Comparison of Curve Fitting Method for Hyperspectral Data Classification with Nonlinear Based Feature Extraction Methods

摘要

Hyperspectral imagery is one of the most important tools in remote sensing. Increasing the number of bands, lack of training samples, correlation of spectral samples and redundancy of data make conventional image classification methods without reducing the dimensions of the feature vector, not applicable. Dimension reducing as a preprocessing step could be done in two approach: feature selection, and feature extraction. In this paper, recently rational function curve fitting-feature extraction (RFCF-FE) method is analyzed and compared with some nonlinear non-kernel based feature extraction methods such as locally linear embedding (LLE), piecewise constant function approximations (PCFA) and the proposed feature extraction Based on breakpoints (BPB). The maximum likelihood (ML) classification results demonstrate that RFCF-FE provides better classification accuracies compared to competing methods. Also, in this paper, we propose a method for determining the separability of data classes based on specific breakpoints of the spectral response curve (SRC) of pixels.