction methods (PrincipalComponent Analysis and Linear Discriminant Analysis), local preservingdimensionality reduction method (Local Preserving Projection) and the method of acombination of linear discriminant analysis and Local Preserving Projection.(2) Because of image spectral discontinuity caused by the terrain, the shadow isremoved using shadow detection technology based on the color space transform andimage chroma, saturarion and brightness space of hyperspectral data. At last, remove theimage boundary after the restoration of the shadow.(3) In the experiment, OMSI-I and Hyperion data are used. The support vector machine (SVM) are used to classify the feature subsets. Then, the accuracy is evaluated.Experiment results show that the classification result based on LPP feature subset getsthe highest classification accuracy; classification accuracy based on LDA-LPP featuresubset is second to the result based on LPP feature subset,but LDA-LPP compress moreinformation into the first few bands,namely, higher compression efficiency; theclassification result based on PCA and LDA methods is poor than the LPP andLDA-LPP methods, and the classification accuracy based on the LDA methods issuperior to the result based on PCA, because PCA dimensionality reduction method isaim to protect the maximum energy of the original information, while the LDAdimensionality is to diatinguish the different types of classes.The experiment proved that the LPP can better find the internal nonlinearcharacteristic of hyperspectral remote sensing data and achieve the good effect of datacompression. The effect of LDA-lPP is only inferior to the LPP algorithm, but itcompressed the more information into the front few bands, tha is to say, it has highercompression efficiency.