1. Introduction
Precipitation is an important climate variable that has a huge impact on our lives in terms of drinking water availability, agriculture, and the risk assessment of floods. A better understanding of the global water cycle and its evolution over the last 100 years is thus of great interest. The assessment of a parameter such as precipitation, which is spatially and temporally very variable, requires a particularly reliable data basis.
It is necessary to have highly reliable gridded datasets with a high spatial and temporal resolution over a long period of time. This can be provided either through multidecadal reanalysis or through the reprocessing of in situ data. Reanalyses are calculated from observations by means of a numerical weather forecast model. Based on the observations, a complete gridded dataset is created, in which the individual climate variables are physically consistent with each other. Global reanalyses have been produced by the European Centre for Medium-Range Weather Forecasts (ECMWF; Dee et al. 2014), the National Centers for Environmental Prediction and the National Center for Atmospheric Research (NCEP–NCAR reanalysis; Kalnay et al. 1996), the National Oceanic and Atmospheric Administration and the Cooperative Institute for Research in Environmental Sciences (NOAA–CIRES reanalysis; Compo et al. 2011), the National Aeronautics and Space Administration [Modern-Era Retrospective Analysis for Research and Applications (MERRA) reanalysis; Rienecker et al. 2011], and the Japan Meteorological Agency (JRA reanalysis; Onogi et al. 2007; Kobayashi et al. 2015).
The European Reanalysis of Global Climate Observations (ERA-Clim) and ERA-Clim2 projects, headed by the ECMWF, worked on the research and development of a global, long-term reanalysis covering more than a century. The ECMWF is widely respected for its long-term experience in the field of reanalyses [see overview given at ECMWF (2018a)] and has produced well-known reanalyses such as ERA-15 (Gibson et al. 1999), ERA-40 (Uppala et al. 2005), and ERA-Interim (Dee et al. 2011). Thanks to achievements in the area of data rescue and based on newly digitized historical observations, it has recently become possible to evaluate increasingly large time periods and to create reanalyses that cover more than 100 years—for example, the Twentieth-Century Reanalysis (20CR) project of NOAA–CIRES (Compo et al. 2011). During the ERA-Clim and ERA-Clim2 projects, vast amounts of data were digitized, and the ECMWF twentieth-century (ERA-20C) reanalysis (Poli et al. 2013), which covers 111 years, was produced by ECMWF. This reanalysis is based on the ECMWF prediction model IFS version cy38r1 with a spatial resolution of approximately 125 km and covers the period of 111 years from 1900 to 2010. Only surface observations such as air pressure and marine winds were assimilated, together with sea surface temperatures and sea ice conditions (Poli et al. 2013, 2015). All other parameters including precipitation are therefore derived variables and must be verified with independent observations.
This paper focuses on the assessment of precipitation in the ERA-20C reanalysis. Precipitation is one of the most difficult parameters to predict, which is also shown by previous studies of earlier reanalyses: it has, for example, been shown in a study (Dee et al. 2014) that compares ERA-Interim with the Full Data Monthly Version 6 (FDM-V6; Schneider et al. 2014) by the Global Precipitation Climatology Centre (GPCC) that Africa and the intertropical convergence zone (ITCZ), or at least the anomalies of 2010 related to the period 1981–2010, are much drier in the reanalysis than in the FDM-V6. Simmons et al. (2010) provided a detailed insight into surface air temperature, humidity, and precipitation of the ECMWF reanalyses ERA-Interim and ERA-40 in comparison with several gridded in situ datasets. All of these comparisons are based on 12-months running means of precipitation anomalies. To start with, the reanalyses data were adjusted to the GPCC Full Data Monthly Version 4 (FDM-V4) product on a continental scale, so that the average for the period 1989–98 is the same. The study shows that in general, the FDM-V4 has similar interannual continental-scale variations compared to the ERA reanalyses. ERA-Interim is closer to the GPCC reference than ERA-40, and generally the Northern Hemisphere is better reproduced than the Southern Hemisphere, and there is better consistency for Australia than for South America and Africa, with a clear shift of the mean between GPCC and ERA-Interim starting in the 1990s. For the last decade, a decline of precipitation values was found in ERA-Interim relative to FDM-V4. Regarding precipitation, ERA-Interim and ERA-40 perform similar when compared to GPCC, although correlations are distinctly higher with ERA-Interim than with ERA-40, especially for Africa and South America. Precipitation between 55° and 80° in the Northern and Southern Hemispheres was analyzed by Behrangi et al. (2016), which, among other findings, showed that ERA-Interim is consistent with the observational datasets. Other studies that focus on precipitation mainly used the MERRA reanalysis (Rienecker et al. 2011; Bosilovich et al. 2008), and there have also been investigations of specific regions such as Antarctica (Bromwich et al. 2011).
A first assessment of the quality of precipitation data in ERA-20C was provided by Poli et al. (2015), who examined the results for Europe and North America more closely and compared the data with the GPCC in situ product FDM-V6 (Schneider et al. 2011). They are convincing for Europe, where the internal annual fluctuations are well represented and a strong improvement around 1945 has been demonstrated, with additional improvements for North America in 1925 and 1960. However, the data before 1925 still show large differences.
This paper provides a comparison of the monthly precipitation in ERA-20C and the gridded GPCC Full Data Monthly Version 7 (FDM-V7) and sheds light onto the strengths and weaknesses of the precipitation generated in ERA-20C. The GPCC FDM-V7 gridded product is based on the monthly land surface observations of in situ measurements between 1901 and 2013 with a spatial resolution of 0.5°, 1.0°, and 2.5°. For the validation of ERA-20C, the most appropriate resolution of 1° was used. These precipitation in situ data have not been assimilated into the ERA-20C model and thus allow for an independent evaluation of precipitation in the reanalyses.
The document outline is as follows: In section 2, the database for the evaluation is described, consisting of the aforementioned ERA-20C reanalysis and the gridded GPCC FDM-V7 product. The evaluation methods are presented in section 3 and the results are presented in section 4. The latter consists of two parts, the global evaluation and a summarized evaluation according to continents. Finally, there is a summarizing discussion in section 5.
