Application of nearest-neighbor resampling techniques for homogenizing temperature records on a daily to sub-daily level

Nearest-neighbor resampling is introduced as a means for homogenizing temperature records on a daily to sub-daily
level. Homogenization refers here to the problem of calculating daily mean and sub-daily temperatures from a time
series subject to irregular observation frequencies and changing observation schedules. The method resamples diurnal
temperature cycles from an observed hourly temperature subrecord at the station. Unlike other methods, the technique
maintains the variance in a natural way. This property is especially important for the analysis of trends and variability of
extremes. For a given day, the resampling technique does not generate a single-valued solution but this peculiarity is of
no effect in the applications considered here. The skills of the nearest-neighbor resampling technique, in terms of bias,
RMSE, and variability, are compared with those of four other methods: a sine-exponential model, a model that uses the
climatological mean daily cycle, a regression model for calculating daily values, and a deterministic version of the nearestneighbor
technique. The series used in the tests is the 1951–2000 meteorological record of De Bilt (The Netherlands). The
emphasis in the comparisons is on the reconstruction of daily mean temperatures. The analysis shows important differences
in performance between the models. The regression-based method performs best with respect to the calculation of the
individual daily mean temperatures; the day-to-day variability is best reproduced with the nearest-neighbor resampling
technique. The performance of the models improves when cloudiness is used as an extra predictor. The improvement is,
however, small compared to the intermodel differences. The type of model that should be used depends on the desired
application. For trend and variability studies, the nearest-neighbor resampling technique performs best. Nearest-neighbor
resampling can successfully be performed even in situations where the length of the hourly subrecord is an order of
magnitude less than the length of the series to be homogenized.