Patterns of Land Surface Errors and Biases in the Global Forecast System

Abstract

ne year?s worth of Global Forecast System (GFS) predictions of surface meteorological variables (wind speed, air temperature, dewpoint temperature, sea level pressure) are validated for land-based stations over the entire planet for forecasts extending from 0 h into the future (an analysis) to 7 days. Approximately 12 000 surface stations worldwide were included in this analysis. Root-mean-square errors (RMSEs) increased as the forecast period increased from 0 to 36 h, but the initial RMSEs were almost as large as the 36-h forecast RMSEs for all variables. Typical RMSEs were 3°C for air temperature, 2?3 mb for sea level pressure, 3.5°C for dewpoint temperature, and 2.5 m s?1 for wind speed.An analysis of the biases at each station shows that the biggest errors are associated with mountain ranges and other areas of steep topography, with land?sea contrasts also playing a role. When the error is decomposed into the bias, variance, and correlation terms, the large initial RMSEs for the 0-h forecasts are seen to be due to a large forecast bias (which persisted into the longer forecasts) with errors in forecast correlation also making a large contribution.A validation of two subdomains showed results similar to the global validation, but the dependence of the biases on the forecast time was clearer. Finally, the RMSE values climb as forecasts go out when validated out to a period of 7 days as the correlation error term grows.