Kernel-Based Nonparametric Fisher Classifier for Hyperspectral Remote Sensing Imagery

Hyperspectral Imagery Sensing (HIS) is widely gained tremendous popularity in many research areas such as remotely sensed satellite imaging and aerial reconnaissance. HIS is an important technique with the measurement, analysis, and interpretation of spectra acquired sensing scene an airborne or satellite sensor. The development of sensor technology brought the developing of collecting image data using hyperspectral instruments with hundreds of contiguous spectral channels. The preprocessing of hyperspectral sensing data is a feasible way through machine learning technology. Among these machine learning methods, kernel learning is a feasible nonlinear feature extraction on hyperspectral sensing data. The nonlinear problems are solved with kernel function, and system performances such as recognition accuracy, prediction accuracy are largely increased. In this paper, we present a novel Kernel-Based Nonparametric Fisher Classifier (KNFC) for hyperspectral remote sensing imagery. Firstly, we have a comprehensive theoretical analysis on Nonparametric Discriminant Analysis. And NDA has its limitations on extracting the nonlinear features owing to the high nonlinear and complex distribution of the hyperspectral imagery data. In order to improve the limitation of NDA on hyperspectral remote sensing imagery, we introduce the kernel trick to NDA to develop Kernel-Based Nonparametric Fisher Classifier to enhance its ability on hyperspectral Imagery Sensing data. The feasibility of the KNFC is testified on the ORL and YALE databases, and then some experiments are implemented on two real data sets including Indian Pines and Washington, D.C. Mall, with various spectral and spatial resolutions reflecting different environments of remote sensing.