2. Database
Multidecadal reanalyses provide long-term highly resolved climate information across multiple parameters that are physically consistent with each other. Recent improvements were possible because of enhancements in the data assimilation and numerical model systems to corrected and extended input data as well as to increasingly powerful computer systems, which allow for more calculations at a finer spatial and temporal grid. Nevertheless, reanalyses are not perfect and an exact analysis of their performance is essential to be able to interpret the data and to draw reliable conclusions. The uncertainty varies according to the time period and the area covered, but depends also on the weather and the forecast model used. To estimate the model uncertainty, the ERA-20C deterministic reanalysis is compared to independent surface observations, namely, the gridded precipitation product FDM-V7 provided by the GPCC.
A compilation of the characteristic features of the ERA-20C reanalysis and the GPCC reference dataset, which was used, is shown in Table 1. Moreover, both gridded datasets are presented in detail below.
Short description of both datasets used: ERA-20C reanalyses and GPCC’s FDM-V7.
a. ERA-20C deterministic
ERA-20C deterministic is an outcome of the project ERA-Clim, documented in Poli et al. (2013). This is Europe’s first attempt of a reanalysis covering the twentieth-century weather and constitutes a pioneering project carried out by the ECMWF. It was succeeded by other reanalyses such as the ECMWF twentieth-century reanalysis dataset (ERA-20CM; Hersbach et al. 2015), with forcing terms in the model radiation scheme that follow CMIP5 recommendations, or the Coupled ECMWF Reanalysis (CERA-20C; Buizza et al. 2018). The general aims of reanalyses are, amongst others, to reliably represent the global weather and to reproduce the evolution of the global climate, as stated in Poli et al. (2015). Poli et al. (2013, 2015) describe the development of ECMWF’s reanalysis, that is, the ERA-20C ensemble and the improved ERA-20C deterministic run. We are focusing on the ERA-20C deterministic run, which is abbreviated as ERA-20C in this paper.
The ERA-20C run is a pilot analysis for the ECMWF, which is based on surface observations only and which covers a time period of more than a century. This reanalysis represents a test bed, which makes it possible to examine the benefits of new model developments that have taken place since the production of the last reanalysis, such as enhanced representation of moist physics or improved data assimilation methods.
ERA-20C is based on the ECMWF forecast model IFS version cy38r1 with a horizontal spatial resolution of T159 or approximately 125 km. The surface forcings are the same as those in the final product ERA-5, currently calculated at the ECMWF (Poli et al. 2013). The deterministic run was completed in February 2014 (Poli et al. 2015) in 22 parallel streams, of which each had calculated five years of the reanalysis with one year of spinup. The configuration of the deterministic reanalysis is almost the same as for the ensemble runs, but with the following exceptions: there are now 91 atmospheric vertical model levels between the surface and 0.01 hPa and four soil layers, and ocean waves are accounted for with different frequencies and directions. The output is available with different layers for pressure, potential temperature, and potential vorticity units. In this paper, the precipitation at the surface is compared with the precipitation of FDM-V7. The combined ERA-20C streams cover a time period of 111 years (1900–2010), with a model step of 30 min, although the time step for the precipitation output is 3 h (and 6 h for some other parameters).
The four-dimensional variational data assimilation (4D-Var) scheme used has been described in detail in Poli et al. (2013). The only parameters that were assimilated are surface observations, namely, wind and pressure, global sea surface temperatures from HadISST 2.1.0.0 (Rayner et al. 2003), and sea ice conditions. The surface and mean sea level pressure observations are provided by the International Surface Pressure Databank (ISPDv3.2.6; (Compo et al. 2015) and the surface marine winds are provided by ICOADSv2.5.1 (Woodruff et al. 2011). Precipitation can be compared to the reanalysis independently, as such data were not assimilated. It has to be remembered that changes in the quality and quantity of the input datasets can cause breaks in the resulting time series of the reanalysis. The input data coverage is provided in Poli et al. (2013).
b. FDM-V7
The GPCC provides, among other products, a centennial analysis of precipitation gauge reprocessings: the FDM-V7 (Schneider et al. 2017). This multidecadal analysis is available on a regular 0.5°, 1.0°, and 2.5° latitude–longitude grid for the period from 1901 to 2013. The land surface precipitation product was prepared by interpolation of the anomalies, using a modified SPHEREMAP scheme (Willmott et al. 1985; Becker et al. 2013). This means that the gridded anomalies are added to the gridded long-term monthly average to obtain the final gridded product. This anomaly interpolation method or climate-aided interpolation (CAI) method is fully described by Willmott and Robeson (1995). The FDM-V7 dataset is based on more than 75 000 stations that provide at least 10 years of observations.
The FDM-V7 is part of the portfolio of observational gridded precipitation products produced by the GPCC on behalf of the WMO’s World Climate Research Programme (WCRP) and the Global Climate Observing System (GCOS). At the time of the ERA-CLIM2 project, the FDM-V7 and the Climatology V2015 (Meyer-Christoffer et al. 2015) were the most recent GPCC datasets and were therefore used in this study.
3. Methodology
For a global assessment of the ERA-20C reanalysis, we compared the precipitation analyses with the FDM-V7 reference analysis through metrics that were examined grid cell by grid cell. Moreover, we studied the quality of the ERA-20C reanalysis across the time period with metrics representing either particular continental scale areas, or we examined it at grid points where particular strong deviations occur. In doing so, the areas were adapted from those defined in Bromwich et al. (2011) and displayed in Fig. 1.
Six continental regions considered for evaluation loosely based on Bromwich et al. (2011). Color shows the annual ERA-20C climatology.
Citation: Journal of Hydrometeorology 20, 2; 10.1175/JHM-D-17-0239.1
The overlapping time period of ERA-20C and FDM-V7 is 1901–2010 (see Table 1). This comparison was carried out with a spatial resolution of 1°, which is one of the native resolutions of FDM-V7 (Schneider et al. 2014). As the native resolution of ERA-20C is T159, it was regridded on the FDM-V7 grid of the Meteorological Archival and Retrieval System (MARS; Maass 2018), applying a bilinear interpolation scheme (ECMWF 2018b). Therefore, when comparing parameters such as spatial variability, one should bear in mind that the ERA-20C reanalysis is associated with a model run with a lower spatial resolution. The difference in the spatial resolution between FDM-V7 and ERA-20C is relatively small at the equator but amounts to about a factor of 2 at 50° latitude because of the spectral native grid of the numerical model.
a. Legates-corrected data and uncorrected data
The FDM-V7 is a product based on precipitation gauge measurements. Precipitation gauges exhibit systematic undercatch depending on, for example, gauge type, precipitation form, or wind force (Sevruk 1996). The undercatch might cause false differences between ERA-20C and FDM-V7, especially in areas exposed to wind or dominated by snow (Auer 1992). Legates and Willmott (1990) introduced an empirical correction factor based on wind data, anemometer height, roughness length, gauge type, and precipitation form or temperature, among others. The correction factor is based on long-term mean monthly wind speed data including 722 terrestrial and 6972 ocean grid cells. The GPCC provides this factor on a monthly basis via its visualizer (https://kunden.dwd.de/GPCC/Visualizer), described by Schneider et al. (2014), which is consistent with the FDM-V7.
