Estimating alpine snow cover with unsupervised spectral unmixing

An unsupervised algorithm estimates snow-covered area at subpixel resolution from multispectral image data. Classification trees fragment the data set along boundaries of distinct land and cloud cover classes. The dimensionality and number of endmembers for each image fragment are determined from principal components analysis. Each fragment is unmixed with all endmember sets on its convex hull, and the best set is selected. Endmember spectra are converted to surface reflectance with an atmospheric radiative transfer code, and the endmembers are identified by automated search of a spectral library. The final snow cover estimate is a composite of the best mixture model per pixel, adjusted for endmember impurity. The algorithm is tested on Landsat Thematic Mapper data against high resolution aerial photographs