首页 | 本学科首页   官方微博 | 高级检索  
相似文献
 共查询到20条相似文献，搜索用时 421 毫秒
1.
High‐resolution snow depth (SD) maps (1 × 1 m) obtained from terrestrial laser scanner measurements in a small catchment (0.55 km2) in the Pyrenees were used to assess small‐scale variability of the snowpack at the catchment and sub‐grid scales. The coefficients of variation are compared for various plot resolutions (5 × 5, 25 × 25, 49 × 49, and 99 × 99 m) and eight different days in two snow seasons (2011–2012 and 2012–2013). We also studied the relation between snow variability at the small scale and SD, topographic variables, small‐scale variability in topographic variables. The results showed that there was marked variability in SD, and it increased with increasing scales. Days of seasonal maximum snow accumulation showed the least small‐scale variability, but this increased sharply with the onset of melting. The coefficient of variation (CV) in snowpack depth showed statistically significant consistency amongst the various spatial resolutions studied, although it declined progressively with increasing difference between the grid sizes being compared. SD best explained the spatial distribution of sub‐grid variability. Topographic variables including slope, wind sheltering, sub‐grid variability in elevation, and potential incoming solar radiation were also significantly correlated with the CV of the snowpack, with the greatest correlation occurring at the 99 × 99 m resolution. At this resolution, stepwise multiple regression models explained more than 70% of the variance, whereas at the 25 × 25 m resolution they explained slightly more than 50%. The results highlight the importance of considering small‐scale variability of the SD for comprehensively representing the distribution of snowpack from available punctual information, and the potential for using SD and other predictors to design optimized surveys for acquiring distributed SD data. Copyright © 2014 John Wiley & Sons, Ltd.  相似文献   

2.
Seasonal snowpacks in marginal snow environments are typically warm and nearly isothermal, exhibiting high inter‐ and intra‐annual variability. Measurements of snow depth and snow water equivalent were made across a small subalpine catchment in the Australian Alps over two snow seasons in order to investigate the extent and implications of snowpack spatial variability in this marginal setting. The distribution and dynamics of the snowpack were found to be influenced by upwind terrain, vegetation, solar radiation, and slope. The role of upwind vegetation was quantified using a novel parameter based on gridded vegetation height. The elevation range of the catchment was relatively modest (185 m), and elevation impacted distribution but not dynamics. Two characteristic features of marginal snowpack behaviour are presented. Firstly, the evolution of the snowpack is described in terms of a relatively unstable accumulation state and a highly stable ablation state, as revealed by temporal variations in the mean and standard deviation of snow water equivalent. Secondly, the validity of partitioning the snow season into distinct accumulation and ablation phases is shown to be compromised in such a setting. Snow at the most marginal locations may undergo complete melt several times during a season and, even where snow cover is more persistent, ablation processes begin to have an effect on the distribution of the snowpack early in the season. Our results are consistent with previous research showing that individual point measurements are unable to fully represent the variability in the snowpack across a catchment, and we show that recognising and addressing this variability are particularly important for studies in marginal snow environments.  相似文献   

3.
A generalized additive model (GAM) was used to model the spatial distribution of snow depth in the central Spanish Pyrenees. Statistically significant non‐linear relationships were found between distinct location and topographical variables and the average depth of the April snowpack at 76 snow poles from 1985 to 2000. The joint effect of the predictor variables explained more than 73% of the variance of the dependent variable. The performance of the model was assessed by applying a number of quantitative approaches to the residuals from a cross‐validation test. The relatively low estimated errors and the possibility of understanding the processes that control snow accumulation, through the response curves of each independent variable, indicate that GAMs may be a useful tool for interpolating local snow depth or other climate parameters. Copyright © 2005 John Wiley & Sons, Ltd.  相似文献   

