Estimating the error of the BCDG analysis of surface data

As part of the Localized Aviation MOS Program (LAMP), the Meteorological Development Laboratory is analyzing surface data reports on an hourly basis. The Bergthorssen-Cressman-Doos-Glahn (BCDG) analysis program being used for gridding MOS forecasts has been tailored to analyze quasi-continuous surface observations. It is desired to know the errors involved in these analyses. While the actual errors are unknowable for several reasons, they can be estimated. Withheld data tests can be made to get an estimate of overall errors, or more correctly put–differences between the data values and the analyses. On any given analysis, one would expect the errors to be a function of some knowable parameters, such as distances between the reporting locations and the gridpoints, the terrain roughness, the density of reporting locations in the immediate vicinity, and the variability of the data values in the immediate vicinity.

We have made analyses of surface temperature and dewpoint each hour for six months of data. On each analysis, 20 land stations and one water station were randomly withheld from the analysis. For each withheld reporting site, the analysis value at that site was estimated by bilinear interpolation. The differences between these interpolated values and the actual observations were related to knowable parameters through linear regression, one relationship for land and another for water. These regression equations were then applied, respectively, to each land and water gridpoint. This, then, produced an estimate of the error of the analysis for each gridpoint for that specific time. This paper will explain the process in detail and show results.