1. Introduction
Observations during the twentieth century suggest that the increase in global air temperature cannot be solely attributed to natural climate variability (Hegerl et al. 2004); research also suggests an increase in global land precipitation during this time period (Ziegler et al. 2003). These changes are not only in the mean values but also in the occurrence of extreme meteorological events (e.g., floods, heat waves, and cold spells) (Zhang et al. 2011). Assessing the changes in the frequency and intensity of these events is necessary to inform stakeholders and provide adaptation strategies.
Changes in the distribution of meteorological variables (e.g., temperature, precipitation, and wind speed) can have a direct impact on a population’s food, water, energy, shelter, and transportation needs. For example, heavier rainfall events drive higher overland flow and an increase in flash flooding; prolonged extreme heat events affect human health, crop yields, and food security (Karl and Knight 1997; Schar and Jendritzky 2004) and a decrease in cold days leads to a decrease in water availability in snowpack fed regions (Barnett et al. 2005).
Global evaluations of climate extremes indicate that there has been an increase in the number of warm days and a decrease in cold days since 1950 (Donat et al. 2013a,c; Zhang et al. 2011). Wet spells produce higher rainfall totals nowadays than just a few decades ago, which have been coupled with an increase in frequency of heavy rainfall (Alexander et al. 2006; Donat et al. 2013c; Frich et al. 2002).
Although understanding the changes in global patterns of extreme events is important, there is also a need to assess regional and local impacts. One area of particular interest is sub-Saharan Africa (here we focus on the latitudes between 5°S and 25°N), as shown in Fig. 1. For this study, this region comprises the Sahel (semiarid transition between the Sahara and wetter regions of equatorial Africa to the south), the Greater Horn of Africa, the southern Sahara Desert, the northern Congo basin, and the southern Arabian Peninsula. The region is susceptible to high variability of water availability, because of distinct wet and dry seasons, and high interannual variability in rainfall that can manifest in devastating droughts and floods (Held et al. 2005). Understanding how the distribution of precipitation and temperature is evolving over the region, particularly its extremes, can potentially contribute to inform adaptation measures under a changing climate.
The entire domain (ED), which we define as sub-Saharan Africa, encompasses the northern Congo, Sahel, southern Sahara, southern Arabian Peninsula, and Horn of Africa. The domain is split into six subregions: northwest (NW), north (N), northeast (NE), southwest (SW), south (S), and southeast (SE).
Citation: Journal of Climate 27, 15; 10.1175/JCLI-D-13-00423.1
Few studies have analyzed climate extremes over Africa and the Arabian Peninsula because of the limited availability of long-term daily observations. New et al. (2006) analyze the trends in extremes of temperature and precipitation from station data over southern and western Africa between 1961 and 2000. Their results indicate a decrease in cold days and nights combined with an increase in warm days and nights. Over western central Africa and Guinea Conakry, Aguilar et al. (2009) report similar changes: a decrease in cold extremes and an increase in warm extremes. They also indicate a decrease in total annual precipitation and the amount of precipitation from extreme events. Analysis of precipitation over South Africa between 1910 and 2004 suggest a lack of significant regional changes (Kruger 2006). Over the greater Arab region, numerous studies report an increase in the frequency of warm days and nights and a decrease in cold days and nights over the second half of the twentieth century (Athar 2014; Donat et al. 2013b; Zhang et al. 2005). Over the same region Donat et al. (2013b) report an increase in precipitation in the west and a decrease in the east. These studies highlight statistically significant changes occurring over Africa and the Middle East but also the problems with inconsistent trends and lack of data.
Zhang et al. (2011) suggest that reanalysis model products might be a viable alternative for monitoring changes of extremes in data-sparse regions. However, the coarse resolution, biases, and spurious trends in these products can lead to errors in their use for studies of changes in climate means and extremes and evaluating impacts on hydrology and agriculture. Therefore, to ensure a reliable analysis of changes in extremes, special care must be taken to address these concerns, while at the same time providing the spatial coverage and resolution that are of practical use.
In this paper, we analyze climatic extremes over sub-Saharan Africa using a newly developed high-temporal- and high-spatial-resolution dataset of temporally consistent near-surface meteorology. The dataset is developed by downscaling global datasets, correcting for temporal inhomogeneities and merging with available station observations to improve local accuracy. The global datasets are taken from the Princeton University global meteorological dataset of Sheffield et al. (2006). The final dataset is developed at 0.1° daily resolution for 1979–2005.