4.
Snowpacks and forests have complex interactions throughout the large range of altitudes where they co-occur. However, there are no reliable data on the spatial and temporal interactions of forests with snowpacks, such as those that occur in nearby areas that have different environmental conditions and those that occur during different snow seasons. This study monitored the interactions of forests with snowpacks in four forest stands in a single valley of the central Spanish Pyrenees during three consecutive snow seasons (2015/2016, 2016/2017 and 2017/2018). Daily snow depth data from time-lapse cameras were compared with snow data from field surveys that were performed every 10–15 days. These data thus provided information on the spatial and temporal changes of snow–water equivalent (SWE). The results indicated that forest had the same general effects on snowpack in each forest stand and during each snow season. On average, forest cover reduced the duration of snowpack by 17 days, reduced the cumulative SWE of the snowpack by about 60% and increased the spatial heterogeneity of snowpack by 190%. Overall, forest cover reduced SWE total accumulation by 40% and the rate of SWE accumulation by 25%. The forest-mediated reduction of the accumulation rate, in combination with the occasional forest-mediated enhancement of melting rate, explained the reduced duration of snowpacks beneath forest canopies. However, the magnitude and timing of certain forest effects on snowpack had significant spatial and temporal variations. This variability must be considered when selecting the location of an experimental site in a mountainous area, because the study site should be representative of surrounding areas. The same considerations apply when selecting a time period for study.  相似文献   

5.
This study analyzes spatial variability of snow depth and density from measurements made in February and April of 2010 and 2011 in three 1–2 km2 areas within a valley of the central Spanish Pyrenees. Snow density was correlated with snow depth and different terrain characteristics. Regression models were used to predict the spatial variability of snow density, and to assess how the error in computed densities might influence estimates of snow water equivalent (SWE).The variability in snow depth was much greater than that of snow density. The average snow density was much greater in April than in February. The correlations between snow depth and density were generally statistically significant but typically not very high, and their magnitudes and signs were highly variable among sites and surveys. The correlation with other topographic variables showed the same variability in magnitude and sign, and consequently the resulting regression models were very inconsistent, and in general explained little of the variance. Antecedent climatic and snow conditions prior to each survey help highlight the main causes of the contrasting relation shown between snow depth, density and terrain. As a consequence of the moderate spatial variability of snow density relative to snow depth, the absolute error in the SWE estimated from computed densities using the regression models was generally less than 15%. The error was similar to that obtained by relating snow density measurements directly to adjacent snow depths.  相似文献   

6.
Snowpack water equivalent (SWE) is a key variable for water resource management in snow-dominated catchments. While it is not feasible to quantify SWE at the catchment scale using either field surveys or remotely sensed data, technologies such as airborne LiDAR (light detection and ranging) support the mapping of snow depth at scales relevant to operational water management. To convert snow depth to water equivalent, models have been developed to predict SWE or snowpack density based on snow depth and additional predictor variables. This study builds upon previous models that relate snowpack density to snow depth by including additional predictor variables to account for (1) long-term climatologies that describe the prevailing conditions influencing regional snowpack properties, and (2) the effect of intra- and inter-year variability in meteorological conditions on densification through a cumulative degree-day index derived from North American Regional Reanalysis products. A non-linear model was fit to 114 506 snow survey measurements spanning 41 years from 1166 snow courses across western North America. Under spatial cross-validation, the predicted densities had a root-mean-square error of 47.1 kg m−3, a mean bias of −0.039 kg m−3, and a Nash-Sutcliffe Efficiency of 0.70. The model developed in this study had similar overall performance compared to a similar regression-based model reported in the literature, but had reduced seasonal biases. When applied to predict SWE from simulated depths with random errors consistent with those obtained from LiDAR or Structure-from-Motion, 50% of the SWE estimates for April and May fell within −45 to 49 mm of the observed SWE, representing prediction errors of −15% to 20%.  相似文献   

