Hyperspectral image classification based on multi-feature fusion and different dimensionality reduction methods

Abstract

The rich band information of hyperspectral images provides data support for accurate classification, but the information redundancy also brings bad effects to image classification. In order to improve the classification accuracy, this paper takes Liutang Town of Guilin City, Guangxi Province as the research area, and "OVS-1A/B" hyperspectral image as the data source. Based on two different dimensionality reduction methods, Principal Components Analysis (PCA) and CfsSubsetEval GreedyStepwise (CG), combined with Support Vector Machine (SVM), classifiers are used to classify hyperspectral images based on four multi-feature fusion schemes, and their classification accuracy is compared. The results show that: (1) In the two classification schemes with different dimensionality reduction methods, the classification accuracy of ground objects increases continuously with the increase of the types of fused features. (2) In the same combination scheme, the classification accuracy obtained by CG feature selection is higher than that obtained by PCA feature extraction. (3) The classification accuracy of scheme 4 using CG method for feature selection was the highest, the overall classification accuracy was 97.12%, and the Kappa coefficient was 0.9556.