NASA's Land Information System (LIS) provides an integrated framework with multiple choices of land surface models, forcing datasets and parameter information to effectively support various applications such as water resources management and disaster management. In this work, we diagnostically evaluate the snow predictions from several land surface models with multiple forcing data sources to assess the impact of forcing uncertainty vs. that of model uncertainty on snow predictions, as well as the interactions between forcing uncertainty and model uncertainty.

Four models have been chosen for this assessment, including the Community Land Model (CLM version 2), the Noah land surface models version 2.7.1 and version 3.2, and the operational SNOW-17 model of the National Weather Service. Four different sets of precipitation forcing data have been used to drive the models to explore the impacts of forcing uncertainty. These include the forcing data from the Global Data Assimilation System (GDAS) of the National Centers for Environmental Prediction (NCEP), the Merged Analysis of Precipitation (CMAP) of NCEP, the operational North American Mesoscale (NAM) forecast dataset of NCEP, and the in-situ precipitation observations from NRC SNOpack TELemetry (SNOTEL) stations and NOAA Cooperative Observer Program (COOP) stations. For the assessment, a retrospective analysis of 10 years (2001~2010) is performed using LIS (version 6) for the central Alaska, at 0.01o and hourly resolutions. The model predictions of snow quantities including snow cover fraction (SCF), snow water equivalent (SWE), and snow depth are evaluated against SCF observations from the Moderate Resolution Imaging Spectroradiometer (MODIS), SWE observations from the SNOTEL stations, and snow depth observations from the COOP sites, respectively.

As expected, all models generally perform best with the observed in-situ precipitation forcing. There however exist large discrepancies among snow predictions from the four different models, which also exhibit dramatically different responses to uncertainty in the different precipitation forcing sources. For example, during the snow accumulation season where precipitation uncertainty dominates, the CLM seems to favor the CMAP precipitation forcing instead of the GDAS precipitation forcing, outperforming the Noah models which otherwise would exhibit superior performance than the CLM. During the snowmelt season where model uncertainty dominates, the CLM appears to perform the worst regardless of the precipitation forcing source. Overall, the statistics indicate that snow predictions by the mean of the multi-model multi-forcing ensemble consistently outperform all individual members for some observation stations, while consistently outperforming most individual members for most stations. The insights learned from this study on the interplay between model and forcing uncertainties and their implications for ensemble-based snow predictions provide valuable diagnostic information for guiding the ongoing multi-sensor snow data assimilation research.