7.
Multivariate statistical analysis was used to explore relationships between catchment topography and spatial variability in snow accumulation and melt processes in a small headwater catchment in the Spanish Pyrenees. Manual surveys of snow depth and density provided information on the spatial distribution of snow water equivalent (SWE) and its depletion over the course of the 1997 and 1998 melt seasons. A number of indices expressing the topographic control on snow processes were extracted from a detailed digital elevation model of the catchment. Bivariate screening was used to assess the relative importance of these topographic indices in controlling snow accumulation at the start of the melt season, average melt rates and the timing of snow disappearance. This suggested that topographic controls on the redistribution of snow by wind are the most important influence on snow distribution at the start of the melt season. Furthermore, it appeared that spatial patterns of snow disappearance were largely determined by the distribution of snow water equivalent (SWE) at the start of the melt season, rather than by spatial variability in melt rates during the melt season. Binary regression tree models relating snow depth and disappearance date to terrain indices were then constructed. These explained 70–80% of the variance in the observed data. As well as providing insights into the influence of topography on snow processes, it is suggested that the techniques presented herein could be used in the parameterization of distributed snowmelt models, or in the design of efficient stratified snow surveys. Copyright © 2003 John Wiley & Sons, Ltd.  相似文献   

8.
The spatial and temporal distribution of snow accumulation is complex and significantly influences the hydrological characteristics of mountain catchments. Many snow redistribution processes, such as avalanching, slushflow or wind drift, are controlled by topography, but their modelling remains challenging. In situ measurements of snow accumulation are laborious and generally have a coarse spatial or temporal resolution. In this respect, time‐lapse photography shows itself as a powerful tool for collecting information at relatively low cost and without the need for direct field access. In this paper, the snow accumulation distribution of an Alpine catchment is inferred by adjusting a simple snow accumulation model combined with a temperature index melt model to match the modelled melt‐out pattern evolution to the pattern monitored during an ablation season through terrestrial oblique photography. The comparison of the resulting end‐of‐winter snow water equivalent distribution with direct measurements shows that the achieved accuracy is comparable with that obtained with an inverse distance interpolation of the point measurements. On average over the ablation season, the observed melt‐out pattern can be reproduced correctly in 93% of the area visible from the fixed camera. The relations between inferred snow accumulation distribution and topographic variables indicate large scatter. However, a significant correlation with local slope is found and terrain curvature is detected as a factor limiting the maximal snow accumulation. Copyright © 2010 John Wiley & Sons, Ltd.  相似文献   

9.
We report a methodology for reconstructing the daily snow depth distribution at high spatial resolution in a small Pyrenean catchment using time‐lapse photographs and snow depletion rates derived from an on‐site measuring meteorological station. The results were compared with the observed snow depth distribution, determined on a number of separate occasions using a terrestrial laser scanner (TLS). The time‐lapse photographs were projected onto a digital elevation model of the study site, and converted into snow presence/absence information. The melt‐out date (MOD; first occurrence of melt out after peak snow accumulation) was obtained from the projected photograph series. Commencing the backward reconstruction for each grid cell at the MOD, the method uses simulated snow depth depletion rates using a temperature index approach, which are extrapolated to the grid cells of the domain to arrive at the snow distribution of the previous day. Two variants of the reconstruction techniques were applied (1) using a spatially constant degree day factor (DDF) for calculating the daily expected snow depth depletion rate, and (2) allowing a spatially distributed DDF calculated from two consecutive TLS acquisitions compared to the snow depth depletion rate observed at the meteorological station. Validation revealed that both methods performed well (average R2 = 0.68; standard RMSE = 0.58), with better results obtained from the spatially distributed approach. Nevertheless, the spatially corrected DDF reconstruction, which requires TLS data, suggests that the constant DDF approach is an efficient, and for most applications sufficiently accurate and easily reproducible method. The results highlight the usefulness of time‐lapse photography for not only determining the snow covered area, but also for estimating the spatial distribution of snow depth. Copyright © 2016 John Wiley & Sons, Ltd.  相似文献   