2. Data and methods
a. Datasets
1) Princeton University global meteorological forcing dataset
The dataset is based on the Princeton University global meteorological forcing (PGF) dataset (Sheffield et al. 2006), which consists of 3-hourly, 1.0°-resolution fields of near-surface meteorology for global land areas for 1948–2010. The dataset is used primarily for forcing land surface models and other terrestrial models but also for analysis of climate variability and change. The dataset merges data from the National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP–NCAR) reanalysis (Kalnay et al. 1996) with the Global Precipitation Climatology Project (GPCP; Adler et al. 2003) and Tropical Rainfall Measuring Mission (TRMM) Multisatellite Precipitation Analysis (TMPA; Huffman et al. 2007) observation-based datasets of precipitation, the Climatic Research Unit (CRU) temperature dataset (New et al. 2000; Harris et al. 2013), and the Surface Radiation Budget (SRB) radiation dataset (Stackhouse 2004). This dataset is used as the baseline to construct the high-resolution dataset used in this study and follows other successful uses of the dataset for regional studies of climate and hydrology (Demaria et al. 2013; Sheffield and Wood 2008; Troy et al. 2011; Wang et al. 2011). All variables (precipitation, temperature, pressure, downward surface shortwave and longwave radiation, specific humidity, and wind speed) are spatially downscaled to 0.1° resolution using bilinear interpolation (with corrections to temperature, humidity, and downward surface longwave radiation based on changes in elevation; see Sheffield et al. 2006). Our focus here is on precipitation, temperature, humidity, and wind speed, given the lack of station observations of surface radiation for the region.
2) Station data
In situ station data are taken from the U.S. National Climatic Data Center (NCDC) Integrated Surface Database (ISD), which comprises worldwide surface weather observations from approximately 20 000 stations historically (Smith et al. 2011). This study uses the Global Summary of the Day (GSOD) dataset that summarizes the ISD data into a global daily product from 1929 to the present (
Although each station undergoes extensive quality control, errors remain in the data because of instrument error and undetected station relocation, among other reasons. As a result, previous studies have implemented additional quality-control tests to provide a higher-quality dataset [Hadley Centre Integrated Surface Database (HadISD)] for climate studies (Dunn et al. 2012). To maximize the number of in situ data, minimize error, and provide a framework for real-time updates, we also implement additional tests for outliers and inhomogeneities [discussed in section 2b(2)] to provide a higher-quality dataset.
b. Methods
1) Detection and correction of temporal inhomogeneities
The datasets that contribute to the PGF gridded product contain biases and spurious trends because of changes in instruments, observing practices, station location, station environment, satellite sensors, and network configuration and density (Klein Tank et al. 2009; Peterson et al. 1998; Reeves et al. 2007). These alterations can create artificial temporal inhomogeneities in the time series, including shifts in mean and variance that propagate through to the gridded datasets upon interpolation. When possible, these inhomogeneities must be identified and removed to not distort the analysis of climate extremes (Zhang et al. 2011).
Different approaches to address the homogenization problem have been proposed and developed (for a recent discussion, consult Venema et al. 2012, among others). In particular, we can highlight two main approaches, absolute and relative homogenization. In the former, abrupt changes are detected based only on the information in the record of interest. In the latter, we examine the presence of shifts with respect to a reference (homogeneous) time series. While the relative approach may be preferable, it suffers from the shortcoming that a reference record located near the site of interest needs to be identified. Because of this and because we deal with extremes and data from skewed distributions, we resort to the nonparametric Pettitt test (Pettitt 1979) to detect step changes in the downscaled PGF dataset. This test detects a single abrupt change (and its corresponding statistical significance) in the mean of the distribution of the variable of interest at an unknown point in time. The successful use of the Pettitt test in several previous hydrometeorological studies serves as a basis for using it here (Busuioc and von Storch 1996; Ferguson and Villarini 2012; Villarini et al. 2009).
Before applying the Pettitt test to the annual averaged temperature, precipitation, wind speed, and specific humidity at each grid cell in the domain, the series is detrended with respect to the Kendall–Theil robust line (Theil 1950; Ziegler et al. 2003). For the Pettitt test, we set a significance level of 1%. After detecting the step change in the annual time series, the data are homogenized. The cumulative distribution matching technique (Gao et al. 2007; Reichle and Koster 2004) is used to match the distribution of the data after the step change to the data before the step change (the ordering is chosen for consistency). The homogenized data are used to scale the original 3-hourly dataset over the domain between 1979 and 2005.
2) Merging the station data using state-space estimation
The global gridded PGF dataset merges reanalysis data with observation-based gridded datasets, such as the CRU data. The latter are derived using a limited number of stations whose number and location vary in time, which are particularly sparse over sub-Saharan Africa (on average 0.0153 stations per 0.5° × 0.5° grid cell). The observation-based gridded dataset products lose temporal and spatial information as they are interpolated and upscaled to coarser temporal (1 month) and spatial (0.5°–1.0°) resolutions. As a result, the original downscaled PGF dataset rarely agrees with the in situ measurements.
