Distribution of Snow in the Upper Marble Fork Basin, California
Al Leydecker1 and James O. Sickman1
We have measured snow-water-equivalent (SWE) of the spring snowpack in the Upper Marble Fork
basin (1900 ha) on the western slope of the Sierra Nevada since 1993. Over a thousand snow-depth
measurements were taken at maximum accumulation each year in grid patterns, or along traverses;
sampling points for the earlier surveys were spatially located by mapping or with corrected and
noncorrected gps coordinates; since 1994 corrected gps coordinates were used throughout. Two to
five sampling points, usually spaced 610 m apart, were used to estimate the mean depth over a
DEM pixel (typically 30 30 m). We were unable to spatially distribute snow density due to the
difficulty of accumulating sufficient measurements in the deep (26 m) snowpack, and our analysis
of snow distribution is based solely on depth. This overstates any correlation between SWE and
terrain because the deeper snow is typically less dense; however, the error is minor because density
varies within a narrower range than snow depth (coefficient of variation of ∼0.08 for density vs.
∼0.39 for snow depth).
Snow-depths in the basin, the headwater catchments, and along individual traverses were usually
normally distributed, and aside from broad generalities of deeper snow below avalanching slopes
and shallower snow at lower elevations and on slopes subject to winter melt, snow depth in the basin
was predominately random; net winter solar radiation and other terrain-based parameters, e.g., slope,
aspect, and elevation, typically explained less than 15% of the variability. Only for the steep topog-
raphy of the Emerald (120 ha) and Pear (136 ha) glacial cirques could a linear regression model be
a major dichotomy between the northern south-facing portion and the remainder of the catchment;
within these two subregions depth was more or less random. Regression tree models were inferior to
linear regression; although generating high r2 values (> 0.6), tree models consistently failed cross-
validation tests, i.e., trees formed with randomly chosen subsets of 75% of the data were used to
model snow depth at nonselected points, which were then correlated with actual measurements (typ-
ically, r2 < 0.15). Kriging performed poorly as a method of distributing snow depth because vario-
grams, derived from numerous traverses over the years, typically had an auto-correlation range of 30
to 40 m, i.e., rarely extending beyond the adjacent pixel. Contour mapping of snow depth with
various methods, e.g., linear kriging, variogram kriging, minimum curvature, produced cross-vali-
dation correlation coefficients (r) of < 0.15. In contrast to the poor correlation of snow depth with
terrain, there was a consistent year-to-year relationship between the mean basin snow depth and
mean depths in the subbasins, e.g., the Emerald catchment had a mean depth equal to 131 1% (
SE) of the overall mean, which lends promise to snow distribution modeling based on stratified
sampling of the major subregions and random redistribution based on the mean and standard devia-
tion.
1Institute
for Computational Earth System Science, University of California, Santa Barbara, California
93106, USA
103
Previous Page