We applied this correction factor to the monthly data to compensate for the undercatch error on the interpolated FDM-V7 grid before comparing them to ERA-20C, in order to have a better understanding of the errors induced by undercatch, but we will subsequently not use this factor in the main part of the study.
b. Temporal resolution and statistical methods
ERA-20C was compared to the FDM-V7 with corrected undercatch at different temporal aggregations, namely, monthly, seasonal, and annual totals.
Areas that turn out to be particularly conspicuous in these investigations are then studied in detail. These could be mountainous regions, coastal regions, or even deserts. In addition, the quality and quantity of the underlying data can also have an impact.
To obtain a spatial understanding of the data, we combined grid points into continental groups. As shown in Fig. 1, the continental groups are roughly based on the division of Bromwich et al. (2011). The continuous comparison of the two datasets basically consists of two parts. To begin with, the grid points are examined individually, and temporal scores [frequency bias (FBIAS) and threat score] are calculated and displayed as boxplots (see Figs. 5 and 6). In the second step, temporal changes in the data quality are examined.
Both scores were calculated with different thresholds in order to determine the performance of the datasets at different precipitation intensities. Quantiles were calculated with continent-specific threshold values (Table 2), as those vary strongly regarding the rainfall totals.
In a second step, the continent-specific average spatial time series of both datasets were compared with each other at a monthly, seasonal, and annual time scale. For a better understanding of outliers, trends, and breaks, we applied a homogeneity test, the Craddock test (Craddock 1979; Böhm 1992; Peterson et al. 1998). Since the test is bias-independent, it was used to determine whether significant changes or breakpoints occurred in one or both datasets. The station density of the underlying pressure time series assimilated in ERA-20C as well as of the gauges used for FDM-V7 were further analyzed for a period around these events.
Although it would be common to do so, neither of the two time series was used as the reference time series. However, both time series are plotted side by side, since it is assumed that none of them are homogeneous. Shown in this way, it can be easily seen in which of the time series the irregularities occur.
4. Results
To present the results obtained from the methods described above, we will first discuss the undercatch problem of precipitation gauges and its impact on the evaluation. In a second step, we will then provide a global seasonal overview of the performance of the ERA-20C reanalysis, while paying particular attention to individual regions, which attract attention because of weak performance. Finally, we will summarize the results continent by continent.
a. Effects of the gauge undercatch error in the in situ data on the comparison
If the Legates correction factor (Legates and Willmott 1990) is applied to the FDM-V7 (shown in Fig. 2b), a better agreement between the FDM-V7 and ERA-20C is found in terms of bias and ratio than without the correction factor (Fig. 2a). This is especially evident during the winter in the Northern Hemisphere with the exception of Siberia. For example, the winter precipitation between Canada and the United States in the Rocky Mountains shows clear systematic differences. By applying the correction factor to the FDM-V7, these differences are clearly reduced. The stronger correction on the U.S. side of the border compared to the Canadian side leads to spatial homogeneous data on both sides of the border. When corrected and uncorrected FDM-V7 data for Siberia are directly compared with ERA-20C, the Legates correction has the opposite effect. In this case, the correction factor seems to overcorrect the data during the winter season. In addition, the systematic difference in Siberia between the datasets is probably not due to the type of precipitation. Siberia shows a much better agreement of the values in the winter than in the summer months. In the Southern Hemisphere, the influence of the Legates correction is much weaker, as the fraction of snow is smaller than in the Northern Hemisphere. In contrast to the Northern Hemisphere, the changes are less extensive, especially when comparing the difference between the two datasets.
Mean seasonal bias between ERA-20C and FDM-V7 for the period 1901–2010. (a) FDM-V7 original data as reference; (b) the Legates-corrected FDM-V7.
Citation: Journal of Hydrometeorology 20, 2; 10.1175/JHM-D-17-0239.1
Despite its overall positive impact on the comparison, the undercatchment correction factor is not applied in the further analyses, unless explicitly noted otherwise. This is due to the obvious overcorrections, in particular in the area of eastern Russia, leading to spatial inhomogeneities that should not impact data and results.
An alternative weather-dependent undercatch correction was provided by Schneider et al. (2017), based on Ungersbock et al. (2001) and Rubel and Hantel (1999), using synoptic weather reports. For many parts of Asia as well as for other regions, the correction factors in this new estimation are lower than the ones from Legates and Willmott (1990). This new approach was not available at the time when the analyses presented in this paper were carried out.
b. Global
1) Difference and ratio
The difference between the two datasets, ERA-20C and FDM-V7, is most prominent in high-rainfall areas, as shown in Fig. 2. This in particular concerns coastal, mountainous, and monsoon areas. Most of these regions show more precipitation in the ERA-20C data. Above all, this affects the Andes in South America, but also the Canadian west coast and the Drakensberg in South Africa. ERA-20C also shows higher precipitation for the Rocky Mountains in Canada and the northern part of the United States, but to a lesser extent. All of these areas are located close to the ocean. Examples for inland mountains causing higher precipitation in ERA-20C are the Tibetan Plateau and the Daxue Range in China, the latter especially during MAM. In eastern Russia the same pattern can be found, but it is not as pronounced as in the examples listed before: ERA-20C shows more precipitation in the Siberian highlands, while lowlands exhibit similar or less precipitation. On the other hand, parts of western Russia have less precipitation during DJF and during parts of SON in ERA-20C than in FDM-V7.
However, the strongest differences on an annual scale occur in the region of the ITCZ, as the ITCZ in ERA-20C is drier than in FDM-V7. This applies in particular to the ITCZ area in northern South America, as well as in northern Brazil and the more northern countries except the Andes, but also to the ITCZ region in Africa, except the Virunga Mountains, and to Indonesia. This supports the previous findings that ERA-20C generally shows more precipitation in mountainous regions. As the ITCZ is shifting during the seasons, so does the strong precipitation associated with the ITCZ, and therefore also the areas with high differences between ERA-20C and FDM-V7.