Although still sparse over sub-Saharan Africa, there are many more stations in the GSOD dataset than those used in the coarser data products. After additional quality checks, the stations are assimilated into the PGF dataset to improve its accuracy in space and time. The quality checks consist of removing detectable outliers [values outside the acceptable regional and global ranges defined in Dunn et al. (2012) and Durre et al. (2010)] and discarding stations with inhomogeneities (detected using the Pettitt test with a significance level of 1%). Acceptable stations must have a 20%+ temporal coverage between 1979 and 2005. Finally, to ensure temporal consistency, candidate stations are gap filled when necessary (for details, see the appendix).
The method used to assimilate the station data is as follows:
Let xbj be the value of the 0.1° downscaled PGF global gridded dataset (or background field) at a grid cell at time t and ymi and hxbi be all in situ observations and their corresponding collocated gridcell values in the vicinity of xbj for the variable at the same time step.
Let the vector of departures d be the difference between y, the vector of observations in the vicinity of the grid cell, and 
This implies that any point within the grid cell represents the gridcell spatial average. The combination of daily time step and 0.1° spatial resolution in this study make this assumption reasonable for meteorological variables (Beek et al. 1992).
Following the Hollingworth–Lonnberg method (Hollingsworth et al. 1985), 
Constructing the sample semivariogram at each time step requires a high network density, which is not possible in data-sparse regions, such as sub-Saharan Africa. To overcome this, in this study, space is traded for time and the climatologies of the covariance function are developed for each grid cell in the domain (see also Van de Beek et al. 2011). All stations within a radius r of a grid cell are used to construct the covariance function. We found that a value of r = 1000 km is an appropriate compromise between using as many stations (minimum of 20) as possible while ensuring that the spatial homogeneity and stationarity assumptions are still reasonable. When a grid cell does not meet these conditions, inverse distance weighting is used to spatially interpolate the covariance function parameters.
c. Indices for evaluating changes in extremes
The added value in the development of a bias-corrected, downscaled, temporally homogenous dataset is a higher degree of confidence in the analysis of daily meteorological extremes. Merging the in situ measurements provides the benefits of the local measurements and of the ability to assess spatial patterns. Following previous studies, the Expert Team on Climate Change Detection and Indices (ETCCDI) indices are used to quantify the evolution of meteorological extremes over the domain between 1979 and 2005. These indices were developed to standardize the detection of climatological extreme events (Klein Tank et al. 2009; Peterson et al. 2008). Table 1 describes the subset of ETCCDI indices used in this study. For a more detailed background on the indices, consult Zhang et al. (2011).
Summary of the temperature and precipitation extremes indices used in this study. The definitions are taken from Zhang et al. (2011). The precise definitions for all the indices are online (at http://etccdi.pacificclimate.org/list_27_indices.shtml).
d. Detection of trends
Trends in the extreme indices are calculated as the slope m of the Kendall–Theil robust line (Theil 1950). Previous studies of climate extremes suggest using this nonparametric approach to determine trends over simple linear regression (Aguilar et al. 2009; Peterson et al. 2008). We then compute the 99% confidence interval following the method described in Sen (1968).
Because of the presence of nonlinear trends, the Mann–Kendall test (Kendall 1975; Mann 1945) is also used to determine the statistical significance of the presence of monotonic patterns by determining the tendency for the data to increase or decrease monotonically in time. This test has been used previously to investigate hydroclimatological signals of climate change and variability (Hisdal et al. 2001; Lettenmaier et al. 1994; Lins and Slack 1999; Mitosek 1995; Ziegler et al. 2003). The nonparametric nature of the test and its ability to detect both linear and nonlinear trends makes it a valuable tool in assessing changes in climate extremes. We set a significance level of 1% for the Mann–Kendall test.
3. Results
a. Homogenization of the PGF dataset
The Pettitt test with a significance level of 1% is applied to the time series of annual values of the downscaled PGF dataset for each grid cell. Figure 2 shows the year when a step change is detected in specific humidity, wind speed, air temperature, and precipitation. Air temperature, specific humidity, and wind speed show similar step change years in the region surrounding Lake Victoria during the 1990s. Air temperature shows distinct step changes in the 1990s across the north and central sub-Saharan Africa. Specific humidity shows distinct shifts in the western sub-Sahara (mid to late 1990s) and north (mid to late 1980s). Wind speed shows a shift from the mid to late 1980s in the northwestern sub-Sahara (mid to late 1990s) and throughout the southern Arabian Peninsula. One noticeable feature is the limited number of step changes in precipitation. Temperature and, indirectly, specific humidity are the only variables corrected when CRU is merged with the original PGF dataset. This likely explains the correlation between these two variables and the differences with wind speed and precipitation.
