In developing homogeneous, daily time series of extreme temperatures, a number of obstacles related to station irregularities are encountered. Previous researchers have implemented a method based on difference series from nearby, correlated stations to adjust for such inhomogeneities in mean temperature series. This test, unfortunately, falls short in regions of low station density where the distance between the closest neighboring station is large and between station correlation is low. This is particularly a problem in long time series where few stations have records back to the turn of the century. An alternative test, which uses only the station in question to detect data inhomogeneities, is described. Such a method removes the restrictions of identifying nearby, correlated stations with homogeneous records, providing a more versatile adjustment scheme for data discontinuities, particularly those associated with temperature extremes.
The single-station test is applied to annual temperature threshold exceedence (e.g., days with maximum temperatures > 32.2 C) series. Station metadata is used to identify inhomogeneities. The 75th and 25th percentiles are computed for the longest homogeneous time period (i.e. the stationary (no significant trend) period before or after the documented inhomogeneity). These values then form the basis of the test which compares the differences between the percentage of years that exceed the 75th percentile and fall below the 25th percentile in both periods. In cases where the discontinuity does not affect the temperature record, it is expected that during the shorter period a similar percentage of years (i.e. approximately 25%) will exceed (fall below) the 75th (25th) percentile of the longer period. Resampling techniques are used the determine the statistical significance of the inhomogeneity. A similar test that can be applied to non-stationary time series is also discussed and evaluated.
A comparable number of known discontinuities are correctly identified by both the single station-test and a similar test using a difference series of annual temperature threshold exceedences formed using a neighboring station. At the alpha = 0.05 level, more than 50% of the 1.1 C (2 F) discontinuities are identified. Discontinuities > 1.7 C (3 F) are detected in nearly all cases. The single station test tends to out-perform the difference series test when the correlation between neighboring stations is less than approximately 0.50