Quantification of uncertainties in snow accumulation, snowmelt, and snow disappearance dates

Abstract

Seasonal mountain snowpack holds hydrologic and ecologic significance worldwide. However, observation networks in complex terrain are typically sparse and provide minimal information about prevailing conditions. Snow patterns and processes in this data sparse environment can be characterized with numerical models and satellite-based remote sensing, and thus it is essential to understand their reliability. This research quantifies model and remote sensing uncertainties in snow accumulation, snowmelt, and snow disappearance as revealed through comparisons with unique ground-based measurements. The relationship between snow accumulation uncertainty and model configuration is assessed through a controlled experiment at 154 snow pillow sites in the western United States. To simulate snow water equivalent (SWE), the National Weather Service SNOW-17 model is tested as (1) a traditional "forward" model based primarily on precipitation, (2) a reconstruction model based on total snowmelt before the snow disappearance date, and (3) a combination of (1) and (2). For peak SWE estimation, the reliability of the parent models was indistinguishable, while the combined model was most reliable. A sensitivity analysis demonstrated that the parent models had opposite sensitivities to temperature that tended to cancel in the combined model. Uncertainty in model forcing and parameters significantly controlled model accuracy. Uncertainty in remotely sensed snow cover and snow disappearance in forested areas is enhanced by canopy obstruction but has been ill-quantified due to the lack of sub-canopy observations. To better quantify this uncertainty, dense networks of near-surface temperature sensors were installed at four study areas (≤ 1 km2) with varying forest cover in the Sierra Nevada, California. Snow presence at each sensor was detected during periods when temperature was damped, which resulted from snow cover insulation. This methodology was verified using time-lapse analysis and high resolution (15m) remote sensing, and then used to test daily 500 m canopy-adjusted MODIS snow cover data. Relative to the ground sensors, MODIS underestimated snow cover by 10-20% in meadows and 10-40% in forests, and showed snow disappearing 12 to 30 days too early in the forested sites. These errors were not detected with operational snow sensors, which have seen frequent use in MODIS validation studies. The link between model forcing and snow model uncertainty is assessed in two studies using measurements at energy balance stations in different snow climates. First, representation of snow surface temperature (T_s) with temperature and humidity is examined because Ts tracks variations in the snowmelt energy balance. At all sites analyzed, the dew point temperature (T_d) represented T_s with lower bias than the dry and wet-bulb temperatures. The potential usefulness of this approximation was demonstrated in a case study where detection of model bias was achieved by comparing daily T_dand modeled T_s. Second, the impact of forcing data availability and empirical data estimation is addressed to understand which types of data most impact physically-based snow modeling and need improved representation. An experiment is conducted at four well-instrumented sites with a series of hypothetical weather stations to determine which measurements (beyond temperature and precipitation) most impact snow model behavior. Radiative forcings had the largest impact on model behavior, but these are typically the least often measured.