A location displacement of precipitation patterns between ERA-20C and FDM-V7 occurs mostly at mountains close to oceans. This can be seen, for example, in the Western Ghats at the Indian west coast, where the areas with highest precipitation amounts are more distant to the ocean in ERA-20C than in FDM-V7, especially during the summer (JJA) monsoon season. Similar phenomena occur at the Canadian east coast and in the Rocky Mountains or at the coast of New Guinea and in the New Guinea Highlands.
The consistency of ERA-20C and FDM-V7 in arid regions was analyzed by means of precipitation ratios, as the ratio is more sensitive to small amounts than the difference. FDM-V7 has more precipitation than ERA-20C in dry areas such as in the United States [e.g., California (JJA, SON)] and Mexico, the Sahara (MAM, JJA, and SON), Saudi Arabia (all year), the Taklamakan, and the Gobi in China and Mongolia (all year), and, as mentioned above, also in Siberia except in the highlands (DJF), as well as in Egypt during all seasons except spring (MAM). But in arid mountains, ERA-20C shows higher precipitation totals than FDM-V7, for example in the Kunlun Mountains in China (all year) and in central Australia (all year), or in mountainous areas in eastern Siberia (especially in MAM and JJA).
2) Correlation
The Kendall correlation coefficients show a mixed result (Fig. 3). Highest correlations are found in Europe during the winter months, exceeding 80%. Other regions with high correlations are the United States, especially the Rocky Mountains in autumn and winter, Japan, and eastern Brazil in DJF and MAM.
As in Fig. 2, but for the Kendall correlation coefficient.
Citation: Journal of Hydrometeorology 20, 2; 10.1175/JHM-D-17-0239.1
On the other hand, there also are areas with poor, sometimes even negative correlations. The largest areas are located in the Southern Hemisphere, but also in the deserts of the Northern Hemisphere, with variable spatial extent, depending on the season. From June to August, Africa in general has very low correlation values of less than 20%, except for a few areas such as South Africa. However, this can also be observed in other areas, such as in Mongolia. All these areas have a low station density. However, scores in Australia are also generally very low—between 0% and 20%—although the station density is higher than in other well-correlating regions.
Even mountainous regions such as the Andes and the Himalayas, which are very conspicuous when looking at the difference and the ratio, show a comparable correlation with the surrounding areas.
3) Trends in ERA-20C and FDM-V7
In this section, trends in both datasets and the differences between them were investigated. Both datasets are based on data collections that strongly vary in size across time and space. Therefore, the datasets are not homogeneous, and discrepancies are to be expected. To reduce the impact of the varying database, the trends were only shown for the period 1951–2010 in Fig. 4. The homogeneity of the trends will be discussed in section 4d(2).
Seasonal Sen trend for the time period 1951–2010. (a) The Sen trend found in the ERA-20C data; (b) the Sen trend found in FDM-V7. The black dots indicate a significance level of 0.025.
Citation: Journal of Hydrometeorology 20, 2; 10.1175/JHM-D-17-0239.1
A comparison of the two datasets shows that they have different trends in terms of magnitude and sign. Furthermore, areas with significant trends only partially overlap between the datasets.
Similar trend patterns are found across Europe in all seasons in FDM-V7 and ERA-20C, such as the well-known dipole of decreasing precipitation amounts in the south and increasing amounts in the north. However, the magnitudes, as well as the line between positive and negative trends, differ. For Australia, the trends are similar in sign but not in magnitude for all seasons except SON. For this season, FDM-V7 has a negative and ERA-20C a positive trend in the southeast of Australia, Tasmania, and New Zealand. In addition, trends in Southeast Asia including China and India have the same sign, with strong positive trends in southeast China and northeast India in JJA, indicating increasing monsoon precipitation amounts and possibly increasing tropical cyclone activity. An opposite sign in these regions is found for SON, which could be caused by different lengths of the monsoon season in the datasets. Similar trend tendencies are found in North America for SON and DJF, with increasing precipitation amounts in the Southeast, potentially caused by increased hurricane activity, and decreasing amounts at the Pacific coast. In the other two seasons, however, the trends differ. For South America, the trend patterns of the datasets differ across many regions, where opposite trends are found for MAM and DJF. Except for the Sahara, which has only minor trends, the trends in Africa differ between the datasets except for MAM. During this season, the regions with the same trend direction are larger than those with opposite trends. Opposite trends are also found in JJA for the region influenced by the African monsoon, where FDM-V7 shows decreasing precipitation totals and ERA-20C shows increasing precipitation totals.
If longer periods such as 1901–2010 are investigated, the differences between the trends of ERA-20C and FDM-V7 are much greater (not shown). However, most trend differences coincide with breaks in the time series, detected by the Craddock test, either in one or in both of the datasets. The Gydan Peninsula and the Taymyr Peninsula in Russia show breakpoints in the FDM-V7 data in about 1953. In the Brazilian Amazon region, an area with sparse station density, a break was found in the annual data in about the year 1970 in the FDM-V7 dataset and a smaller one in ERA-20C in about 1935. Moreover, there are breaks in the MAM time series of FDM-V7 in 1920 and ERA-20C in 1942. Across the Sahara, the two datasets show large differences between each other during the first half of the last century until about 1950, and even less agreement before 1920. The ERA-20C time series is almost constant with a small break around 1962, while FDM-V7 exhibits two breaks in about 1920 and about 1945. The trend differences in Japan are caused by a break in the ERA-20C data in 1946.
c. Conspicuous areas
Overall, areas with strong biases are mountainous regions with their upwind and downwind effects, and coastal areas (Fig. 2). The monsoon regions exhibit strong biases throughout their series and severe shifts in the means. Some of the mountain effects seem to be related to the precipitation phase, especially in the Rocky Mountains, as they can be reduced by applying the Legates correction on the GPCC products.
The number of applicable stations also influences the difference between the datasets other than the mentioned Legates correction. For example, the Alps region, which has many available stations in both datasets, shows a good agreement without Legates correction, in particular from 1930 onward. On the other hand, Tibet shows practically no agreement, as only very few measurements were available for the generation of the compared datasets.
