An Evaluation of Snow Initializations for NCEP Global and Regional Forecasting Models

Abstract:

Snow plays a major role in land-atmosphere interactions, affecting the forecasting of weather, climate, and water resources. At the same time, the strong spatial heterogeneity in snow depth and snow water equivalent (SWE) makes it challenging to evaluate gridded snow quantities using in situ point measurements. First, we have developed a new method to upscale point measurements into gridded datasets. This method is found to be superior to three other methods. It is then used to generate daily snow depth and SWE datasets for water years 2012-2014 at eight 2° X 2° areas using in situ measurements from the COOP and SNOTEL networks. These areas encompass a variety of terrain characteristics over North America.

These datasets are used to quantify the performance of daily snow depth and SWE initialization in the NCEP global forecasting models (GFS and CFS) and regional models (NAM and RAP). Model initializations which utilize AFWA snow depths (GFS, CFS, and NAM) are found to have a too shallow snow depth compared to our area averaged method. Across all areas and water years, our method averaged 0.58m (0.57m) of snow while the models averaged 0.18m (0.19m) with a mean absolute error of 0.42m (0.47m) for the global (regional) models utilizing AFWA data. These models also ended the snow season much too early on average (by more than a month). The RAP model, which cycles snow instead of initializing with AFWA snow depths, underestimates snow depth to a lesser degree and has a mean absolute error of 0.26m while ending the snow season about two weeks early on average. Compared with snow depth errors, SWE errors from GFS, CFS, and NAM are even larger because of their use of globally constant snow densities. Furthermore, we have evaluated the daily snow depth gridded data produced by the Canadian Meteorological Centre (CMC), which has been utilized as the best available ground truth in multiple studies. It is found that the CMC product underestimates snow depth and has a mean absolute error of 0.40m on average across the eight areas and ended the snow season more than one month early on average. In other words, it is similar to the GFS, CFS, and NAM, but worse than RAP.

In this presentation, we will discuss these results, along with our current work on the development of a simple, physically based, snow density model for snow initialization.