The first year of the step change identified in time series of detrended annual averages from the PGF dataset for temperature, specific humidity, wind speed, and precipitation. Step changes are identified using a Pettitt test at the 1% significance level.
Citation: Journal of Climate 27, 15; 10.1175/JCLI-D-13-00423.1
Based on the step changes detected at each grid cell, the data are temporally homogenized as described in section 2b(1). An example of the homogenization is shown in Fig. 3. The Pettitt test detects the upward shift in annual temperature, which is then removed. The corrected annual values are then used to scale the 3-hourly data of the PGF dataset.
An example of the identification and removal of a step change in annual air temperature time series for the grid cell at 22°N, 5°E. The Pettitt test with a test significance level of 1% is applied to the annual temperature to detect a step change shown as the vertical line. The step change is removed by using the cumulative distribution matching technique.
Citation: Journal of Climate 27, 15; 10.1175/JCLI-D-13-00423.1
b. Gap filling the GSOD station data
The cumulative distribution function (cdf) matching technique discussed in section 2b(1) is applied to all candidate station’s collocated gridcell time series for daily mean wind speed, minimum and maximum temperature, and specific humidity. The Kolmogorov–Smirnov statistic is used as a metric to compare the distribution of each variable at each collocated grid cell to the distribution of the corresponding station (Mood et al. 1974). Through cdf matching, we aim to minimize Dn to ensure our collocated gridcell data come from the same distribution as the station data. Figures 4 and 5 provide an assessment of how well the bias-corrected collocated grid cells perform against the available station data.
The Kolmogorov–Smirnov statistic (left) before and (right) after the bias correction of (top)–(bottom) daily maximum temperature, daily minimum temperature, daily mean specific humidity, and daily mean wind speed. The bias correction is applied to the collocated grid cells of all qualifying GSOD stations using all available data for the station.
Citation: Journal of Climate 27, 15; 10.1175/JCLI-D-13-00423.1
Error statistics of the gridded precipitation dataset at collocated stations. Shown are (top)–(bottom) the Kolmogorov–Smirnov statistic, the ratio of mean number of rain days, the ratio of no rain–no rain transition probability, and the ratio of rain–rain transition probability of the (left) original and (right) bias-corrected gridded data compared to available collocated station data.
Citation: Journal of Climate 27, 15; 10.1175/JCLI-D-13-00423.1
As shown in Fig. 4, the technique effectively removes the biases in the collocated gridcell data of daily minimum temperature, maximum temperature, specific humidity, and mean wind speed over the GSOD network. The original distribution of both daily mean wind speed and specific humidity of collocated grid cells is in general more biased in the southern region of the domain. In all cases, after cdf matching the grid cell’s distribution closely resembles the station’s distribution.
The rain day statistics over the GSOD network domain before and after correcting for the mean number of rain days are shown in Fig. 5. The improvement in the first-order transition probabilities and the Kolmogorov–Smirnov statistic after cdf matching using a gamma distribution are also shown. As expected, the mean number of days and the distribution show the largest improvement while the transition probabilities only see a marginal improvement.
An example of the bias correction for all variables is shown in Fig. 6 for a station in Mali. The method effectively removes the biases in all the variables. The missing data are then infilled with the bias-corrected gridcell data. Note that the random errors in the original collocated gridcell data are not improved using this method.
An example of the bias correction of the time series of meteorological variables for the grid cell collocated with the station at Kenieba, Mali (12.85°N, 11.233°E). The homogenized and downscaled PGF gridded data are bias corrected using cdf matching. The rain day statistics are corrected prior to cdf matching. The missing time steps in the station data are then infilled with the bias-corrected data.
Citation: Journal of Climate 27, 15; 10.1175/JCLI-D-13-00423.1
c. Merging the station data into the gridded dataset
Following Chirlin and Wood (1982) and Hollingsworth et al. (1985), the spatial covariance functions of the departures are used to calculate the error background covariance matrix. Assuming that the measurement error is 0 (
In practice, 
Ability for surrounding stations to correct a given grid cell (min = 0; max = 1) for (top)–(bottom) daily minimum temperature, daily maximum temperature, specific humidity, mean wind speed, and precipitation for (left) January and (right) July. The influence of surrounding stations on a grid cell is calculated by summing all the weights of stations.
Citation: Journal of Climate 27, 15; 10.1175/JCLI-D-13-00423.1
The noticeable features are the mainly invariant spatial fields between the months. This suggests that the monthly climatology can be replaced with an annual mean with little decrease in performance. The relationship between the spatial correlation of the departures and their corresponding variables is apparent in the comparison between regions of high topographic relief (e.g., Rift Valley) and low topographic relief (e.g., Sahel). This is also apparent when comparing the precipitation departures to other variables.
