Introducing a Method for Classification of AVIRIS Hyperspectral Data, using Feature Extraction and Classifier Ensemble Methods

One of the most applied supervised classification method is Maximum Likelihood (ML) in which a series of statistical parameters such as variance-covariance matrices are estimated. In Hyperspectral remote sensing images, due to the limited number of training samples and their high spectral dimensions, the probability of having singular matrices and/or reduction the accuracy of classification is plausible. To solve this problem, different approaches such as reduction of number of features or ensembling of classifiers can be used. In theory, the acquisition of large number of training data set is feasible, but it is very time consuming. Then in practice there are always some limitations where we believe methods such as Feature Extraction algorithms or ensemble of classifiers for dimensionality reduction can solve that. In this research, Nonparametric Weighted Feature Extraction (NWFE), as well as classifier ensembles, are used simultaneously. For constructing multiple classifiers, manipulation of input features are done in which input feature space is divided into multiple subspaces by using class based NWFE method. A ML classifier is applied on each of the prepared feature subsets, and finally a combination scheme was used to combine the outputs of each individual classifier. In order to fuse multiple classifiers, a measurement level method is suggested using the mean rule. The results show an overall accuracy of 85.67% for NWFE method and 89.26% for classifier ensemble. For method suggested in this paper the overall classification accuracy of 89.34% was achieved. The results indicate significant improvement in classification accuracy, compared to the two methods on which this method is based upon. Despite the closeness of these two accuracies, because of less complexities and feasibility of parallel calculation, the suggested method is preferred.