Dimensionality Reduction of Hyperspectral Images Using Empirical Mode Decompositions and Wavelets

This paper presents a novel method for dimensionality reduction of hyperspectral images. It combines Empirical Mode Decomposition (EMD) with wavelets in order to generate the smallest set of features that leads to the best classification accuracy. The introduced method exploits both spatial and spectral information of the image which leads to more and better class separability and hence to better classification accuracy. Specifically, the 2D-EMD is applied to each hyperspectral band to enhance spatial information and then 1D-DWT is applied to each EMD feature´s signatures to enhance spectral information. The reduced Wavelet-based Intrinsic Mode Function Features (WIMF) are obtained by selecting coefficients. Then, new features are generated by summing up the lower order WIMF features. Support Vector Machine (SVM) based classification is performed with two different hyperspectral data sets, namely AVIRIS Indian Pine and ROSIS Pavia. The experimental results show that features obtained by the proposed method significantly increase the classification accuracy compared to features obtained by other frequency based direct dimensionality reduction methods.