Comparison of linear least squares unmixing methods and Gaussian maximum likelihood classification

The study addresses the lack of controlled linear least squares unmixing studies comparing with standard per-pixel classification techniques at the same resolution. A general forward model for scene reflectance is specified and a constrained total least squares solution is described. This solution encompasses previous least squares unmixing methods. The Bayes risk for per-pixel classification is developed to indicate when classification may be inaccurate. Experiments comparing the accuracy of per-pixel classification and least squares unmixing suggest that unmixing is significantly more accurate when complete mixing occurs in the scene