Quaternion-Based Multiscale Analysis for Feature Extraction of Hyperspectral Images

Abstract

This paper proposes a new method called multiscale quaternion Weber local descriptor histogram (MQWLDH) for feature extraction of hyperspectral images (HSIs), which is used to model spatial information based on the corresponding spectral features. The proposed method first transforms spectral data into an orthogonal space using principal component analysis, and extracts the first three principal components (PCs) based on the maximum variance theory. Then, construct the MQWLDH to extract spatial features based on those first three PCs. The proposed method uses the algebraic structure of quaternions to unify the process of processing the first three PCs, which reduces the computational cost and the dimensionality of the extracted spatial feature vector. Moreover, the constructed quaternion Weber local descriptor effectively characterizes the variations of each pixel neighborhood and detects the edges of HSIs. To capture more intrinsic spatial information contained in homogeneous regions of different sizes and shapes, multiscale feature histograms are constructed. Finally, a feature fusion framework is proposed to fuse spectral and spatial features so that spectral information can be fully utilized. The experimental results on three HSI datasets demonstrate that the proposed method provides effective features to different classifiers and achieves excellent classification performance.