a. Study site

The Green Lake 4 study area is part of the Niwot Ridge (NWT) Long-Term Ecological Research (LTER) project site located approximately 35 km northwest of Boulder, Colorado (Fig. 1). The Green Lake 4 basin flanks the eastern side of the Continental Divide, functioning as a headwater catchment for North Boulder Creek, a primary water source for the city of Boulder. The research area is 2.25 km² in size, has an elevation range of 3560–4090 m, and is entirely above treeline. The terrain is steep and rugged, with an average slope of 28°. Seasonal snow accumulation typically begins in October. Mean monthly minimum temperatures are below 0°C from October through May, with mean monthly maxima below freezing from November to April (Berg 1986). The continental location and high altitude combine to generate low-density snowfall events and atmospheric conditions that retard thermal densification of the surface snow, creating an environment where snow transport is rarely limited by snow cover surface conditions (Berg 1986). The basin's characteristic high winds, steep topography, and continental snow climate produce an extremely heterogeneous snow cover governed largely by wind redistribution.

b. Snow data

The snow survey data were collected, prepared, and made available for this study by the National Science Foundation (NSF) supported NWT LTER project and the University of Colorado Mountain Research Station (MRS). The snow survey was conducted as part of an annual assessment of snow distribution in the study basin and was not designed specifically for the study described herein. Depth samples were taken at 504 sites in mid- to late May 1999 (Fig. 1). The locations of the sample sites were recorded and differentially corrected using a global positioning system (GPS). Whenever possible, measurements were taken at approximately 50-m intervals. Steep terrain and avalanche danger, however, limited data collection to safe and accessible areas. The 1 May estimate of snowpack conditions in the Boulder Creek drainage was 112% of average (NRCS 1999).

The modeled dependent variable in this study was snow depth. Snow-water-equivalence (SWE), the state variable of primary interest to researchers and forecasters, is obtained from the product of snow depth and snow density. Snow density is substantially more conservative than snow depth, and the variability of snow density continues to decrease as the snowpack ripens into the late spring (Elder et al. 1991). We have used snow depth as a surrogate for SWE and have assumed that distributed SWE would follow the pattern of distributed depths in this mid- to late May modeling time frame.

c. Wind data

Wind data were likewise made available by the NSF-supported NWTLTER project and the MRS. Wind measurements were made at the D-1 meteorological station located on Niwot Ridge (3744 m), an exposed east–west-trending rise in the northeast corner of the Green Lake 4 basin (Fig. 1). The wind instrumentation is located approximately 9 m above the ground surface in an exposed location. Hence, there is little snow accumulation and little variability in the instrument height through the year. The beginning of the snow accumulation season in October and relatively low May wind speeds, combined with the presence of moist spring snow, typically limit significant blowing snow events to a 6.5-month period from mid-October to early May (Berg 1986). Thus, the wind data gathered for this study were restricted to the 15 October–7 May time period.

The wind direction data consisted of the daily resultant means of instantaneous recordings taken at 30-s intervals for the accumulation season. Unfortunately, wind direction data from the D-1 site were unavailable for approximately 1 month (14 February 1999–14 March 1999) during the accumulation season prior to the survey dates. Based on historic data, wind patterns during the mid-February to mid-March period are consistent with other wintertime observations. Figure 2 illustrates the characteristic wintertime dominance of western flow at the D-1 site during the study period. The mean resultant wind direction for the study period was 265°.