10.
KEVIN SHOOK  D. M. GRAY 《水文研究》1996,10(10):1283-1292
The results of a field study of the small-scale spatial structure of the depth of shallow seasonal snowcovers in prairie and arctic environments are presented. It is shown that the spatial distribution of snow depth is fractal at small scales, becoming random at scales beyond some limiting length. This is due to the autocorrelation of depth at small sampling distances. The transition of fractal to random behaviour is indexed by a ‘cutoff length’, which is defined by the intersection of the ‘fractal’ slope and horizontal tangent of a logarithmic plot of the standard deviation of depth versus sampling distance. The magnitude of the cutoff length is related to the degree of macroscopic variability of the underlying topography. An increase in length due to the effects of macroscopic topographic variability on snowcover accumulation is confirmed by de-trending field measurements. The de-trended data shown a cutoff length for wheat stubble and fallow surfaces of approximately 30 m, which is consistent with the distance determined from measurements on ‘flat’ fields. The implications of the transition of snow depth from fractal to random structure on the scales of snow sampling and modelling are presented. The cutoff length may provide a statistic for stratifying shallow snowcovers, by linking snowcover properties to the underlying topography.  相似文献   

11.
Comprehensive snow depth data, collected using georadar and hand probing, were used for statistical analyses of snow depths inside 1 km grid cells. The sub‐grid cell spatial scale was 100 m. Statistical distribution functions were found to have varying parameters, and an attempt was made to connect these statistical parameters to different terrain variables. The results showed that the two parameters mean and standard deviation of snow depth were significantly related to the sub‐grid terrain characteristics. Linear regression models could explain up to 50% of the variation for both of the snowcover parameters mentioned. Copyright © 2004 John Wiley & Sons, Ltd.  相似文献   

12.
This paper reports on a study analysing the spatial distribution functions, the correlation structures, and the power spectral densities of high‐resolution LIDAR snow depths (～1 m) in two adjacent 500 m × 500 m areas in the Colorado Rocky Mountains, one a sub‐alpine forest the other an alpine tundra. It is shown how and why differences in the controlling physical processes induced by variations in vegetation cover and wind patterns lead to the observed differences in spatial organization between the snow depth fields of these environments. In the sub‐alpine forest area, the mean of snow depth increases with elevation, while its standard deviation remains uniform. In the tundra subarea, the mean of snow depth decreases with elevation, while its standard deviation varies over a wide range. The two‐dimensional correlations of snow depth in the forested area indicate little spatial memory and isotropic conditions, while in the tundra they indicate a marked directional bias that is consistent with the predominant wind directions and the location of topographic ridges and depressions. The power spectral densities exhibit a power law behaviour in two frequency intervals separated by a break located at a scale of around 12 m in the forested subarea, and 65 m in the tundra subarea. The spectral exponents obtained indicate that the snow depth fields are highly variable over scales larger than the scale break, while highly correlated below. Based on the observations and on synthetic snow depth fields generated with one‐ and two‐dimensional spectral techniques, it is shown that the scale at which the break occurs increases with the separation distance between snow depth maxima. In addition, the breaks in the forested area coincide with those of the corresponding vegetation height field, while in the tundra subarea they are displaced towards larger scales than those observed in the corresponding vegetation height field. Copyright © 2009 John Wiley & Sons, Ltd.  相似文献   

13.
Collecting spatially representative data over large areas is a challenge within snow monitoring frameworks. Identifying consistent trends in snow accumulation properties enables increased sampling efficiency by minimizing field collection time and/or remote sensing costs. Seasonal snowpack depth estimations during mid-winter and melt onset conditions were derived from airborne Lidar over the West Castle Watershed in the southern Canadian Rockies on three dates. Each dataset was divided into five sets of snow depth driver classes: elevation, aspect, topographic position index, canopy cover and slope. Datasets were quality controlled by eliminating snow depth values above the 99th percentile value, which had a negligible effect on average snow depths. Consistent trends were observed among driver classes with peak snow accumulation occurring within the treeline ecotone, north-facing aspects, open canopies, topographic depressions and areas with low slope angle. Although mid-winter class trends for each driver were similar and watershed-scale snow depth distributions were significantly correlated (0.76, p < .01), depth distributions within the same driver class of the three datasets were not correlated due to recent snowfall events, redistribution and settling processes. Trends in driver classes during late season melt onset were similar to mid-winter conditions but watershed scale distribution correlation results varied with seasonality (0.68 mid-winter 2014 and melt onset 2016; 0.65 mid-winter 2017 and melt onset 2016, p < .1). This is due to the differing stages of accumulation or ablation and the upward migration in the 0°C isotherm during spring, when snow depth can be declining in valley bottoms while still increasing at higher elevations. The observed consistency in depth driver controls can be used to guide future integrated snow monitoring frameworks.  相似文献   

