Obtaining reference spectrums for hyperspectral matching using elitist non-dominated sorting genetic algorithm

Matching approaches, as characteristic hyperspectral classification methods, have been utilized more and more frequently in many relevant fields. To avoid complicated spectral calibration and correction, obtaining reference spectrums from remote sensing image is often adopted. The commonly used way is to calculate mean spectrum of a certain class after collecting a training set for it. However, mean spectrum is just a statistical descriptor and can not guarantee high matching accuracy. In this presentation, a new intelligent method of obtaining reference spectrums from image is put forward. Starting from the assumption that every entity in training set can become reference spectrum, we convert the task into a Multi-Objective optimization problem. Then elitist non-dominated sorting genetic algorithm (NSGA-II), analytical hierarchical process (AHP), and fuzzy evaluation are implemented step by step to finally get the reference spectrums through selecting entities from training sets. Experiment results indicate that the reference spectrums obtained by this new method are superior to mean spectrums and average improvement of matching accuracy is 6.04%similar to 8.15% in the case of two-class separation. When the new method is extended to solve multi-class separation using one vs. one approach, accuracy enhancement is as large as 33.52%similar to 54.83%.