New Techniques for Detection and Adjustment of Shifts in Daily Precipitation Data Series

This study integrates a Box-Cox power transformation procedure into a common trend two-phase regression model based test (the PMFred algorithm) for detecting changepoints, to make the test applicable to non-Gaussian data series, such as non-zero daily precipitation amounts. The detection power aspects of the transformed method (transPMFred) were assessed using Monte Carlo simulations, which show that this new algorithm is much better than the corresponding untransformed method for non-Gaussian data series. The transPMFred algorithm was also shown to be able to detect three changepoints in a non-zero precipitation series recorded at a Canadian station, with the detected changepoints being in good agreement with documented times of changes.

A set of functions for implementing the transPMFred algorithm to detect changepoints in non-zero daily precipitation amounts were developed and made available online free of charge, along with a quantile matching (QM) algorithm for adjusting shifts in non-zero daily precipitation series, which should work well in absence of any discontinuity in the frequency of precipitation measured (i.e., frequency discontinuity) and should work well for continuous variables such as daily temperature series.

However, frequency discontinuities are often inevitable, especially in the measurement of small precipitation due to changes in the measuring precision etc. Thus, it was recommended to use the transPMFred to test the series of daily precipitation amounts that are larger than a threshold, using a set of different small thresholds, and to use the PMFred algorithm to check the homogeneity of the series of monthly or annual frequency of various small precipitations measured. These would help gain some insight into the characteristics of discontinuity. It was also noted that, when a frequency discontinuity is present, adjustments derived from the measured daily precipitation amounts, regardless of how they were derived, could make the data deviate more from the truth. In this case, one shall adjust for the frequency discontinuities first.