Figure 8 shows an example of the assimilation of station data to correct the daily maximum temperature on 1 February 2000 in the Gabon region in central Africa. The background field provides the spatial distribution of the field while the station data provide the ground truth. After merging the stations into the gridded background field, the spatial properties of the background field are maintained. In the southwest of the region, where there is a higher density of stations, the merged gridded dataset shows significant differences with the original background field. In the northeast of the region there is a large decrease in the extent of the very warm region, while the northwest is left relatively untouched because of a lack of stations in the area.
An example of the assimilation of GSOD station daily maximum temperature data into the gridded PGF dataset for the Gabon region on 1 Feb 2000. (left) The downscaled and homogenized PGF dataset provides the spatial distribution of the temperature field, whereas (middle) the GSOD stations provide the ground truth observations at point scale. (right) The observations are merged into the background field by interpolating the departures using simple kriging and superimposing them onto the background field.
Citation: Journal of Climate 27, 15; 10.1175/JCLI-D-13-00423.1
We compare the annual averages between 1979 and 2005 of the variables of the different versions of the PGF dataset (original, homogenized, and bias corrected) over different regions in the domain. As shown in Fig. 9 and based on the change in the regional means (original PGF compared to bias-corrected dataset) shown in Table 2, the largest changes tend to occur after assimilating the GSOD station measurements except when the spatial coverage of in situ measurements is inadequate to have a large impact. For a majority of the regions, specific humidity is overestimated in the original dataset and wind speed and precipitation are underestimated. In high-density station regions in western Africa, the inclusion of the station data has a small effect on the daily maximum and minimum temperatures. This suggests that the CRU dataset that is used to scale the PGF dataset already used data from many of the stations merged into the homogenized dataset.
Comparison of regional annual averages of the original downscaled PGF dataset, the homogenized version, and the final bias-corrected version. Areal averages are estimated for each year and variable over the entire domain and the six subdomains defined in Fig. 1.
Citation: Journal of Climate 27, 15; 10.1175/JCLI-D-13-00423.1
Change in the mean of the regional average annual time series after homogenization and bias correction.
d. Trend analysis of daily extremes
Using the bias-corrected daily dataset (PGF-BC), the extreme indices are calculated for each year between 1979 and 2005 over the study domain. Figure 10 shows the estimated slope of the Kendall–Theil robust line over the domain for the ETCCDI indices used in this study, areas that show a statistically significant (1%) monotonic trend according to the Mann–Kendall test, regions affected by the assimilation of the GSOD stations, and homogenized grid cells. Estimated linear trends over regions where step changes are removed through homogenization should be disregarded because of the presence of nonlinearities and imperfections in the homogenization process.
The slope of the Kendall–Theil robust line of annual ETCCDI indices (see Table 1). The presence of a monotonic trend with a 1% significance level is shown by slash (/) line hatching. Grid cells that are impacted by the assimilation of GSOD data are shown by circular hatching and grid cells that are homogenized are shown by backslash (\) line hatching.
Citation: Journal of Climate 27, 15; 10.1175/JCLI-D-13-00423.1
The spatial trend information is complemented by the regional analysis as shown in Fig. 11 and their corresponding linear trends, confidence intervals, and statistical significance of a monotonic trend as shown in Table 3. In the following subsection, we summarize the results over the entire study region and discuss large regional differences.
The areal averages of the ETCCDI indices described in Table 1 for all the regions defined in Fig. 1 including the entire domain between 1979 and 2005.
Citation: Journal of Climate 27, 15; 10.1175/JCLI-D-13-00423.1
The trends (m) and corresponding 99% confidence intervals (±) of the regional annual time series of the ETCCDI indices as shown in Fig. 11. The trends are calculated using the Kendall–Theil robust line. The significance level (α) of the Mann–Kendall test is computed to determine the existence of monotonic trends (linear or nonlinear). The regions are defined in Fig. 1.
1) Temperature
An increase in warm days (TX90p) and warm nights (TN90p) has been accompanied by a decrease in cold days (TX10p) and cold nights (TN10p). There has also been an increase in the number of warm spells (WSDI) and a decrease in cold spells (CSDI). In most regions, the absolute value of the increase in warm days, warm nights, and warm spells is larger than the absolute value of the decrease in cold days, cold nights, and cold spells.
