Analysis of thin snow spectral characteristic and retrieval algorithm construction of the fractional snow cover in Qilian Binggou Basin

The terrain in Qilian Binggou Basin is relatively complex, the snow depth is generally thin and snow distribution is fragmentized. Aiming at the problem of poor monitoring accuracy of MODIS snow products in this area, This study explored the influence of spectral characteristic of thin snow on the retrieval accuracy of MODIS data based on thin snow spectral characteristic analysis and combined with field survey experience; Then three MODIS fractional snow cover retrieval models are constructed through linear regression, linear mixed pixel unmixing and artificial neural network, and the snow map retrieved from Landsat 8 OLI image is taken as the ground truth to validate the three models’ accuracy respectively. The results show that: 1) The spectral reflectance of thin snow almost has no effect on MODIS snow retrieval accuracy based on the NDSI threshold method in the area. The poor accuracy of MODIS retrieval from thin snow is mainly due to the fragmentation of snow distribution caused by the complex terrain in the area, that is, the existence of a quantity of mixed pixels. 2) The best input parameters combination of BP artificial neural network model for MODIS fractional snow cover retrieval is(ρ1~ρ7)+NDSI+DEM. 3) The linear mixed pixel unmixing model has the lowest accuracy and the BP artificial neural network model has the best accuracy in terms of snow cover extraction in the study area. 4) Multi-factor model (BP artificial neural network) has better accuracy and stability of snow coverage extraction compared with the single factor model (unary linear regression model) in complex terrain, it is an ideal method for the retrieval of fractional snow cover in the study area.