Subspace-based Support Vector Machines For Hyperspectral Image Classification Technology

The development of hyperspectral remote sensing makes the remote sensing imageclassification study be again on a new step, and the hyperspectral remote sensing hasbecome the frontier in the field of remote sensing technology of which the hyperspectralimage classification technology has become the research hotspot. Numbers ofcharacteristics of hyperspectral remote sensing ensure the hyperspectral image beingapplied to the terrain classification of high accuracies and effectiveness but some of whichmake it has to face obstacles in the hyperspectral classification process, such as “Curse ofdimensionality”, noise and limited training samples. Currently, many scholars, at home orabroad, are trying to solve the difficulty and contradiction in the hyperspectralclassification, and some of them have made all aspects of valuable progress.This paper introduces a new supervised classification algorithm for remotely sensedhyperspectral image data based on the statement above. Support vector machine was firstlyproposed by Cortes and Vapnik in1995which has shown many unique advantages in thesmall sample, nonlinear situation and high dimensional pattern recognition. The subspaceprojection method was originally the product of linear feature extraction and datacompression whose function is to compress the vector data to the principal axis on whichthe energy concentrate to realize the mapping from high dimension space to the low one.This paper has combined the SVM with the subspace projection method and successfullyimplement the hyperspectral image data classification using the LIBSVM toolbox.Rigorous validation methods on the experimental results prove our method is scientific andfeasible, and has high practical values.