Hyperspectral image classification based on Gabor features and decision fusion

Traditional methods for hyperspectral image classification typically use raw spectral signatures without considering spatial characteristics. In this work, a classification algorithm based on Gabor features and decision fusion is proposed. First, the adjacent and high correlated spectral bands are intelligently grouped by coefficient correlation matrix. Following that, Gabor features in each group are extracted in PCA-projected subspaces to quantify local orientation and scale characteristics. Afterwards, locality-preserving non-negative matrix factorization is incorporated to reduce the dimensionalities of these feature subspaces. Finally, the classification results from Gaussian-mixture-model classifiers are merged by a decision fusion rule. Experimental results show that the proposed algorithms substantially outperforms the traditional and state-of-the-art methods.