Hyperspectral Image Classification Based on Segmented Local Binary Patterns

Recently, local binary patterns (LBP) coupled with principal component analysis has been developed for feature extraction of hyperspectral imagery, which has shown success over traditional methods but is limited in physical meaning representation due to the noise bands existing in hyperspectral data. In order to preserve the intrinsic geometrical structure of original data, we propose a segmented LBP (SLBP) to group correlative bands and then extract spatial-spectral features from each band group. The proposed approach employs the LBP operator on independent subspaces to characterize local texture information and distinct spectral signatures, along with a decision fusion system further improving discriminant power. The proposed approach is compared with several traditional and state-of-the-art methods on two benchmark datasets (i.e., the Indian Pines dataset and the Salinas Valley dataset). Experimental results demonstrate that the proposed SLBP strategy can yield superior classification performance (96.8% for the Indian Pines dataset with an improvement of approximately 6.4% and 4.2% when compared with LBP and MELBP, respectively; 98.1% for the Salinas Valley dataset with an improvement of approximately 3.5% and 1.3% compared with LBP and MELBP, respectively).