To look at the temporal evolution of heavy rainfall independently from the amplitude error, the scores mentioned above were also calculated for percentiles. For the 75th percentile, areas with a dense observation network again show better scores than regions with sparse data (not shown). The worst areas are those for which no observational data are available: Tibet, the border area between Niger and Chad, and the Brazil area close to Bolivia. For these regions, the score shows negative values for the period 1901–2010, which means that the agreement of the two datasets is worse than random. Looking at the second half of the period only, there are fewer inconsistencies, and Tibet in particular displays negative values only during the DJF season.
d. Continents
For better comparability, the continental areas as defined in Bromwich et al. (2011) were used (Fig. 1). The defined areas are North America, South America, Europe, northern Asia, Africa, and India and the monsoon area. In this paper, many areas close to the continental borders were not considered. This is true in particular for Mexico, Australia, and the northern coast of Africa.
1) Spatial scores
Contingency table scores were calculated for these six continental areas. For each grid point, TS and FBIAS were calculated with annual, seasonal, and monthly values for ERA-20C, with FDM-V7 as observations. Figure 5 shows the annual scores for different thresholds. These thresholds are the percentiles from 10% to 90% in 10% steps for the individual areas. Table 2 shows the thresholds of the percentiles.
Boxplot of gridpoint-wise annual scores for the continents as marked in Fig. 1 for the period 1901–2010 of ERA-20C and FDM-V7 for different quartiles as thresholds. The boxes display the 0.25th and 0.75th quartiles and hence 50% of the values. The whiskers extend to the lowest and highest values within a 1.5 interquartile range from the box. The circles indicate all values outside the whiskers’ range. (a) The TS; (b) the FBIAS.
Citation: Journal of Hydrometeorology 20, 2; 10.1175/JHM-D-17-0239.1
As the TS is positively oriented, Fig. 5a shows the best agreement for the lower quantiles in India, Europe, and South America. The scores for Africa are at first similar to those of the other continents, though with higher variability, but decline above the 50% threshold. The performance for all other areas decreases more slowly. The FBIAS shows precipitation situations that are over- and underforecast. Most areas show good agreement at the small percentiles, with the exception of Africa, where all values are underforecast. Northern Asia, Europe, and North America show an increasing overforecast with higher thresholds, though much weaker in North America. Africa and India show an underforecast at higher thresholds, which is strongest for India where ERA-20C does not capture the high precipitation measurements. South America shows an almost balanced performance.
Figure 6 depicts the TS and Fig. 7 shows the FBAIS for JJA (Figs. 6a,7a) and DJF (Fig. 6b,7b). Regarding the TS, most areas show little to no seasonal differences in quality, though Europe’s scores are worse during the winter. Africa shows better scores for seasonal than for annual totals.
Threat score as in Fig. 5a, but for (a) JJA and (b) DJF.
Citation: Journal of Hydrometeorology 20, 2; 10.1175/JHM-D-17-0239.1
FBIAS as in Fig. 5b, but for (a) JJA and (b) DJF.
Citation: Journal of Hydrometeorology 20, 2; 10.1175/JHM-D-17-0239.1
Seasonal differences are more pronounced in the FBIAS than in the TS. In general, the scores show similar patterns but with a different magnitude. During DJF, the continental results are more balanced than during JJA. In North America, the ERA-20C values are too high for both seasons. Northern Asia shows an almost perfect score in DJF, but an overforecast in JJA. The exception of the 90th percentile, already known from the annual scores, occurs only in DJF in India, where the ERA-20C values are far too low. Only Africa has values that are too low in all percentiles; however, the results are better during DJF.
2) Time period
Grid points were investigated individually, without taking account of the temporal evolution of the data in section 4d(1). In this section, area mean values were analyzed for the regions described in Fig. 1, which means that the spatial variability is not taken into account. Figure 8 depicts the annual mean time series for the mentioned regions from both datasets.
Mean annual time series for the period 1901–2010. The areas are defined in Fig. 1.
Citation: Journal of Hydrometeorology 20, 2; 10.1175/JHM-D-17-0239.1
The annual series show a very good agreement for Europe. For South America, from 1944 onward, the difference between the two datasets is smaller than before. North America and northern Asia exhibit more precipitation in the ERA-20C reanalysis than in the FDM-V7, but with a rather constant bias. In contrast, the ERA-20C reanalysis underestimates the precipitation across Africa and India/the monsoon area.
The magnitude of the differences between the datasets differs partly with the seasons (not shown). For South America, the main differences occur in the first half of the last century during winter (JJA). In the annual data, northern Asia shows a strong bias between ERA-20C and FDM-V7 during all seasons with the exception of DJF in the latter half of the century, which shows good agreement. Across North America, the bias appears mainly during MAM and JJA, whereas during SON and DJF, the precipitation means are of the same order of magnitude until 1942. Before 1940 and after 1970, FDM-V7 and ERA-20C have similar precipitation totals in Africa during SON, but diverge during the other seasons.
To take a closer look at these averaged precipitation series, the Craddock test as described in section 3 was applied in Fig. 9. As before in Fig. 8, the focus is on changes in the bias of the series. This test is a visualization tool that highlights different trends or breakpoints between series and makes it possible to compare these characteristics between ERA-20C and FDM-V7. Given that the average total precipitation has been subtracted, the test is not sensitive to the systematic errors that are obvious in Fig. 8. Other points of attention are the steps in the Craddock test result. These usually indicate the existence of outliers, regardless of whether they are of natural or artificial origin.
Seasonal Craddock test of the mean time series of the areas described in Fig. 1 for the time period 1901–2010. (a) JJA; (b) DJF.
Citation: Journal of Hydrometeorology 20, 2; 10.1175/JHM-D-17-0239.1
Figure 9 shows the seasonal results for JJA and DJF. Since South America (DJF) and India/the monsoon area (JJA) have higher precipitation totals, the results of the Craddock test also show higher values for this area. In general, it is evident that ERA-20C exhibits stronger changes in the mean than FDM-V7. The exceptions are North America in JJA and northern Asia in DJF. The Craddock plot for northern Asia reaches its maximum around 1950. This is distinctive in both datasets and results from a break in both time series. Since the beginning of the 1950s, FDM-V7 has contained data from more gauges, which might explain the break in the GPCC dataset.
In India/the monsoon area, there is a big step within the FDM-V7 in DJF in the later 1940s. This jump indicates two high values in the area-averaged time series, which are not included in the ERA-20C totals. A similar jump can be seen in South America in JJA between 1970 and 1980, but in that case, both datasets show a decade of high monthly totals. Especially in ERA-20C, these concur with breaks in the time series in 1940, 1970, and 1980 for South America in JJA.