A method of monitoring blowing snow events was not available during the study period. A comparison with wind speeds registered during blowing snow events recorded with an elevated snow particle counter as part of Berg's (1986) study provided insight into the degree of wind-induced snow redistribution during the 1998/99 season. At a meteorological station located on Niwot Ridge, 3 km east of the D-1 site, 53% of the hours monitored over the course of two winters (1973/74 and 1974/75) recorded blowing snow events (Berg 1986). Using wind data taken from a similar instrument height (8.75 m), Berg found that blowing snow occurred during 22% of the hourly averaged wind speeds in the 7–8 m s^–1 category, during 49% of 13–15 m s^–1 winds, and during 75% of winds of 29–30 m s^–1. Over the course of the 1998/99 blowing snow season, 71% of the days had a mean wind speed greater than 7.5 m s^–1, 29% had a mean wind speed greater than 14 m s^–1, and on only two days did a recorded wind speed fail to exceed 7.5 m s^–1. These data attest to the fact that wind-induced snow redistribution occurs on an ongoing basis throughout the winter in this environment.

a. Redistribution parameters

The terrain-based redistribution parameters presented here quantify different aspects of the upwind terrain. Hence, the upwind terrain needed to be defined. In this study, the upwind terrain consisted of a pie-shaped area centered on the seasonal prevailing wind direction (Fig. 3). Based on the wind data gathered at the D-1 meteorological site, an upwind window of analysis centered on a prevailing wind direction of 265° was chosen. Extending upwind of each pixel, the window was laterally bound by two azimuths separated by an arbitrarily chosen width of 60° (A₁ = 235°, A₂ = 295°). The lateral width of the window allows for inclusion of terrain that is not directly upwind of the cell based on the D-1 observations. While this approach is somewhat robust for wind directions that vary from the seasonal mean at D-1, it still contains a natural bias to that mean. Calculated shelter/exposure indices at sites where the prevailing wind direction was systematically different (e.g., due to topographic deflection) may contain some error. Specification of a search distance (dmax) controlled the extent of upwind terrain analyzed and was varied to assess the effects of near and far terrain upon snow accumulation.

The objective of the maximum upwind slope (Sx) parameter was to quantify the extent of shelter or exposure provided by the terrain upwind of each pixel. To determine the value of the Sx parameter, we examined all cells along a vector emanating from the cell of interest, determined the cell that had the greatest upward slope relative to the cell of interest, and returned the slope between this shelter-defining cell and the cell of interest:

View Expanded

where A is the azimuth of the search direction, (x_i, y_i) are the coordinates of the cell of interest, and (x_υ, y_υ) are the set of all cell coordinates located along the search vector defined by (x_i, y_i), A, and dmax. The selection of a maximum shelter-producing pixel based on slope is analogous to the determination of solar shading within the horizon function used in radiation modeling (Dozier et al. 1981). Negative Sx values indicated exposure relative to the shelter-defining pixel (i.e., the cell of interest was higher than the shelter-defining pixel). Examples of Sx determinations are provided in Fig. 4. Theoretically, increasingly negative Sx values correspond to greater constrictions on the approaching wind flow, leading to increasing wind speeds, while increasingly positive Sx values correspond to larger and more proximal landscape obstacles, producing a greater degree of shelter and lower wind speeds. Values of Sx were determined along vectors at 5° increments within the upwind window and averaged to formulate the mean maximum upwind slope parameter, Sx, defined by

View Expanded

where n_υ is the number of search vectors in the window defined by A₁, A₂, and dmax. Search distances of 50 m (Sx₅₀), 100 m (Sx₁₀₀), 300 m (Sx₃₀₀), 500 m (Sx₅₀₀), 1000 m (Sx₁₀₀₀), and 2000 m (Sx₂₀₀₀) were used to test the influence of near and far terrain.

It was felt that Sx calculated for within-basin sites would have been only slightly affected by the 10-m DEM's restricted view of the western topography, given the considerable western obstruction presented by the Continental Divide. Hence, Sx was calculated from the 10-m DEM.