2) Precipitation
Over the entire domain there has been on average an increase in the total annual precipitation of around 50 mm. Although all regions have seen an increase, the changes have taken place mainly in the northwest, north, southwest, and southeast. The same tends to be true regarding the increase in occurrence of heavy (R95p) and very heavy rainfall events (R99p). The increase in the annual number of consecutive wet days (CWD) is statistically insignificant while there has been a more significant decrease in the annual number of consecutive dry days (CDD).
4. Discussion
a. Evaluation of the merging scheme over a high-density network: Oklahoma
Although the GSOD station data provide larger spatial coverage than other in situ observation datasets over sub-Saharan Africa, it can still be considered sparse and insufficient to validate the assimilation scheme used in this study. For this reason, we validate the scheme over the Mesoscale Network (MESONET) in situ network (McPherson et al. 2007) in the United States. The MESONET is a high-density, high-quality in situ network (110+ stations). A synthetic “truth” dataset is developed by merging the MESONET daily average temperature into the North American Land Data Assimilation System (NLDAS) gridded dataset (Mitchell et al. 2004) between 2000 and 2010.
To reproduce the conditions found over sub-Saharan Africa (network sparsity and changing network), we construct the semivariogram climatology of the departures for each month by allowing only 10 stations to be available, which we choose at random for each time step between 2000 and 2010. For a given number of stations, we extract a random subset of stations from the network and choose 50 random time steps in the 2000–10 time period. A set of different interpolation schemes is used to calculate the interpolated departures field to superimpose on the background field: inverse distance weighting (p = 2) (IDW), simple kriging (SK), ordinary kriging (OK), simple kriging with the semivariogram climatology (SK svclim), and ordinary kriging with the semivariogram climatology (OK svclim). We repeat the exercise for 50 different random subsets for a given number of stations. The normalized root-mean-square deviation (RMSD) is computed and averaged over the different time steps and network configurations.
The results are shown in Fig. 12. As expected, the original NLDAS-2 data are always the worst-case scenario (i.e., assimilating any error-free station data is better than none). As the network density decreases below 35 stations, inverse distance weighting performs better than both simple and ordinary kriging. This is due to the uncertainties in the constructed semivariogram model. However, when each time step’s semivariogram model is replaced with the precomputed semivariogram climatology, simple and ordinary kriging on average outperform inverse distance weighting regardless of the network density and configuration.
Mean normalized RMSE of assimilated daily average temperature fields compared to the baseline synthetic truth using a suite of assimilation schemes. Daily average temperature in the MESONET is merged into the NLDAS-2 gridded dataset over Oklahoma from 2000 to 2010. The network density and configuration of MESONET are allowed to change to determine the ability to reproduce the baseline.
Citation: Journal of Climate 27, 15; 10.1175/JCLI-D-13-00423.1
Assuming that these results are qualitatively applicable to sub-Saharan Africa, this assimilation scheme substantially increases the reliability and accuracy of the downscaled homogenized PGF dataset and outperforms more simplistic methods such as inverse distance weighting.
b. Validation of the high-resolution dataset and future improvements
We validate the bias-corrected dataset with the GSOD stations that were used in the assimilation (≥20% temporal coverage) and those that were not (<20% temporal coverage). For each of the station’s collocated grid cells, we compute the Kling–Gupta efficiency (KGE) for both the homogenized (before assimilation) and bias-corrected (after assimilation) datasets with the observations (Gupta et al. 2009).
In general, as shown in Fig. 13, the data from the stations that are assimilated into the dataset are replicated in the bias-corrected downscaled version. In the case of stations that were not assimilated, the results are mixed. Specific humidity shows the best performance in reducing the error at these sites. The other variables except for precipitation have some limited skill. Precipitation shows the worst performance with practically no skill in improving the nonassimilated sites. The errors in capturing the nonassimilated sites are most likely related to errors from infilling, low spatial correlation of the errors, and effects of multiple stations on a single point.
The KGE between the collocated grid cells of the (left) homogenized and (right) bias-corrected datasets. The sum of the weights annual mean that ranges from 0 to 1 is shown as shading underneath. The darker the shading the more influence the stations have on bias correcting the homogenized dataset. When the KGE value is equal to 1, the collocated gridcell data are equal to the station data.
Citation: Journal of Climate 27, 15; 10.1175/JCLI-D-13-00423.1
In future validation efforts, high-resolution satellite precipitation products should be used to assess the spatial properties of the downscaled precipitation product. We envision using a suite of product such as TMPA (Huffman et al. 2007), Precipitation Estimation from Remotely Sensed Information Using Artificial Neural Networks (PERSIANN; Sorooshian et al. 2000), Climate Prediction Center (CPC) morphing technique (CMORPH; Joyce et al. 2004), and Global Satellite Mapping of Precipitation (GsMAP; Kubota et al. 2007) These efforts could then implement techniques to incorporate the spatial properties of these products into the downscaled rainfall fields through conditional stochastic simulations methods such as lower upper (LU) decomposition and simulated annealing (Webster and Oliver 2007).
