Feature generation of hyperspectral images for fuzzy support vector machine classification

Feature generation (i.e. feature selection/extraction) for hyperspectral images in pattern classification is one of the key elements in improving the accuracy of classifier. Nonetheless, most existing techniques encounter difficulties in extracting essential features of non-stationary and/or non-linear signal, let alone hyperspectral images. Here, an alternative technique motivated by bi-dimensional empirical mode decomposition (BEMD) is presented. By virtue of BEMD, the given signal is adaptively decomposed into a series of bi-dimensional intrinsic mode functions (BIMFs) with different oscillations. Furthermore, those BIMFs are integrated into new features of the original signal. Additionally, the recently developed fuzzy support vector machine (FSVM) is exhibited to classify those features so as to reduce effects of outliers or noises. Experimental results on the widely used 92AV3C hyperspectral dataset demonstrate the efficiency of the proposed approach.