The Sb parameter was designed to delineate topographic features capable of generating flow separation. To obtain Sb, independent representations of near and far topographic features were necessary. The difference between relative slope measures of the local and outlying topography supplied a measure of the break in upwind slope. To determine Sb, two applications of the Sx algorithm were used to determine a local Sx (Sx₁) and an outlying Sx (Sx_o) A user-defined separation distance (sepdist) determined the cutoff between the two regional Sx calculations, with Sx was determined along an upwind search vector with a maximum search distance (dmax) equal to the separation distance (sepdist) to determine Sx₁, defined as

_1Ax_iy_iA,sepdistx_iy_i(3)

The distant or outlaying terrain described by Sx₀ was defined as the Sx_A,1000 of the cell located the separation distance upwind of the cell of interest along each respective search vector:

_oAx_iy_iA,1000x_oy_o(4)

where [(x_o − x_i)² + (y_o − y_i)²]^0.5 ≅ sepdist and (x_o, y_o) ∈ (x_υ, y_υ). Subtraction of Sx_o from Sx₁ yielded Sb. The average of each vector calculation of Sb within the upwind window formed a mean measure of upwind slope break, Sb, defined by

View Expanded

In order to determine Sx_o for many pixels located near the Continental Divide it was necessary to analyze terrain beyond the western extent of the 10-m DEM. Therefore, Sx_o was determined from the 30-m DEM, while calculation of Sx₁ remained at the 10-m scale to retain high-resolution depictions of proximal topography.

In addition, Sx_o was used to measure the relative upwind exposure of each potential slope break. With Sb providing a means of locating cells downwind of considerable slope breaks, Sx_o [calculated from Eq. (2) with Sx_o substituted for Sx] was independently evaluated to determine whether the described slope break was subject to the high winds and assumed correspondent high snow-transport fluxes needed to produce substantial downwind drifts. Based on threshold values of Sb and Sx_o, a drift delineator variable (D₀) was established whereby each sample was classified as either being within or outside of a modeled lee-slope drift zone. Applied in this manner, D₀ was a first approximation that broadly defined drift areas; intradrift accumulation differences were not explicitly represented.

b. Data analysis

We sought to quantify how these redistribution parameters (Sx and D₀) related to measured snow depths in two ways. First, we sought to answer the question, “Are Sx and D₀ significant predictors of snow depth?” while also examining the best length scales to use for calculating Sx and D₀. Linear regression and analysis of variance (ANOVA) were used to perform the aforementioned analyses. Second, we asked the question, “What effects do the addition of Sx and D₀ have on a multivariate model of snow distribution?” In response to this latter question, regression tree modeling was used to formulate and compare two tree models of snow distribution: one that incorporated the new parameters with solar radiation, elevation, and local slope data as predictors, and one based solely on solar radiation, elevation, and slope.

The first analysis step was to assess the performance of the redistribution parameters as snow depth predictors. The inspection of the mean maximum upwind slope (Sx) parameter included analysis at varying spatial scales. Standard linear analysis techniques were used to determine the strength of the relationships between the parameters and the sampled snow depths. Though linear models are oftentimes inappropriate for modeling snow distribution (Elder et al. 1995), they do provide a means of comparing the relative strengths of predictor–response relationships. Comparisons were also drawn between the variance in depth explained by Sx and that accounted for by elevation, net potential radiation, and slope to test for the relative importance of each of these variables in explaining the observed snow distribution. Raster grids were then produced to visually compare the distribution of Sx values with the distribution of sampled depths. The drift delineator (D₀) was applied as a binary factor or dummy variable alongside the Sx wind scalar parameter within a general linear model. F tests on the drift delineator variable would indicate whether observations with similar upwind shelter contained significantly more snow if they were in a modeled drift zone.

The second analysis step was to assess the effect that the addition of the redistribution parameters had on a model of snow distribution. Based on the aforementioned analyses, redistribution parameters were chosen for inclusion as predictor variables along with the commonly used independent variables of snow distribution modeling—elevation, radiation input, and slope—to form a “redistribution” regression tree model (Breiman et al. 1984) of the May 1999 snow distribution. A “nonredistribution” regression tree model, based on just the latter three predictors was also fit to the sample dataset. Quantitative and qualitative comparisons between the regression tree models trained on the two sets of explanatory variables were then conducted. Regression trees have been successfully used to model snow distribution (Elder et al. 1995, 1998; Balk and Elder 2000) and to determine subpixel resolution snow cover (Rosenthal and Dozier 1996).

