On the estimation of spatial-spectral mixing with classifier likelihood functions

Conventional, hard classification algorithms that decide one class per pixel ignore the fact that many pixels in a remote sensing image represent a spatial average of spectral signatures from two or more surface categories. The mixing of signatures arises from the intrinsic, spatially-mixed nature of most natural land cover categories, the physical continuum that may exist between discrete category labels, resampling for geometric rectification, and by the spatial integration defined by the sensor's point spread function. By allowing for multiple classes per pixel, each with a relative membership likelihood, soft classification algorithms have the potential to ''unmix'' the pixel data into the proportions of individual components. The potential and limitations of this approach are explored in this paper by empirical examples and analyses. A major conclusion is that the use of likelihood functions as estimators of mixing is valid for classes with high spectral signature separability, but problematic otherwise.