Non-linear mixture model and application for enhanced resolution multispectral classification

A new mixture model based on a probabilistic formulation is described. The model is non-linear, and unlike previous mixture models, does not need to be inverted to compute mixtures from spectra. Instead, mixtures are estimated using an interpolation approach. There are no constraints between the number of bands and endmembers. The mixture model is used to enhance the spatial resolution of classification images-increasing the classification accuracy by about 20% in the example presented. The enhanced resolution classifier is better able to detect features near the resolution of the sensor, to identify subtle patterns not fully resolved in the original data, and to more precisely delineate the boundaries of areal features