14.
We analyse spatial variability and different evolution patterns of snowpack in a mixed beech–fir stand in the central Pyrenees. Snow depth and density were surveyed weekly along six transects of contrasting forest cover during a complete accumulation and melting season; we also surveyed a sector unaffected by canopy cover. Forest density was measured using the sky view factor (SVF) obtained from digital hemispherical photographs. During periods of snow accumulation and melting, noticeable differences in snow depth and density were found between the open site and those areas covered by forest canopy. Principal component analysis provided valuable information in explaining these observations. The results indicate a high variability in snow accumulation within forest areas related to differences in canopy density. Maximum snow water equivalent (SWE) was reduced by more than 50% beneath dense canopies compared with clearings, and this difference increased during the melting period. We also found significant temporal variations: when melting began in sectors with low SVF, most of the snow had already thawed in areas with high SVF. However, specific conditions occasionally produced a different response of SWE to forest cover, with lower melting rates observed beneath dense canopies. The high values of correlation coefficients for SWE and SVF (r > 0·9) indicate the reliability of predicting the spatial distribution of SWE in forests when only a moderate number of observations are available. Digital hemispherical photographs provide an appropriate tool for this type of analysis, especially for zenith angles in the range 35–55 . Copyright © 2007 John Wiley & Sons, Ltd.  相似文献   

15.
W. T. Sloan  C. G. Kilsby  R. Lunn 《水文研究》2004,18(17):3371-3390
General circulation models (GCMs), or stand‐alone models that are forced by the output from GCMs, are increasingly being used to simulate the interactions between snow cover, snowmelt, climate and water resources. The variation in snowpack extent, and hence albedo, through time in a cell is likely to be substantial, especially in mid‐latitude mountainous regions. As a consequence, the energy budget simulation by a GCM relies on a realistic representation of snowpack extent. Similarly, from a water resource perspective, the spatial extent of the pack is key in predicting meltwater discharges into rivers. In this paper a simple computationally efficient regional snow model has been developed, which is based on a degree‐day approach and simulates the fraction of the model domain covered by snow, the spatially averaged melt rate and the mean snowpack depth. Computational efficiency is achieved through a novel spatial averaging procedure, which relies on the assumptions that precipitation and temperature scale linearly with elevation and that the distribution of elevations in the domain can be modelled by a continuous function. The resulting spatially averaged model is compared with both observations of the duration of snow cover throughout Austria and with results from a distributed model based on the same underlying assumptions but applied at a fine spatial resolution. The new spatially averaged model successfully simulated the seasonal snow duration observations and reproduced the daily dynamics of snow cover extent, the spatially averaged melt rate and mean pack depth simulated by the distributed model. It, therefore, offers a computationally efficient and easily applied alternative to the current crop of regional snow models. Copyright © 2004 John Wiley & Sons, Ltd.  相似文献   

16.
The spatio‐temporal distribution of snow in a catchment during ablation reflects changes in the total amount of snow water equivalent and is thus a key parameter for the estimation of melt water run‐off. This study explores possible rules behind the spatial variability of snow depth during the ablation season in a small Alpine catchment with complex topography. The snow depth observations are based on more than 160 000 terrestrial laser scanner data points with a spatial resolution of 1 m, which were obtained from 11 scanning campaigns of two consecutive ablation seasons. The analysis suggests that for estimating cumulative snow melt dynamics from the catchment investigated, assessing the initial snow distribution prior to the melt season is more important than addressing spatial differences in the melt behaviour. Snow volume and snow‐covered area could be predicted well using a conceptual melt model assuming spatially uniform melt rates. However, accurate results were only obtained if the model was initialized with a pre‐melt snow distribution that reflected measured mean and standard deviation. Using stratified melt rates on the other hand did not improve the model results. At least for sites with similar meteorological and topographical conditions, the model approach presented here comprises an efficient way to estimate snow depletion dynamics, especially if persistent snow accumulation pattern between years facilitate the characterization of the initial snow distribution prior to the melt. Copyright © 2011 John Wiley & Sons, Ltd.  相似文献   

