The development of Adapted Caussinus–Mestre Algorithm for homogenising Networks of Temperature series (ACMANT), one of the most successful homogenisation methods tested by the European project COST ES0601 (HOME) has been continued. The third generation of the software package ‘ACMANT3’ contains six programmes for homogenising temperature values or precipitation totals. These incorporate two models of the annual cycle of temperature biases and homogenisation either on a monthly or daily time scale. All ACMANT3 programmes are fully automatic and the method is therefore suitable for homogenising large datasets. This paper describes the theoretical background of ACMANT and the recent developments, which extend the capabilities, and hence, the application of the method. The most important novelties in ACMANT3 are: the ensemble pre‐homogenisation with the exclusion of one potential reference composite in each ensemble member; the use of ordinary kriging for weighting reference composites; the assessment of seasonal cycle of temperature biases in case of irregular‐shaped seasonal cycles. ACMANT3 also allows for homogenisation on the daily scale including for break timing assessment, gap filling and ANOVA application on the daily time scale. Examples of efficiency tests of monthly temperature homogenisation using artificially developed but realistic test datasets are presented. ACMANT3 can be characterized by improved efficiency in comparison with earlier ACMANT versions, high missing data tolerance and improved user friendliness. Discussion concerning when the use of an automatic homogenisation method is recommended is included, and some caveats in relation to how and when ACMANT3 should be applied are provided.

Sub-daily meteorological observations are needed for input to and assessment of high-resolution reanalysis products to improve understanding of weather and climate variability. While there are millions of such weather observations that have been collected by various organisations, many are yet to be transcribed into a useable format.Under the auspices of the Uncertainties in Ensembles of Regional ReAnalyses (UERRA) project, we describe the compilation and development of a digital dataset of 8.8 million meteorological observations of essential climate variables (ECVs) rescued across the European and southern Mediterranean region. By presenting the entire chain of data preparation, from the identification of regions lacking in digitised sub-daily data and the location of original sources, through the digitisation of the observations to the quality control procedures applied, we provide a rescued dataset that is as traceable as possible for use by the research community.Data from 127 stations and of 15 climate variables in the northern African and European sectors have been prepared for the period 1877 to 2012. Quality control of the data using a two-step semi-automatic statistical approach identified 3.5 % of observations that required correction or removal, on par with previous data rescue efforts.In addition to providing a new sub-daily meteorological dataset for the research community, our experience in the development of this sub-daily dataset gives us an opportunity to share some suggestions for future data rescue projects.All versions of the dataset, from the raw digitised data to data that have been quality controlled and converted to standard units, are available on PANGAEA: https://doi.org/10.1594/PANGAEA.886511 (Ashcroft et al., 2018).

Time series homogenization for 299 of the available precipitation records for the island of Ireland (IENet) was performed. Four modern relative homogenization methods, that is, HOMER, ACMANT, CLIMATOL and AHOPS were applied to this network of station series where contiguous intact monthly records range from 30 to 70 years within the base period 1941–2010. Break detection results are compared between homogenization methods, and coincidences with available documentary information (metadata) were analysed. The lowest (highest) number of breaks were detected with HOMER (ACMANT). Large differences of break frequency were found, namely ACMANT and AHOPS detected 8 times as many breaks than HOMER, while the break frequency with CLIMATOL was intermediate. Also, the ratio of series classified to be homogeneous varies widely between the methods. It is 85% with HOMER, 60% with CLIMATOL, 31% with AHOPS, while only 22% with ACMANT. In a further experiment, all the available time series for Ireland and Northern Ireland, (910 series) were used with ACMANT and CLIMATOL to explore the stability of break frequency for the same 299 series examined in the base experiment. While overall break frequency slightly increased (by 6–13%), the break positions notably changed for individual time series. The number of breaks changed for 59% (23%) of the series with ACMANT (CLIMATOL). For the breaks detected coincidentally by at least three methods including ACMANT and CLIMATOL in the base experiment, the second experiment confirmed the break positions in 86–87% of the breaks. The consequences of these results in relation to the reliability of statistical homogenization are discussed. Sometimes markedly different step functions provide comparable good approaches. However, the accuracy of homogenized time series cannot be related directly to the instability of break detection results.

The aim of time series homogenization is to remove nonclimatic effects, such as changes in station location, instrumentation, observation practices, and so on, from observed data. Statistical homogenization usually reduces the nonclimatic 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 nine versions, based on five homogenization methods—the adapted Caussinus-Mestre algorithm for the homogenization of networks of climatic time series (ACMANT), “Climatol,” multiple analysis of series for homogenization (MASH), the pairwise homogenization algorithm (PHA), and “RHtests”—are presented and evaluated. The tests were executed with 12 synthetic/surrogate monthly temperature test datasets containing 100–500 networks with 5–40 time series in each. Residual centered 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 nonclimatic 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), where 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. In general, ACMANT gave the most accurate homogenization results. In the accuracy of individual time series ACMANT is closely followed by Climatol, and for the accurate calculation of mean climatic trends over large geographical regions both PHA and ACMANT are recommended.

We assess the performance of different break detection methods on three sets of benchmark data sets, each consisting of 120 daily time series of integrated water vapor differences. These differences are generated from the Global Positioning System (GPS) measurements at 120 sites worldwide, and the numerical weather prediction reanalysis (ERA‐Interim) integrated water vapor output, which serves as the reference series here. The benchmark includes homogeneous and inhomogeneous sections with added nonclimatic shifts (breaks) in the latter. Three different variants of the benchmark time series are produced, with increasing complexity, by adding autoregressive noise of the first order to the white noise model and the periodic behavior and consecutively by adding gaps and allowing nonclimatic trends. The purpose of this “complex experiment” is to examine the performance of break detection methods in a more realistic case when the reference series are not homogeneous. We evaluate the performance of break detection methods with skill scores, centered root mean square errors (CRMSE), and trend differences relative to the trends of the homogeneous series. We found that most methods underestimate the number of breaks and have a significant number of false detections. Despite this, the degree of CRMSE reduction is significant (roughly between 40% and 80%) in the easy to moderate experiments, with the ratio of trend bias reduction is even exceeding the 90% of the raw data error. For the complex experiment, the improvement ranges between 15% and 35% with respect to the raw data, both in terms of RMSE and trend estimations.