Estimating the spatial distribution of snow water equivalent in an alpine basin using binary regression tree models: the impact of digital elevation data and independent variable selection

Regression tree models have been shown to provide the most accurate estimates of distributed snow water equivalent (SWE) when intensive field observations are available. This work presents a comparison of regression tree models using different source digital elevation models (DEMs) and different combinations of independent variables. Different residual interpolation techniques are also compared. The analysis was performed in the 19Ð1 km2 Tokopah Basin, located in the southern Sierra Nevada of California. Snow depth, the dependent variable of the statistical models, was derived from three snow surveys (April, May and June 1997), with an average of 328 depth measurements per survey. Estimates of distributed SWE were derived from the product of the snow depth surfaces, the average snow density (54 measurements on average) and the fractional snow covered area (obtained from the Landsat Thematic Mapper and the Airborne Visible/Infrared Imaging Spectrometer). Independent variables derived from the standard US Geological Survey DEM yielded the lowest overall model deviance and lowest error in snow depth prediction. Simulations using
the Shuttle Radar Topography Mission DEM and the National Elevation Dataset DEM were improved when northness was substituted for solar radiation in five of six cases. Co-kriging with maximum upwind slope and elevation proved to be the best method for distributing residuals for April and June, respectively. Inverse distance weighting was the best residual distribution method for May. Copyright 2004 John Wiley & Sons, Ltd.