Improving subpixel classification by incorporating prior information in linear mixture models

This paper introduces a new subpixel classification algorithm that incorporates prior information from known class proportions in the linear mixture model. The prior information is expressed in terms of the occurrence probabilities of each land-cover class in a pixel. The use of different error cost functions that measure the similarity between the model-derived mixed spectra and the observed spectra is also investigated. Under these assumptions, the maximum a posteriori (MAP) methodology is employed for optimization. Finally, optimization problems under the MAP criteria for different error cost functions are formulated and solved. Our numerical results illustrate that the performance of the subpixel classification algorithm can be significantly improved by incorporating prior information from the known class proportions. Furthermore, there are marginal differences in accuracy when different error cost functions are used.