Classification of hyperdimensional data based on feature and decision fusion approaches using projection pursuit, majority voting, and neural networks

Hyperspectral sensors provide a large amount of data. The inherent characteristics of hyperspectral feature space still require the development of information extraction algorithms with a high degree of accuracy. Data fusion techniques can enable one to analyze high-dimensional data that is provided by hyperspectral sensors. There are two levels of fusion that will be discussed in the present paper: feature fusion and decision fusion. Feature fusion is a projection from one feature vector space (spectral bands) to another. An example of this is multispectral data feature extraction. In decision fusion, a local discrimination is performed at each sensor. Then the set of decisions is combined in a decision fusion center. This center has a set of algorithms to integrate the individual and local decisions of each sensor. The algorithms are based on different techniques such as majority voting, max rule, min rule, average rule and neural network. Experiments show that feature and decision fusion schemes enhance the classification accuracy of hyperspectral data