Improving the accuracy of linear pixel unmixing via appropriate endmember dimensionality reduction

Spectral unmixing is a quantitative analysis procedure used to recognize constituent ground cover materials (or endmembers) and obtain their mixing proportions (or abundances) from a mixed pixel. The endmember abundances may be estimated using the least squares estimation (LSE) method based on the linear mixture model. This paper investigates the use of spectral dimensionality reduction as a preprocessing tool for hyperspectral linear unmixing. Four dimensionality reduction methods are investigated and compared; these include methods based on the discrete wavelet transform (DWT), discrete cosine transform, principal component transform, and linear discriminant transform (LDT). Three sets of experiments are designed and implemented for evaluating the effects of the dimensionality reduction techniques on the LSE of endmember abundances. Experimental results show that the use of the DWT and LDT-based features extracted from the original hyperspectral signals can greatly improve the abundance estimation of endmembers. On average with these methods, the root-mean-square of the abundance estimation error is reduced by 20%.