Efficiency of Time Series Homogenization: Method Comparison with 12 Monthly Temperature Test Datasets

Abstract

The aim of time series homogenization is to remove non-climatic effects, such as changes in station location, instrumentation, observation practices, etc., from observed data. Statistical homogenization usually reduces the non-climatic effects, but does not remove them completely. In the Spanish MULTITEST project, the efficiencies of automatic homogenization methods were tested on large benchmark datasets of a wide range of statistical properties. In this study, test results for 9 versions, based on 5 homogenization methods (ACMANT, Climatol, MASH, PHA and RHtests) are presented and evaluated. The tests were executed with 12 synthetic/surrogate monthly temperature test datasets containing 100 to 500 networks with 5 to 40 time series in each. Residual centred root mean square errors and residual trend biases were calculated both for individual station series and for network mean series.The results show that a larger fraction of the non-climatic biases can be removed from station series than from network-mean series. The largest error reduction is found for the long-term linear trends of individual time series in datasets with a high signal-to-noise ratio (SNR), there the mean residual error is only 14 – 36% of the raw data error. When the SNR is low, most of the results still indicate error reductions, although with smaller ratios than for large SNR. Generally, ACMANT gave the most accurate homogenization results. In the accuracy of individual time series ACMANT is closely followed by Climatol, while for the accurate calculation of mean climatic trends over large geographical regions both PHA and ACMANT are recommended.

Combination of Using Pairwise Comparisons and Composite Reference Series: A New Approach in the Homogenization of Climatic Time Series

“…The efficiency tests of the European project COST ES0601 ("HOME" [19] found much lower efficiency of the PHA method than MULTITEST [5], and the PRODIGE method [20] was notably more accurate than PHA both for individual and network mean time series, in spite of PRODIGE includes an iterative removal of inhomogeneities. The large differences between the HOME results and MULTITEST results may have three reasons: a) The HOME benchmark is relatively small, its surrogate temperature section consists of only 15 networks, in addition, 9 of the 15 networks are small, comprising only 5 time series in each.…”

“…This technique may provide favourably high accuracy of homogenized data of individual time series, but it treats less effectively two problems: a) In iterative procedures the temporal evolution of individual series becomes more and more similar to each-other, but the convergence of the iteration might fail to find the true temporal variation of the climate signal, and this might result in relatively large errors in areaaverage trends; b) When a notable portion of time series is affected by similar inhomogeneities, both the averaging of neighbour series and the mentioned iterations have reduced efficiency in the inhomogeneity removal. In the efficiency tests of the Spanish MULTITEST project [5] the results of Climatol homogenization method [6] represented well all these attributes of composite reference series use with iterative removal of inhomogeneities. Note that the generally good results for individual time series and lower accuracy for area mean characteristics may seem a contradiction, which, however, only a seeming contradiction.…”

“…Menne and Williams [10] constructed an automatic homogenization method with pairwise comparisons, it is the Pairwise Homogenization Algorithm (PHA), and this has been applied to some large global or regional datasets [11][12][13]. [5] reported that PHA provides one of the most accurate regional mean trends among the tested homogenization methods, except for low SNR datasets. However, the homogenization accuracy for individual station series is notably lower with PHA than with some other tested methods.…”

“…for datasets with flat climatic trends the HOMER results were good or even excellent, while for datasets with enhanced climatic trends the HOMER results were poor. For this reason, we excluded HOMER from the publication [5], and thus ACMANT remained the only tested method including the ANOVA correction model.…”

“…In [5], the results for 12 temperature test datasets are published. Considering the 12 datasets and 5 efficiency measures mentioned, 60 pieces of results are compared for the tested methods, and generally a notable and statistically significant advantage of ACMANTv4 is shown by them.…”