17.
The spatial distribution of snow water equivalent (SWE) is a key variable in many regional‐scale land surface models. Currently, the assimilation of point‐scale snow sensor data into these models is commonly performed without consideration of the spatial representativeness of the point data with respect to the model grid‐scale SWE. To improve the understanding of the relationship between point‐scale snow measurements and surrounding areas, we characterized the spatial distribution of snow depth and SWE within 1‐, 4‐ and 16‐km2 grids surrounding 15 snow stations (snowpack telemetry and California snow sensors) in California, Colorado, Wyoming, Idaho and Oregon during the 2008 and 2009 snow seasons. More than 30 000 field observations of snowpack properties were used with binary regression tree models to relate SWE at the sensor site to the surrounding area SWE to evaluate the sensor representativeness of larger‐scale conditions. Unlike previous research, we did not find consistent high biases in snow sensor depth values as biases over all sites ranged from 74% overestimates to 77% underestimates. Of the 53 assessments, 27 surveys indicated snow station biases of less than 10% of the surrounding mean observed snow depth. Depth biases were largely dictated by the physiographic relationship between the snow sensor locations and the mean characteristics of the surrounding grid, in particular, elevation, solar radiation index and vegetation density. These scaling relationships may improve snow sensor data assimilation; an example application is illustrated for the National Operational Hydrologic Remote Sensing Center National Snow Analysis SWE product. The snow sensor bias information indicated that the assimilation of point data into the National Operational Hydrologic Remote Sensing Center model was often unnecessary and reduced model accuracy. Copyright © 2012 John Wiley & Sons, Ltd.  相似文献   

18.
Snow temperature is a major component of many physical processes in a snowpack. The temperature and the change in temperature across a layer have a dominant effect on physical properties of snow grains as well as its hardness, strength, and failure resistance. In this study, temperature and snow cover thickness were measured during the snow season of 2007–2008 in 11 elevation classes and in three different sampling locations, one in an open area and two under different forest canopy covers for each class along Kartalkaya road, Bolu. Each sampling site was visited 44 times to collect data including snow depth, snow surface temperature, ground temperature, and temperature within snowpack at 20‐cm intervals. Seven different models are developed to determine snowpack temperature variations under forest canopy covers and in an open area with different leaf area index values. All models were performed using a multilayer perceptron (MP) method for the Bolu–Kartalkaya area, Turkey. MP approach constitutes a standard form of neural network modeling and can modify two‐layer linear perceptron methods using three and more layers. The ability of MP is to handle complex nonlinear interactions, which ease the natural process of modeling. This method can overcome complex computations using neuron networks, and they can easily nonlinearly link input and output variables. The predictive errors are determined on the basis of mean absolute error and mean square error criteria. The Nash–Sutcliffe sufficiency score showing compliance between observed and predicted values is also calculated. According to the mean absolute error, the mean square error, and the Nash–Sutcliffe sufficiency score criteria, the predictive errors are within reasonable error intervals, justifying the use of the developed MP models for engineering applications. Copyright © 2012 John Wiley & Sons, Ltd.  相似文献   

