Decision fusion for hyperspectral image classification based on multiple features and locality-preserving analysis

A novel fusion-classification system is proposed for hyperspectral image classification. Firstly, spectral derivatives are used to capture salient spectral features for different land-cover classes and a Gabor filter is applied to extract useful spatial features at neighbouring locations. Then, two locality-preserving dimensionality reduction methods are employed to reduce the dimensionality of data and preserve the local structure of neighbouring samples in the original image, derivative-feature and Gabor-feature domains. Finally, the classification results from Gaussian-mixture-model classifiers are fused by a decision-fusion approach. We have compared the proposed system to several traditional and state-of-the-art methods on two benchmark classification data sets. In both cases, our system achieved improved accuracy than the current best performing methods. Especially in the case of classification for the Indian Pines data set, the proposed DG-locality-preserving nonnegative matrix factorization (LPNMF) has 7% higher accuracy than support vector machine-Markov random field and LPNMF-G-MRF, and also has about 4% higher accuracy than Gabor-LPNMF. The proposed method has practical relevance as it shows potential to create smooth classification maps even for complex, spatially heterogeneous land-cover classes.