Study Of A Nonlinear Mixing Spectral Model And Some Preblems Of Vegetation Classification Using Hyperspectral Remote Sensing

This paper focuses on the establishment of a nonlinear mixing spectral model and its application to the vegetation classification using hyperspectral remote sensing, and discusses the application of vegetation classification methods of hyperspectral images, the selection of training samples and unmixing of mixed pixels. First of all, for a variety of vegetation classification methods applied to hyperspetral remote sensing, especially for the classification methods based on spectral features, the paper performs classification trainings and compares their performance, finds application laws of these methods used in vegetation classification using hyperspectral remote sensing, and gets hold of the skills of selecting optimum methods in vegetation classification to contribute to distinguishing and classifying vegetation types in hyperspectral remote sensing, then presents a classification method combining the traditional classification method and the classification method based on spectral features on the basis of these laws. So the problems of vegetation classification using hyperspectral remote sensing are primarily quested and a certain basis is formed for the study of unmixing of mixed pixels. For the selection of training samples, the study investigates and analyzes the sources of reference spectra and the selection methods of training samples in common use, gives two new selection (purity) methods of reference spectra, and testifies the validity of these methods using specific hyperspectral remote sensing data. For the unmixing of mixed pixels, the paper describes its concept and several unmixing methods available, presents a new nonlinear mixing spectral model and its calculating methods, then, taking an OMISI image as an example, performs tests to compare this nonlinear model and the original linear model applying to classification of mixed pixels in vegetation classification using hyperspectral remote sensing and testify the practicability and validity of the new model.