Snow cover area is a very critical parameter for hydrologic cycle of the Earth. Furthermore, it will be a key factor for the
effect of the climate change. Most research on estimating snow cover area is binary: pixels are verified either "snow" or
"not snow". Most pixels, however, are mixed with snow, vegetation, soil, rock or water. This paper presents a spectral
unmixing to estimate sub pixel snow cover. Firstly, a manmade selection for endmember was set up based on PCA method.
Then an automatic selection of snow endmember and nonsnow endmember based on NDSI and NDVI can be achieved.
The algorithm was tested on several different MODIS scenes in Tibetan Plateau. The efficiency and precision of
classification equals that obtainable from the PCA method but is faster, cheaper. Lastly, Two sub pixel snow cover mapping
means (regression method based on NDSI and spectral unmixing method based on the endmember automatic selection)
was compared and analysised. And it takes the ASTER 15m data as ground true data to calculate the percentage of snow
cover for 500m cells. It shows that the spectral unmixing can map fractional snow cover more precision and the automatic
selection mean is stable and robost.