Breaks in JJA in Africa are detected at different times in both datasets (Fig. 9a). After 1970, the number of underlying stations for the GPCC analysis is higher than before. A lot of these additional stations are located in the northern Sahel region, which moves the dry–wet border of the African monsoon farther to the south and reduces the area mean precipitation amount, which can also be seen in Fig. 8. Figure 8 also shows that the ERA-20C area mean precipitation drops around 1940 and then increases again to reach the level before this drop around 1970. These are the breaks seen in Fig. 9. We can only speculate on the reasons, which are probably associated with changes in the assimilated datasets.
The other differences between the two datasets are mostly related to breaks or short trends in the time series, but not to a series of different values as observed before. Such shifts occur in almost all areas. In Europe, the Craddock test shows that the timing of trends and outliers are very similar in both datasets. For annual data, the test indicates—as seen in the mean time series—that in the first half of the twentieth century, the FDM-V7 values are slightly higher than ERA-20C, but that they match more closely in the second half of the last century. The same is true for seasonal data, although JJA and DJF seasons agree much better than MAM and SON. The underlying station density is very high for both datasets. The FDM-V7 station density is never below 2000 stations and, after an increase in the 1950s and 1960s, even exceeds 10 000. The differences in the trends are much more pronounced for the other continents. In northern Asia, both datasets show breaks, but the breakpoint occurs in the same year (about 1950) only in DJF, but with a different magnitude (Fig. 9). After 1950, the series are pretty close, as can also be seen in the time series plot (Fig. 8). In JJA, both mean series show breaks in the mid-1950s, which are stronger in the FDM-V7 series. But a bias remains, as can be seen by comparing them with the JJA time series. In northern Asia, few stations are included in the analyses and there are never more than 2000. Their number increases between 1935 and 1965, which might explain the aforementioned breakpoints.
The creation of different production streams of the reanalysis every five years [see Table 3 in Poli et al. (2015)] does not cause breaks in the data on a monthly and seasonal scale (not shown). Only in annual data, very small breaks can be found in a 5-yr pattern, for example in India/the monsoon area. However, these are generally small breaks when taking into account the natural variability and compared to other artificial breaks. There are some strong breaks, such as in South America during DJF, that also coincide with the 5-yr pattern, while others, such as those in North America, occur in different years. It is therefore difficult to determine the reasons for these stronger breaks.
5. Discussion
Reanalyses provide unique opportunities to build a physically consistent global data grid over a long period of time. Although some very good results have already been obtained, there is nonetheless still room for more research to be undertaken to meet the high-level objectives and expectations of a reanalysis. It is difficult to reproduce the past with a database that still does not contain sufficient data for the beginning of the twentieth century. This is especially true for the Tibetan area. Because of the Himalayas, this is a challenging area for a model because of its orography, and therefore needs a reliable database. The quality of the GPCC’s product FDM-V7 database is also affected in some regions by the sparse station density.
Mountainous regions in general exhibit difficulties. This applies in particular to areas where ERA-20C shows much more precipitation than FDM-V7. This is especially noticeable in the Andes in South America. This effect is further enhanced when the mountains are located close to coastal regions, and in monsoon areas. Even in dry areas, ERA-20C shows more precipitation in mountainous regions, although FDM-V7 has in general more precipitation in dry areas. These findings also agree with Dee et al. (2014), who compared the ERA-Interim reanalysis with the GPCC Full Data Monthly Version 6 and noted that the anomalies of 2010 in relation to the period 1981–2010 showed that Africa or the ITCZ was much drier in ERA-Interim than in the GPCC data. If, however, the Kendall correlation coefficient is applied in the mountains (Andes, Himalayas), the values are not particularly high, but do not stand out from the surrounding areas. When comparing the precipitation of ERA-Interim and ERA-40 with the GPCC Product Full Data Monthly Version 4 (FDM-V4) on a continental basis, Simmons et al. (2010) found similar interannual variations. This pattern has also been detected when comparing ERA-20C and FDM-V7. Exceptions are Africa, with much higher values between February and September, and northern Asia, with a markedly smaller difference, but still higher values in summer.
A comparison between the Northern and the Southern Hemispheres reveals that the Kendall correlation coefficient is higher in the Northern Hemisphere than in the Southern Hemisphere. This coincides with Simmons et al. (2010), who also found that there was a better agreement in the Northern Hemisphere in the preceding reanalyses (ERA-15, ERA-40). In DJF, the highest correlations between ERA-20C and FDM-V7 are found in Europe (correlation > 80%), as well as in Japan and in large parts of the United States. Negative correlations between ERA-20C and FDM-V7 have been detected in some regions in Africa and some small areas in South America. When looking at the mean time series for all these areas, the consistency in Europe is also particularly good.
North America and northern Asia show more rainfall in the reanalysis than in the FDM-V7—however, with a rather constant bias. In contrast, Africa and India/the monsoon area show less rainfall. For northern Asia, one could assume that FDM-V7 shows less precipitation because of the undercatch of the precipitation gauge, which is especially pronounced when solid precipitation coincides with wind. However, this is not the case, since the bias is large in the summer and in the winter from 1943 onward.
One aim of producing reanalyses and of running an unchanged prediction model over a long time period is to achieve temporal homogeneity, which is needed for the evaluation of trends. ERA-Interim, for example, shows a decline of precipitation in comparison with FDM-V4 (Simmons et al. 2010). In addition, Simmons et al. (2010) found a clear shift of the mean of GPCC minus ERA-Interim between the 1990s and the last decade, which is presumably due to additional satellite data for ERA-Interim or to fewer underlying stations for the GPCC analysis. The evaluation of ERA-20C and FDM-V7 also shows breaks, which may make it impossible to evaluate any trends.
These breaks often occur in areas with scarce data and the breakpoints often coincide with a change in the underlying observation data, as they can be observed in both datasets. On a larger (continental) scale, all areas suffer from a number of breaks. Nevertheless, no dominant leaps could be found in the 1990s in ERA-20C and FDM-V7, also found by Simmons et al. (2010) for ERA-40 and ERA-Interim. Large breaks occurred mainly in 1948 and 1971 in Africa, in 1930, 1955, and 1971 in India/the monsoon area, and in South America in 1942. In Europe, both datasets show very similar trends and variability.
The setup of the reanalysis with different production streams every five years does not cause breaks in the data on a monthly and seasonal scale. Only in annual data, very small breaks can be found in a 5-yr pattern. These breaks are generally small compared to the natural variability and other artificial breaks.