The fitting of a regression tree model is a two-step process of growing an intentionally overfit tree with subsequent pruning to a statistically defensible size. In this application, the dependent variable was sampled snow depth, and the set of independent variables consisted of the coregistered predictor values at each sample point. The model is grown using a method of binary recursive partitioning whereby a dataset is successively split, based on values of the predictor variables, into increasingly homogeneous subsets of the dependent variable (Clark and Pregibon 1992). These subsets are termed nodes. At each split the mean of the sampled depths within the subset was the value assigned to that node. As the number of subsets (i.e., terminal nodes or tree size) increases, model R² values correspondingly increase. Plots of R² versus tree size provided the means to assess the comparative accuracies of the redistribution and nonredistribution models on a node-by-node basis.

Cross validation of tree models is one recommended method of determining the size of tree with the highest defensible predictive accuracy (Clark and Pregibon 1992; Venables and Ripley 1994). In this research, each snow depth sample, together with its coregistered explanatory variables, was randomly split into 10 mutually exclusive subsets for the cross-validation procedure. Nine of the subsets were combined to grow a tree and then the 10th subset was run through the tree with the model deviance recorded at each step through the tree. Cross-validation results for each of the 10 possible ways of holding a set out were accumulated and averaged, producing a tenfold cross validation (Clark and Pregibon 1992). A common feature of the cross-validation plot (model deviance versus tree size) is a lowering of model deviance (i.e., rising R²) with increasing tree size until a flattening out point is reached, followed by a subsequent increase in model deviance (i.e., decreasing R²) resulting from differences between the full dataset and the subsets used in the cross-validation procedure. The flat minimum of the model deviance plot is a suggested range for the optimal amount of terminal nodes in the final pruned tree (Clark and Pregibon 1992). Since the results of a single tenfold cross-validation run have a random element dependent on the initial splitting of the training data, successive cross-validation runs can produce inconsistent results (Venables and Ripley 1994). For a more robust estimation, the results of 100 tenfold cross-validation runs were averaged to produce the results reported here. Analysis of the cross-validation plots for the two models assessed the respective stability of each model.

The regression trees also provided the means to spatially distribute depth values throughout the basin. The spatial representations illustrated the effect that the redistribution parameters had on extrapolation of the information contained in the sampled data to the population of grid cells. Grids representing the spatial distribution of each of the predictor variables were passed through the respective tree models with the depth assigned to each cell determined by the value within the terminal node it was routed to. Regression trees essentially split the data into discrete classes of snow depth, the number of classes dependent on the amount of terminal nodes in the model, and the values contained within them. As applied in this study, the inherent structure of the regression trees served to substantially increase model scale relative to both the grid and process scales.

a. Redistribution parameters

All of the tested Sx parameters were significant predictors of snow depth (Table 1). The seemingly low, yet highly significant, linear R² values are reflective of the potential for complex nonlinear interaction of the controlling processes in high-elevation basins (Elder et al. 1991) and may also be due in part to scaling effects caused by differences in the model and process scales (Blöschl 1999). These factors were accentuated in this rugged, steep basin where sampled snow depths were observed to vary by an order of magnitude over very small distances.

The level of significance of the Sx parameter slightly increased as dmax increased from 50 to 100 m and then slowly decreased as dmax increased beyond 100 m. Based on maximum upwind slopes, the Sx parameter was only affected by the inclusion of additional upwind terrain if the distant terrain provided a higher degree of shading than the near terrain previously analyzed. Therefore, as dmax increased, a sample's Sx value could either increase or remain unchanged. For example, a sample located on a perfectly flat area with an abrupt obstacle located 1000 m upwind would have Sx values of 0 for all Sx with dmax < 1000 m, and Sx values greater than 0 for dmax ≥ 1000 m. Analysis of the coregistered Sx₁₀₀ data indicated that the samples most affected by increases in dmax were those with low Sx₁₀₀ values, so that areas with low proximal shelter tended to be classified with greater relative shelter as the extent of upwind terrain analyzed increased (Fig. 5). The increase in the level of significance of Sx as dmax increased from 50 to 100 m suggested that, based on this dataset, the sheltering effects of upwind features in this region provided additional information on snow accumulation. The decrease in significance levels as dmax increased beyond 100 m suggested that the sheltering effects of terrain in this region may not affect snow accumulation as much as in the near terrain. Based on these results, only the Sx₁₀₀ parameter is applied in the succeeding analyses.

