Wavelet-Based Feature Extraction for Improved Endmember Abundance Estimation in Linear Unmixing of Hyperspectral Signals

Abstract

This paper shows that the use of appropriate features, such as discrete wavelet transform (DWT)-based features, can improve the least squares estimation of endmember abundances using remotely sensed hyperspectral signals. On average, the abundance estimation deviation is reduced by 30% to 50% when using the DWT-based features, as compared to the use of original hyperspectral signals or conventional principal component analysis (PCA)-based features. Theoretical analyses further reveal that the increase of endmember separability is a fundamental reason leading to this improvement. In addition, the robustness of the DWT-based features is verified experimentally. Finally, the idea is generalized as a point that the remote sensing community needs to investigate feature extraction (or dimensionality reduction) methods that are based on signal classification, such as the DWT approach, for linear unmixing problems, rather than using feature extraction methods that are based on signal representation, such as the conventional PCA approach.