Spectral Unmixing for the Classification of Hyperspectral Images

Spectral mixing is inherent in any finite-resolution digital imagery of a heterogeneous surface, so that mixed pixels are inevitably created when multispectral images are scanned. Solving the spectral mixture problem is, therefore, involved in image classification, referring to the techniques of spectral unmixing. The invention of imaging spectrometers especially promotes the potential of applying spectral unmixing for sub-pixel classification. This paper investigates two spectral unmixing techniques: the least squares (LS) unmixing and the matched filter (MF) unmixing. Experiments with a set of AVIRIS data were carried out to evaluate the performance of spectral unmixing. The MF unmixing method proved itself to be an effective technique in classifying a hyperspectral image by showing a 90% classification accuracy. Whereas, the LS unmixing technique did not show promising results, when it was applied to the original bands of the test image. The maximum noise fraction (MNF) transformation, however, is found to be helpful to promote the performance of the LS unmixing. Applying the LS unmixing to the MNF transformed images can improve the classification accuracy for about 20%.