As a function of Sx₁₀₀ the data were arbitrarily classified into exposed (Sx₁₀₀ < 0°), slightly sheltered (0° < Sx₁₀₀ < 15°), moderately sheltered (15° < Sx₁₀₀ < 30°), and extremely sheltered samples (Sx₁₀₀ > 30°) to observe the ability of Sx₁₀₀ to define zones of varying accumulation. Box plots of depth for each of the Sx₁₀₀ groups are shown in Fig. 6 and distinctly indicate the ability of the Sx parameter to segregate regions of high and low accumulations.

Comparisons were also drawn between the amount of observed variance in depth explained by Sx₁₀₀ and that explained by elevation, net potential radiation, and slope. In linear models of snow depth, Sx₁₀₀ (p < 0.0001, R² = 0.19) had a substantially stronger relationship to the dependent variable than elevation (p = 0.88), net potential radiation (p = 0.38), and slope (p = 0.0002, R² = 0.03). The implication was that Sx₁₀₀ characterized a significant snow distribution process independent of the processes characterized by the other parameters.

Though elevation and radiation input have been shown to be important controlling processes on snow distribution (e.g., U.S. Army Corp of Engineers 1956; Meiman 1970; Elder et al. 1995), Balk et al. (1998) similarly found low levels of significance for elevation and radiation as predictors of snow depth in another Colorado Rockies Front Range headwater basin. While issues of scale and leeward locations may be affecting the significance of elevation as a factor in these Front Range basins, the low significance of radiation input, which exhibits tremendous heterogeneity over the study areas, highlights how complex process interactions can diminish measures of statistical significance taken from individual parameter tests.

The grid of Sx₁₀₀ (Fig. 7) highlights the areas of the basin located downwind of sheltering topography and those areas that lack a proximal sheltering feature. These highlights are particularly evident as contour lines are traced beneath the watershed-defining northern and southern ridges and in the relatively flat area approximately 350 m east of the Continental Divide. The distribution of Sx₁₀₀ appears to closely follow the pattern of sampled depths, particularly in the mid- to lower basin, where the dominant wind pattern may be more closely associated with that observed at the D-1 meteorological station. Due to complex wind patterns in the upper basin, the D-1 data may not be representative of localized conditions and may be the source of some error.

Drift zones were modeled using a separation distance of 300 m and threshold values for Sb and Sx_o of 7° and 5°, respectively. Detailed analysis of the Sb parameter using an earlier dataset that encompassed the current study area indicated that increasing the separation distance effectively increased the extent of downwind terrain affected by each topographic break, and that a 300-m separation distance was a best fit to conditions in this basin (Winstral 1999). The best fit of the 300-m separation distance to these data may reflect coincidental downwind and downslope expansion of the accumulation zones due to avalanching and sloughing. Additional data would be required to separate these redeposition effects, while application of Sb to other basins would require calibrating sepdist to local conditions. Since grid-determined slopes and the process of deriving averaged parameters across an upwind window both tend to smooth outstanding features, Sb-calculated slopes would be less than those obtained from field measurements. Guided by field studies citing minimum requirements for flow separation of 7.5°–12° (Berg 1977; Tabler 1994), the 7° threshold was chosen to determine lee-slope drift zones. The threshold for Sx_o applied here states that the Sb-defined upwind slope break must have an average maximum exposure of less than 5° to be considered sufficiently susceptible to the large transport fluxes necessary to produce significant downwind drifts.

