Abstract: Hyperspectral images does exist redundant information, which brings certain side-effects on image classification. In this study, two dimensionality reduction algorithms, the CB method (CfsSubsetEval evaluator combines Best-First search strategies) and the PCA, and four multi-feature fusion combinations are proposed to construct eight schemes. The eight schemes combining with RF(random forest) classifier are then applied to classily hyperspectral images, and the best scheme for hyperspectral image classification are selected on the bases of the classification accuracy and Kappa coefficient. The results show that:①Multi-feature fusion can improve the classification accuracy of hyperspectral images, the classification accuracy of the hyperspectral image increases with considering geographic characteristics, texture characteristics, and exponential features gradually both in the two dimensionality reduction algorithms.②Considering the two dimensionality reduction algorithms, the classification accuracy based on CB reduction is generally higher than that of PCA dimensionality reduction. In terms of the classification accuracy based on eight schemes, the CB method with spectrum information, geographic characteristics, texture characteristics, and exponential features has the highest classification accuracy with 98.01% of overall classification accuracy, and 0.969 9 of Kappa coefficient.

Key words: hyperspectral image, image classification, dimensionality reduction, feature fusion, random forest