A method for manual endmember selection and spectral unmixing

The number of spectrally unique signatures needed to reproduce the statistically significant variance observed in multispectral and hyperspectral datasets can be estimated from the eigenvalues of a principal component analysis (PCA) of the data. In this article, we describe a multidimensional visualization method for interactively searching for a feasible set of spectral signatures in the space of the PCA eigenvectors that account for most of the variance. These spectral signatures, referred to as endmembers, are input to a linear mixture model which can be inverted to compute endmember abundances for each data spectrum. The visualization method discussed in this article and referred to as the manual endmember selection method (MESM) is based on Inselberg's (1985) parallel coordinate representation of multidimensional spaces. It is novel in the field of multidimensional visualization in that it includes not only a passive representation of higher-dimensional data but also the capability to interact with and move geometrical objects in more than 3 dimensions. The spectral shape of endmembers selected with the MESM may be influenced by processes such as multiple scatterings by surface materials and other factors such as illumination geometry that affect the signal received by the sensor. These processes and factors may produce significant errors in computed endmember abundances, tf not accounted for in the endmember reflectance (Roberts, 1991). The MESM is one method for obtaining endmembers that account for all factors and processes significantly affecting the spectral data.