The general linear model

Yβ₀β₁x₁β₂x₂(6)

where

View Expanded

was fit to the dataset. Results indicated that the p values for the coefficients associated with both Sx and D₀ from the fitting of Eq. (10) to these data were significant (Table 2). Given similar conditions of near shelter (100 m), points located within defined drift zones had significantly higher observed snow depths than points located outside of defined flow separation zones.

The areas of the basin modeled as drift zones can be seen in Fig. 7. The largest modeled drift zone in the lee of the Continental Divide corresponds well with the deeper depth samples taken in the central and southern regions of this zone. The width of the northern extension of this modeled drift zone is, however, questionable. It is possible that winds are being funneled down the Continental Divide, producing a more northerly flow pattern at Navajo Peak not evident in the wind data garnered from the D-1 meteorological station. Another possibility is that lateral convergence of flow, not accounted for in the presented modeling techniques, may be occurring downwind of Navajo Peak, producing higher wind gusts in this area. From the available data, some questions also arise as to the accuracy of the smaller drift zones modeled to the south of the spine between Navajo Peak and the D-1 observation site. Though there were insufficient data available to verify the westernmost of these drift zones, the pattern of deeper samples downslope of it suggest that an accumulation zone may exist upslope of the samples. The other drift zones along this spine have more southerly aspects, and their delineation as drifts is very sensitive to any southerly shift in the prevailing wind direction. The use of a static flow pattern as employed in this study does not account for the natural variance of winds seen through the season. These sites may in fact be subject to alternating periods of deposition and scour based on changes in wind flow, and perhaps a modeling approach that accounts for wind dynamics at a smaller timescale would be more appropriate (see Marks and Winstral 2001).

b. Snow distribution models

Two regression trees were grown to distribute snow depth. In the redistribution tree, the set of predictors contained the variables Sx₁₀₀, D₀, elevation, net potential radiation, and slope, while the nonredistribution tree was based on the latter three predictors. Standard diagnostic checking did not reveal any gross heteroscedasticity or non-normality of the residuals in the regression tree modeling that would have necessitated transformation of the dependent variable, snow depth. The plot of model R² versus tree size (Fig. 8) indicated that the redistribution tree was considerably more accurate than the nonredistribution tree. On a node-by-node basis, the increase in accuracy was approximately 12 percentage points through tree sizes of 30. The cross-validation tests (Fig. 9) for the redistribution tree produced a distinct minimum, with a suggested tree size of 16 (R² = 0.50). Cross-validation results for the nonredistribution tree, however, were inconclusive, producing a range of optimal tree sizes. The nonredistribution tree's cross-validation plot had a minimum at 9 nodes (R² = 0.27) and exhibited slightly increasing deviances as tree size grew to 20 nodes (R² = 0.42), before rising at a rate clearly indicative of model overfit. Based on these results, dependent on the selection of an optimal tree size for the nonredistribution tree, the redistribution tree explained 8%–23% more variance in the observed snow distribution than the tree that did not include a redistribution parameter.

The redistribution regression tree with 16 terminal nodes is shown in Fig. 10. The first three splits of the tree, all based on the Sx₁₀₀ variable, divided the dataset into four groups of increasing depth correlated with increasing wind shelter. As the tree evolved, splitting decisions were increasingly based on elevation, radiation input, and the drift delineator (D₀). None of the splits were based on slope. It was encouraging that all of the splits based on the redistribution parameters were physically meaningful (i.e., increasing shelter contributed to greater modeled depth). The structure of the tree is indicative of the Sx₁₀₀ parameter, explaining the first-order variability of the observations, with elevation, radiation input, and the drift delineator (D₀) having a greater influence on the second-order variability. Evaluation of D₀ in this regard is, however, biased. Though regression trees are capable of handling both continuous and categorical data in the same tree, continuous variables have a distinct competitive advantage over binary classified variables such as D₀ in the determination of node splits (Winstral 1999). The D₀-delineated drift samples largely followed the splits based on increasing values of Sx₁₀₀. Of the original 504 samples, 56 (11%) met the D₀ criteria for drift classification, while, after the first tree split, the 3.75-m node contained 121 sample points, of which 34 (28%) were D₀-drift samples, with the following terminal node of 6.03 m containing 11 points, of which 6 (55%) were D₀-drift samples. Interestingly, while overall results indicated the need for the slope-break exposure qualification (Sx_o) on drift delineation, the five data points at the 6.03-m node that failed to meet the D₀ requirements had sufficient upwind slope breaks yet failed to meet the exposure requirement. Further, more detailed analyses of delineating drifts using the available parameters appears to be warranted.