19.
Seasonal snow is a globally important water resource that contributes substantially to upland and lowland water resources. As such, there is a need to understand the controls on the spatial and temporal variation in snow distribution. This study meets this research need by investigating the topographic controls on snow depth distribution in the upper Jollie catchment in the Southern Alps of New Zealand. Furthermore, inter‐annual variation in the importance of the topographic controls is examined and linked to variation in the dominant synoptic‐scale weather patterns over a 4‐year period (2007–2010). Through the use of regression trees, the relative importance of the topographic controls on snow depth was shown to vary between the four study years. In particular, elevation explained the greatest amount of variance in 2007 and 2008 and east‐exposure explained the greatest variance in 2009 and 2010. The other wind exposure variables also had a large effect on the snow depth distribution in 2009 and 2010. Differences in the frequency and duration of synoptic weather patterns were physically consistent with the changing importance of these variables. In particular, a higher frequency of troughing events in 2009 and 2010 is thought to be associated with a reduced importance of elevation and greater influence of wind exposure on snow depth in these years. These findings demonstrate the importance of using multi‐year data sets, and of considering topographic and climatic influences, when attempting to model alpine snow distribution. Copyright © 2014 John Wiley & Sons, Ltd.  相似文献   

20.
Historically, observing snow depth over large areas has been difficult. When snow depth observations are sparse, regression models can be used to infer the snow depth over a given area. Data sparsity has also left many important questions about such inference unexamined. Improved inference, or estimation, of snow depth and its spatial distribution from a given set of observations can benefit a wide range of applications from water resource management, to ecological studies, to validation of satellite estimates of snow pack. The development of Light Detection and Ranging (LiDAR) technology has provided non‐sparse snow depth measurements, which we use in this study, to address fundamental questions about snow depth inference using both sparse and non‐sparse observations. For example, when are more data needed and when are data redundant? Results apply to both traditional and manual snow depth measurements and to LiDAR observations. Through sampling experiments on high‐resolution LiDAR snow depth observations at six separate 1.17‐km2 sites in the Colorado Rocky Mountains, we provide novel perspectives on a variety of issues affecting the regression estimation of snow depth from sparse observations. We measure the effects of observation count, random selection of observations, quality of predictor variables, and cross‐validation procedures using three skill metrics: percent error in total snow volume, root mean squared error (RMSE), and R2. Extremes of predictor quality are used to understand the range of its effect; how do predictors downloaded from internet perform against more accurate predictors measured by LiDAR? Whereas cross validation remains the only option for validating inference from sparse observations, in our experiments, the full set of LiDAR‐measured snow depths can be considered the ‘true’ spatial distribution and used to understand cross‐validation bias at the spatial scale of inference. We model at the 30‐m resolution of readily available predictors, which is a popular spatial resolution in the literature. Three regression models are also compared, and we briefly examine how sampling design affects model skill. Results quantify the primary dependence of each skill metric on observation count that ranges over three orders of magnitude, doubling at each step from 25 up to 3200. Whereas uncertainty (resulting from random selection of observations) in percent error of true total snow volume is typically well constrained by 100–200 observations, there is considerable uncertainty in the inferred spatial distribution (R2) even at medium observation counts (200–800). We show that percent error in total snow volume is not sensitive to predictor quality, although RMSE and R2 (measures of spatial distribution) often depend critically on it. Inaccuracies of downloaded predictors (most often the vegetation predictors) can easily require a quadrupling of observation count to match RMSE and R2 scores obtained by LiDAR‐measured predictors. Under cross validation, the RMSE and R2 skill measures are consistently biased towards poorer results than their true validations. This is primarily a result of greater variance at the spatial scales of point observations used for cross validation than at the 30‐m resolution of the model. The magnitude of this bias depends on individual site characteristics, observation count (for our experimental design), and sampling design. Sampling designs that maximize independent information maximize cross‐validation bias but also maximize true R2. The bagging tree model is found to generally outperform the other regression models in the study on several criteria. Finally, we discuss and recommend use of LiDAR in conjunction with regression modelling to advance understanding of snow depth spatial distribution at spatial scales of thousands of square kilometres. Copyright © 2012 John Wiley & Sons, Ltd.  相似文献   

设为首页 | 免责声明 | 关于勤云 | 加入收藏

Copyright©北京勤云科技发展有限公司  京ICP备09084417号