To validate the station data outlier detection scheme discussed in section 2b(2), we reprocessed the GSOD dataset using a more aggressive outlier detection scheme where we only accept values that are within 3+ standard deviations from the mean. Although not shown here, the differences after rerunning the bias correction, ETCCDI calculation, and ETCCDI trend analysis were negligible suggesting our results are robust to the outlier detection scheme. Following Dunn et al. (2012), these two outlier detection schemes would cover the range of being too liberal (outliers remain) and too aggressive (real tails removed). Future work could find a better balance by implementing the methods outlined in Durre et al. (2010) and Dunn et al. (2012).
c. Changes in climate extremes: Comparison to HadEX2
The recently released Hadley Centre Global Climate Extremes Index 2 (HadEX2) dataset provides a global product of ETCCDI indices derived from 7000 temperature and 11 000 precipitation meteorological stations. This product covers the period 1901–2010 and is created by calculating the indices at each station and then interpolating them onto a 3.75° × 2.5° grid (Donat et al. 2013c). Since this product partially covers this study’s spatial domain, it is suitable to validate the PGF-BC derived ETCDDI indices and to assess the added value of higher spatial resolution and complete spatial coverage. Following the discussion in section 2b(2), we disregard the spatial mismatch between point and gridcell data and assume that station calculated indices represent their collocated gridcell (0.1°) indices. We understand that over heterogeneous terrain (e.g., mountainous regions) this assumption might fail and lead to erroneous conclusions.
For the comparison, we compute the spatial averages over the regions defined in Fig. 1 using the HadEX2 and PGF-BC datasets. Given the different spatial coverage of the datasets, in addition to the spatial averages calculated using the entire PGF-BC dataset we use the coverage mask of HadEX2 to calculate the regional averages of the PGF-BC derived ETCCDI indices. For each series, the slope of the Kendall–Theil robust line and its accompanying 99% confidence interval are computed.
As shown in Fig. 14, the sign of the slope (and sometimes the magnitude) of the PGF-BC ETCCDI temperature indices match those calculated from HadEX2 with differences in the CSDI index (increase in HadEX2 and decreases in PGF-BC). Differences in the southeast region (Greater Horn of Africa) are most likely due to the inability to bias correct the values in this region because of lack of station data and persisting errors from undetected step changes and uncertainties after homogenization. Although not shown here, the spatial coverage over the domain of the temperature indices (especially TN10p, TX10p, TN90p, and TX90p) in HadEX2 is good. This helps explain the close relationship between the full and HadEX2 masked PGF-BC calculated ETCCDI indices.
Slope of the Kendall–Theil robust line and corresponding confidence intervals (99%) of the ETCCDI indices for the regions defined in Fig. 1. The trends are computed between 1979 and 2005 using the HadEX2 dataset, PGF-BC dataset, and PGF-BC dataset using the HadEX2 mask.
Citation: Journal of Climate 27, 15; 10.1175/JCLI-D-13-00423.1
The sign of the slope (and sometimes the magnitude) of the HadEX2 masked PGF-BC ETCCDI precipitation indices match those calculated from HadEX2. However, this is not the case when compared to the unmasked PGF-BC. In many cases, even the sign of the slope is switched (e.g., the entire domain goes from a negative to a positive trend). The most likely cause of these discrepancies is the similar station data being used in both datasets; thus, when we compare HadEX2 to the masked PGF dataset, the results are similar. It is also possible that errors in the original PGF precipitation product make a favorable comparison only in regions that have many in situ stations that can be used for bias correction. However, this is probably not the case given the low spatial correlation of the departures (i.e., ability to correct) and thus the low-density network cannot have a large impact.
d. Attribution of changes in climate extremes
As shown in Fig. 14, the trends in temperature ETCCDI indices indicate an increase in warm extremes over sub-Saharan Africa between 1979 and 2005. However, for many indices, the confidence intervals also indicate high uncertainty in the magnitude and sometimes the sign of the slope. This uncertainty is most likely due to natural variability and short time span. Donat et al. (2013b) examined the impact of natural variability in ETCCDI indices over a section of this study’s domain (the Arab region) by assessing their relationship to the El Niño–Southern Oscillation (ENSO) and North Atlantic Oscillation (NAO). They found a strong relationship with the NAO in the west and a more significant relationship with ENSO in the east. Future work should look into how these results transfer to the greater sub-Saharan Africa and the connection of the trends in ETCCDI indices to other climate indices and changes in general circulation patterns. Understanding their impact could help understand the signal from the interannual variability and add confidence to the calculated long-term trends. Ultimately, accurately extending the datasets in time will lead to the largest decrease in uncertainty in ETCCDI trends.