Extensive snow surveys, like the one applied in this study, are most often limited to small basins, due to cost, accessibility, and manpower concerns. This limitation necessitates some means of extrapolating these basin-scale findings for regional applications. Examination of the regression tree images of distributed snow depth (Fig. 11) offered insight into model performance outside the immediately sampled area. Scaling issues between the broader tree-based model scale and the finer scales of the snow distribution process and modeling of Sx are evident in comparisons with Fig. 7. Both regression tree images were based on models with 16 terminal nodes, with slopes greater than 50° assumed to be too steep to accumulate snow and assigned depth values of zero. The most striking discrepancies between the redistribution and nonredistribution trees occurred along the north-northwest-facing ridge to the west of Kiowa Peak. In this high-elevation region with high wind exposure and low radiation inputs, the nonredistribution tree, lacking a direct means of accounting for wind exposure, assigned the greatest depths in the basin, 4–8 m. The redistribution tree on the other hand, designated this region as a low-accumulation zone having a predominance of 1–2-m depths, with slightly greater depths behind northerly trending transverse ridges. Visual observations made during the study period indicated that this area is in fact a region of very low snow accumulation. Aerial assessments of snow-covered area were not available for the period of the 1999 snow survey; however, there was a classified photograph from 22 May 1996 [1 May snowpack 167% of average (NRCS 1996)] available to compare with the May 1999 [1 May snowpack 112% of average (NRCS 1999)] modeled snow distributions (Fig. 12). Assuming that the locations of low- and high-accumulation zones are annually consistent regardless of snowfall totals, the classified photograph further supported the redistribution model.

Due to the spatial correlation issues mentioned earlier, there was some degree of uncertainty attached to the statistical tests of significance in the first step of data analysis. It is therefore critical to assess whether the parameters are capturing a physical process or picking up idiosyncrasies in either the data or physical characteristics of the basin. Aside from the clearly evident relationship between modeled wind exposure and snow distribution described in the preceding paragraph, there exists other visible correlations between the parameters and snow distribution that cross radiation loadings, elevation bands, and basin characteristics. An example of this is the high snow-accumulation band that extends roughly from the southernmost point in the basin in a northeasterly direction down to the valley floor and part way up the north wall of the basin (Figs. 7 and 12). Radiation inputs along this band range from low along the northeast-facing slopes in the south to high along the southeast-facing slopes in the north. There is an approximate elevation range of 350 m. While radiation loadings, elevation, and slope vary across this region, Sx₁₀₀values remain consistently high (Fig. 7), suggestive of a wind-sheltered region consistent with the observed higher accumulations. A broad visual analysis of Fig. 12 also corroborates some of the earlier statistical findings—that correlations between solar radiation input, elevation, slope, and snow distribution in this basin are very weak. There are, however, striking similarities between the redistribution parameters (Fig. 7), the redistribution regression tree model (Fig. 11b), and the observed snow patterns (Figs. 7 and 12). In this basin, where wind is widely acknowledged as the dominant control on snow distribution, the terrain-based redistribution parameters have seemingly captured snow distribution patterns unrelated to radiation input, elevation, slope, or any apparent basin physiographic features other than those specifically defined by the parameters.