The production of ERA-20C was meant to serve as a test bed for a centennial reanalysis (Poli et al. 2013). The lessons learned from ERA-20C and even those from this study were used to improve the successor reanalysis CERA-20C, which has become available in the meantime (Laloyaux et al. 2018). The GPCC is currently working on enlarging its database and improving the quality control of its data. New versions of the non-real-time products of the GPCC are released after significant updates of the database (assumed cycle of 3–4 years, as it was done in June 2018). The GPCC has switched to an alternative interpolation procedure for its non-real-time products, which uses climatological normals in regions with sparse data. This infilling would especially reduce the precipitation amount in dry regions at the beginning of the analysis, given that in those cases, there were weather stations only at the coast and not in the inland deserts. Taking this into account could reduce the differences between the two datasets in the first decades, but without further investigation, we cannot be more specific.
After all, general advice as to which dataset should be preferred cannot be given, as they both have their strengths, depending on the users’ application. It can be said that in general, the reliability of both datasets increases in parallel with the increase in the number of used stations. Users are encouraged to contact the authors of the datasets in order to choose the best set for their specific applications.
Acknowledgments
This work has been funded in part by the EU 7th Framework Program collaborative project ERA-CLIM2 (Grant 607029).
REFERENCES
Auer, I., 1992: Experiences with the completion and homogenization of long-term precipitation series in Austria. Central European Research Initiative, Project Group Meteorology Working Paper 1, 7 pp.
Becker, A., P. Finger, A. Meyer-Christoffer, B. Rudolf, K. Schamm, U. Schneider, and M. Ziese, 2013: A description of the global land-surface precipitation data products of the Global Precipitation Climatology Centre with sample applications including centennial (trend) analysis from 1901–present. Earth Syst. Sci. Data, 5, 71–99, https://doi.org/10.5194/essd-5-71-2013.
Behrangi, A., and Coauthors, 2016: Status of high-latitude precipitation estimates from observations and reanalyses. J. Geophys. Res. Atmos., 121, 4468–4486, https://doi.org/10.1002/2015JD024546.
Böhm, R., 1992: Description of the procedure of homogenizing temperature time series in Austria. Central European Research Initiative, Project Group Meteorology Working Paper 2, 14 pp.
Bosilovich, M. G., J. Chen, F. R. Robertson, and R. F. Adler, 2008: Evaluation of global precipitation in reanalyses. J. Appl. Meteor. Climatol., 47, 2279–2299, https://doi.org/10.1175/2008JAMC1921.1.
Bromwich, D. H., J. P. Nicolas, and A. J. Monaghan, 2011: An assessment of precipitation changes over Antarctica and the Southern Ocean since 1989 in contemporary global reanalyses. J. Climate, 24, 4189–4209, https://doi.org/10.1175/2011JCLI4074.1.
Buizza, R., and Coauthors, 2018: The EU-FP7 ERA-CLIM2 project contribution to advancing science and production of Earth system climate reanalyses. Bull. Amer. Meteor. Soc., 99, 1003–1014, https://doi.org/10.1175/BAMS-D-17-0199.1.
Compo, G. P., and Coauthors, 2011: The Twentieth Century Reanalysis Project. Quart. J. Roy. Meteor. Soc., 137, 1–28, https://doi.org/10.1002/qj.776.
Compo, G. P., and Coauthors, 2015: The International Surface Pressure Databank, version 3. Research Data Archive at the National Center for Atmospheric Research, Computational and Information Systems Laboratory, accessed 20 August 2017, https://doi.org/10.5065/D6D50K29.
Craddock, J. M., 1979: Methods of comparing annual rainfall records for climatic purposes. Weather, 34, 332–346, https://doi.org/10.1002/j.1477-8696.1979.tb03465.x.
Dee, D. P., and Coauthors, 2011: The ERA-Interim reanalysis: Configuration and performance of the data assimilation system. Quart. J. Roy. Meteor. Soc., 137, 553–597, https://doi.org/10.1002/qj.828.
Dee, D. P., M. Balmaseda, G. Balsamo, R. Engelen, A. Simmons, and J.-N. Thépaut, 2014: Toward a consistent reanalysis of the climate system. Bull. Amer. Meteor. Soc., 95, 1235–1248, https://doi.org/10.1175/BAMS-D-13-00043.1.
ECMWF, 2018a: Browse reanalysis datasets. ECMWF, accessed 8 June 2018, https://www.ecmwf.int/en/forecasts/datasets/archive-datasets/browse-reanalysis-datasets.
ECMWF, 2018b: Grid point to grid point interpolation. ECMWF, accessed 20 August 2018, https://confluence.ecmwf.int/display/EMOS/Grid+point+to+Grid+point+Interpolation.
Gibson, J., P. Kållberg, S. Uppala, A. Hernandez, A. Nomura, and E. Serrano, 1999: ERA-15 description, version 2. ECMWF ERA Rep. 1, 72 pp.
Hersbach, H., C. Peubey, A. Simmons, P. Berrisford, P. Poli, and D. Dee, 2015: ERA-20CM: A twentieth-century atmospheric model ensemble. Quart. J. Roy. Meteor. Soc., 141, 2350–2375, https://doi.org/10.1002/qj.2528.
Kalnay, E., and Coauthors, 1996: The NCEP/NCAR 40-Year Reanalysis Project. Bull. Amer. Meteor. Soc., 77, 437–471, https://doi.org/10.1175/1520-0477(1996)077<0437:TNYRP>2.0.CO;2.
Kobayashi, S., and Coauthors, 2015: The JRA-55 Reanalysis: General specifications and basic characteristics. J. Meteor. Soc. Japan, 93, 5–48, https://doi.org/10.2151/jmsj.2015-001.
Laloyaux, P., and Coauthors, 2018: CERA-20C: A coupled reanalysis of the twentieth century. J. Adv. Model. Earth Syst., 10, 1172–1195, https://doi.org/10.1029/2018MS001273.
Legates, D. R., and C. J. Willmott, 1990: Mean seasonal and spatial variability in gauge-corrected, global precipitation. Int. J. Climatol., 10, 111–127, https://doi.org/10.1002/joc.3370100202.
Maass, C., 2018: MARS user documentation. ECMWF, accessed 30 January 2019, https://confluence.ecmwf.int/display/UDOC/MARS+user+documentation.
