Since numerous modeling and statistical applications require serially complete temperature data, a method to estimate missing and erroneous temperature data has been developed and implemented by NRCC for Cooperative Network Stations in the northeastern United States. The estimation procedure uses the same general principles used by Steurer (1985). However, compared to Steurer's methods, our procedure significantly improves the accuracy of maximum temperatures estimates and reduces maximum errors by as much as 10°F. This improved accuracy can be attributed to our requirement that observations used in estimating missing temperatures be taken at a time close to the observation time used at the missing-data station. In addition, the stations used in reconstruction are selected based on minimum distance from the missing-data station rather than by their inclusion in the same climate division. Since climate divisions are often based on political rather than climatic boundaries, the distance approach increases the probability that stations used for estimation have experienced the same meteorological conditions on any given day.
Based on the results of validation analyses, it appears that the magnitude of MAE for maximum temperature is influenced by station density. Where several adjacent stations with similar observation times could be found, errors tended to be small, on the order of 1.5°F. In data sparse regions, however, MAE errors increased to approximately 2.5°F.
The accuracy of minimum temperature estimates is also improved somewhat with our procedure. The relatively small increase in the accuracy of minimum temperature estimates most likely results from two factors. Temperature biases resulting from observation time differences tend to be smaller fro minimum temperatures than those associated with maximum temperatures (DeGaetano and Knapp, 1993). In addition, most morning observations in the region are taken at a late enough hour (07:00-08:00) that observation time bias associated with minimum temperature is minimal. In contrast, the majority of afternoon observations are taken near the time when observation time bias is at a maximum. This contributes to the improvement in accuracy obtained when observation time was considered in estimating maximum temperatures.
Microclimate differences between the stations used in the estimation process and the missing-data site may also account for the limited improvement in the minimum temperature estimates. It is likely that, on a given day, the departure of the minimum temperature form normal will be influenced by microclimatic features over short distances. These effects should be most prevalent in mountainous terrain, where temperature inversions and cold air drainage can produce sharp minimum temperature gradients. the results of validation analyses appear to support this conclusion, since the largest MAE values associated with stations from primarily mountainous areas of West Virginia.
9 pp.
To purchase this publication, fill out this form.