Estimating missing daily temperature extremes using an optimized regression approach

A variation of a least squares regression approach to estimate missing daily maximum and minimum temperatures is developed and evaluated, specifically for temperature extremes. The method focuses on obtaining accurate estimates of annual exceedence counts (e.g. the number of days greater than or equal to the 90th percentile of daily maximum temperatures), as well as counts of consecutive exceedences, while limiting the estimation error associated with each individual value. The performance of this method is compared with that of two existing methods developed for the entire temperature distribution. In these existing methods, temperature estimates are based on data from neighbouring stations using either regression or temperature departure‐based approaches. Evaluation of our approach using cold minimum and warm maximum temperatures shows that the median percentage of correctly identified exceedence counts is 97% and the median percentage of correctly identified consecutive exceedence counts is 98%. The other existing methods tend to underestimate both single and consecutive exceedence counts. Using these procedures, the estimated exceedence counts are generally less than 80% of those that actually occurred. Despite the fact that our method is tuned to estimate exceedence counts, the estimation accuracy of individual daily maximum or minimum temperatures is similar to that of the other estimation procedures. The median absolute error (MAE) using all temperatures greater than or equal to the 90th percentile ( T 90 )−1.1°C for ten climatically diverse stations is 1.28°C for our method, while the other methods give MAEs of 1.27 and 1.17°C. In terms of median error, however, the tendency for underprediction by the existing methods is pronounced with −0.77 and −0.61°C biases. Our optimized method is relatively unbiased as the resulting mean error is −0.12°C. Copyright © 2001 Royal Meteorological Society