Marchetto, A., 2014: rkt: Mann-Kendall test, seasonal and regional Kendall tests, version 1.3. R package, http://CRAN.R-project.org/package=rkt.
Meyer-Christoffer, A., A. Becker, P. Finger, B. Rudolf, U. Schneider, and M. Ziese, 2015: GPCC climatology version 2015 (at 0.25°, 0.5°, 1.0°, 2.5°): Monthly land-surface precipitation climatology for every month and the total year from rain-gauges built on GTS-based and historic data. Global Precipitation Climatology Centre hosted at Deutscher Wetterdienst (DWD), accessed 15 June 2017, https://doi.org/10.5676/DWD_GPCC/CLIM_M_V2015_100.
Onogi, K., and Coauthors, 2007: The JRA-25 Reanalysis. J. Meteor. Soc. Japan, 85, 369–432, https://doi.org/10.2151/jmsj.85.369.
Peterson, T. C., and Coauthors, 1998: Homogeneity adjustments of in situ atmospheric climate data: A review. Int. J. Climatol., 18, 1493–1517, https://doi.org/10.1002/(SICI)1097-0088(19981115)18:13<1493::AID-JOC329>3.0.CO;2-T.
Poli, P., and Coauthors, 2013: The data assimilation system and initial performance evaluation of the ECMWF pilot re-analysis of the 20th-century assimilating surface observations only (ERA-20C). ECMWF ERA Rep. 14, 62 pp.
Poli, P., H. Hersbach, P. Berrisford, D. Dee, A. Simmons, and P. Laloyaux, 2015: ERA-20C deterministic. ECMWF ERA Rep. 20, 53 pp.
Rayner, N. A., D. E. Parker, E. B. Horton, C. K. Folland, L. V. Alexander, D. P. Rowell, E. C. Kent, and A. Kaplan, 2003: Global analyses of sea surface temperature, sea ice, and night marine air temperature since the late nineteenth century. J. Geophys. Res., 108, 4407, https://doi.org/10.1029/2002JD002670.
Rienecker, M. M., and Coauthors, 2011: MERRA: NASA’s modern-era retrospective analysis for research and applications. J. Climate, 24, 3624–3648, https://doi.org/10.1175/JCLI-D-11-00015.1.
Rubel, F., and M. Hantel, 1999: Correction of daily rain gauge measurements in the Baltic Sea drainage basin. Nord. Hydrol., 30, 191–208, https://doi.org/10.2166/nh.1999.0011.
Schneider, U., A. Becker, P. Finger, A. Meyer-Christoffer, B. Rudolf, and M. Ziese, 2011: GPCC full data reanalysis version 6.0 at 1.0°: Monthly land-surface precipitation from rain-gauges built on GTS-based and historic data. Global Precipitation Climatology Centre hosted at Deutscher Wetterdienst (DWD), accessed 15 June 2017, https://doi.org/10.5676/DWD_GPCC/FD_M_V6_100.
Schneider, U., A. Becker, P. Finger, A. Meyer-Christoffer, M. Ziese, and B. Rudolf, 2014: GPCC’s new land surface precipitation climatology based on quality-controlled in situ data and its role in quantifying the global water cycle. Theor. Appl. Climatol., 115, 15–40, https://doi.org/10.1007/s00704-013-0860-x.
Schneider, U., A. Becker, P. Finger, A. Meyer-Christoffer, B. Rudolf, and M. Ziese, 2015: GPCC full data reanalysis version 7.0 (at 0.5°, 1.0°, 2.5°): Monthly land-surface precipitation from rain-gauges built on GTS-based and historic data. Global Precipitation Climatology Centre hosted at Deutscher Wetterdienst (DWD), accessed 15 June 2017, https://doi.org/10.5676/DWD_GPCC/FD_M_V7_100.
Schneider, U., P. Finger, A. Meyer-Christoffer, E. Rustemeier, M. Ziese, and A. Becker, 2017: Evaluating the hydrological cycle over land using the newly-corrected precipitation climatology from the Global Precipitation Climatology Centre (GPCC). Atmosphere, 8, 52, https://doi.org/10.3390/atmos8030052.
Sen, P. K., 1968: Estimates of the regression coefficient based on Kendall’s tau. J. Amer. Stat. Assoc., 63, 1379–1389, https://doi.org/10.1080/01621459.1968.10480934.
Sevruk, B., 1996: Adjustment of tipping-bucket precipitation gauge measurements. Atmos. Res., 42, 237–246, https://doi.org/10.1016/0169-8095(95)00066-6.
Simmons, A. J., K. M. Willett, P. D. Jones, P. W. Thorne, and D. P. Dee, 2010: Low-frequency variations in surface atmospheric humidity, temperature, and precipitation: Inferences from reanalyses and monthly gridded observational data sets. J. Geophys. Res., 115, D01110, https://doi.org/10.1029/2009JD012442.
Ungersböck, M., F. Rubel, T. Fuchs, and B. Rudolf, 2001: Bias correction of global daily rain gauge measurements. Phys. Chem. Earth Part B, 26, 411–414, https://doi.org/10.1016/S1464-1909(01)00027-2.
Uppala, S. M., and Coauthors, 2005: The ERA-40 Re-Analysis. Quart. J. Roy. Meteor. Soc., 131, 2961–3012, https://doi.org/10.1256/qj.04.176.
Wilks, D. S., 2006a: Spearman rank correlation and Kendall’s τ. Statistical Methods in the Atmospheric Sciences, 2nd ed. International Geophysics Series, Vol. 100, Academic Press, 55–57.
Wilks, D. S., 2006b: Scalar attributes characterizing 2 × 2 contingency tables. Statistical Methods in the Atmospheric Sciences, 2nd ed. International Geophysics Series, Vol. 100, Academic Press, 262–265.
Willmott, C. J., and S. M. Robeson, 1995: Climatologically aided interpolation (CAI) of terrestrial air temperature. Int. J. Climatol., 15, 221–229, https://doi.org/10.1002/joc.3370150207.
Willmott, C. J., C. M. Rowe, and W. D. Philpot, 1985: Small-scale climate maps: A sensitivity analysis of some common assumptions associated with grid-point interpolation and contouring. Amer. Cart., 12, 5–16, https://doi.org/10.1559/152304085783914686.
Woodruff, S. D., and Coauthors, 2011: ICOADS release 2.5: Extensions and enhancements to the surface marine meteorological archive. Int. J. Climatol., 31, 951–967, https://doi.org/10.1002/joc.2103.
