Uncertainty in homogenized daily temperatures and derived indices of extremes illustrated using parallel observations in Canada

Abstract

Parallel daily temperature observations at site pairs over a 5-year period at 88 locations across Canada were used to derive and validate adjustments required during homogenization process. The data was first 'aligned' for compatible observing times at 12 locations (other locations do not have this problem). Then the homogenization adjustments were obtained using three procedures (Seasonal Bias, Monthly Interpolation and Quantile Matching) and two approaches (using parallel and neighbours observations). The root mean squared error (RMSE) between the daily temperatures of site 1 and site 2, and the percentage of days within 0.5 ∘ C (PD05) between site 1 and site 2 were used to assess the uncertainty in the mean and extreme values, respectively. The instruments were not necessarily collocated as the distance between the two observing sites varied from 0 to 30 km. The results confirm that it is necessary to apply adjustments for known issues first, such as a different observing time. They also show that when a shift between site 1 and site 2 (defined by the annual mean of the daily temperature differences) is small [<0.25 standard deviation (SD)], the adjustments do not reduce the error between site 1 and site 2. When the shift size is between 0.25 and 0.5 SD, the adjustments derived from parallel observations help to reduce the uncertainty. When the shift is large (>0.5 SD), both approaches reduce the error, although the adjustments derived from parallel observations provide better results as compared to those computed from neighbour observations. The results also indicate that Quantile Matching adjustments can provide a better estimate of the adjustments than the other methods evaluated to indices of extreme temperature computed from the adjusted daily values; however, highly correlated neighbours are needed when the adjustments are based on neighbours observations.

Comparison of homogenization methods for daily temperature series against an observation-based benchmark dataset

“…A reliable comparison of homogenization methods requires a set of metrics, defined in advance. The most commonly used metric is the relative RMSE (rRMSE) (Mestre et al 2011;Venema et al 2013;Domonkos 2013;Trewin 2013;Vincent et al 2018;Domonkos and Coll 2017) which allows to measure the distance between the homogenized and the benchmark version of a series. Then, an average of the results on all the benchmark series and an analysis of their distribution allow to compare all the tested methods.…”

“…In several cases, the comparison among methods has been performed as a validation for newly introduced homogenization methods (Caussinus and Mestre 2004;Menne and Williams 2005;Della-Marta and Wanner 2006;Mestre et al 2011;Trewin 2013;Vincent et al 2018). On the other hand, some outstanding studies have inspected the efficiency of automated, semi-automated, and manual homogenization procedures (Venema et al 2013;Domonkos 2013;Li et al 2016), focusing on break detection skills and on the performance on monthly data (Venema et al 2013).…”

“…The temporal frequency and amplitude used for the insertion of such events is chosen following ad-hoc studies, in order to reproduce as much as possible realistic situations. An evolution of this approach considers the creation of ensembles (Domonkos 2011;Vincent et al 2018;Domonkos and Coll 2017) of perturbed test series. This is done by varying amplitude and frequency of the breaks, allowing to test the efficiency of homogenization methods under particular conditions.…”

“…This is done by varying amplitude and frequency of the breaks, allowing to test the efficiency of homogenization methods under particular conditions. Though such an approach is not able to reproduce all the possible artificial signals that might be introduced to inhomogeneous series (Vincent et al 2018), giving more relevance to those signals that are easier to be studied and recognized. Moreover, the events and phenomena that can affect a recording station are so various, complex, and unexpected, that recent works have considered the opportunity to rely only on validated data (instead of simulated ones) for the creation of the test series (Venema et al 2013;Gubler et al 2017;Vincent et al 2018).…”

“…The distance between the homogenized versions and the benchmarks, evaluated with transparent and clear metrics that must be defined in advance, indicates the performance of the homogenization. The most common used metrics involve calculations of relative RMSE (rRMSE) (Mestre et al 2011;Venema et al 2013;Domonkos 2013;Trewin 2013;Domonkos and Coll 2017;Gubler et al 2017;Vincent et al 2018), changes in the trends of indices (Domonkos 2013;Venema et al 2013;Pérez-Zanón et al 2015;Domonkos and Coll 2017), and countings of number of days within fixed absolute thresholds around benchmark values (e.g., PD05, see Section 2.4) (Trewin 2013;Vincent et al 2018). Finally, the calculation of network mean biases of the trends (Domonkos 2013;Domonkos and Coll 2017) are a good indication of systematic behaviors related to certain homogenization methods and provide a good general estimation of the performance.…”