Spatial discrimination based on the ground truth with mixed categories

The authors examined discriminant analysis of land-cover categories based on multivariate Gaussian fields. A parameter-estimation method by the ground-truth data consisting of pure and mixed cells with known category-proportions was discussed, and test data were discriminated by a penalized likelihood. With actual Landsat data, their procedure showed an excellent performance with respect to the linear discriminant function and Switzer´s method. Further, the use of mixels in the ground-truth data improved the discrimination results significantly