For many regions in this study’s domain, the magnitude of the positive trend in TN90p and TX90p indices is larger than the magnitude of negative trend in TN10p and TX10p. This suggests that although the mean of the distribution is shifting toward warmer values, it is not shifting uniformly. This agrees with previous studies that show how the distribution of global daily temperature anomalies has changed and experienced a shift toward warmer temperatures (Donat and Alexander 2012; Hansen et al. 2012). Changes in the distribution are possibly through an increase in skewness while changes in variance are highly spatially heterogeneous (Donat and Alexander 2012). Given the complex behavior of the moments of the temperature distribution, this reinforces the need to provide metrics of the entire distribution (Katz et al. 2013). Although the ETCCDI indices suggest a broadening of the distribution over sub-Saharan Africa, it is unclear if this is due to changes in skewness, variance, or kurtosis. Future work should analyze the PGF-BC dataset developed in this study to provide a more complete picture of the distribution through its different moments.
The increase in precipitation across sub-Saharan Africa can be mainly attributed to the recovery from the droughts of the 1970s and 1980s. This is apparent across all precipitation indices: a decrease in dry spells, an increase in wet spells, an increase in heavy rainfall, and an increase in annual rainfall. This suggests a fairly uniform shift in the probability distribution of rainfall toward wetter conditions. However, questions remain regarding the drivers of the differences between regions. Future work should look into other natural and anthropogenic drivers of trends in the data and determine the connection to changes in the monsoon over the Sahel and the Greater Horn of Africa (e.g., Williams et al. 2012) and connections to changes in general circulation patterns.
5. Conclusions
A high-resolution meteorological dataset for 1979–2005 has been developed for sub-Saharan Africa between 5°S and 25°N. A combination of reanalysis, gridded observation datasets, and in situ networks (GSOD) are used to develop the final downscaled, homogenized, and bias-corrected dataset. The accuracy of the final dataset is dependent on the density and temporal availability of the contributing GSOD station data. The highest confidence in the data is in western Africa and the Greater Horn of Africa, where the GSOD database has the highest density of stations. In regions with few stations, such as Chad and the Democratic Republic of the Congo, the accuracy of the dataset is subject to the biases and any remaining undetected temporal inhomogeneities in the reanalysis and gridded observation datasets.
We use the dataset to evaluate changes in meteorological extremes over the central latitudes of Africa including the Sahel, western Africa, the Greater Horn of Africa, the southern Arabian Peninsula, and the northern Congo basin. ETCCDI indices were calculated for 1979–2005 and the Mann–Kendall test was used to determine the statistical significance of monotonic trends in the indices. The magnitudes of the trends and their confidence intervals were calculated using the Kendall–Theil robust line. There has been a statistically significant increase in the annual number of warm days and nights and a corresponding decrease in cold days and nights. These results agree qualitatively with other regional studies of changes in climate extremes. Changes in precipitation can most likely be attributed to the droughts in the 1970s and 1980s in the Sahel.
The data have been developed for multiple purposes including as input to hydrologic and crop modeling, for flood and drought monitoring, and for climate studies. The high spatial resolution of the gridded dataset offers an unprecedented opportunity to model and capture local features not available in coarser gridded datasets over data-sparse regions. The corrections for topography and the merging of in situ measurements allow for the representation of temperature and precipitation gradients over mountainous regions and water bodies; this is a feature generally not available in coarser datasets. The data can also be used in drought and flood monitoring tools over the region such as Princeton University’s Experimental African Flood and Drought Monitor (Sheffield et al. 2014) and FewsNet (http://www.fews.net).
Acknowledgments
This study was supported by funding from the International Center for Tropical Agriculture (ICAT) Research Program on Climate Change, Agriculture and Food Security (CCAFS), NASA Grant NNX08AN40A, and NOAA Grant NA10OAR4310130.
APPENDIX
Station Data Gap Filling
One source of temporal inhomogeneities in the CRU dataset is the interpolation and upscaling of changing station networks, especially in data-sparse regions, where a few stations can dominate the gridded values. Changes in the station density and spatial configuration can impact the gridded estimates and create artificial step changes in the mean and variance. To avoid similar problems when assimilating the station data into the downscaled and homogenized PGF dataset, the network configuration and density are kept constant in time by gap filling the noncontinuous station records. We do this by bias correcting the data at the collocated grid cell in the downscaled PGF dataset using the available station time series. The bias-corrected data are then used to infill the missing station data. A simple bias correction technique is used that matches the cumulative distribution function (cdf) of the collocated gridcell dataset to the available station data. This technique has been successfully used in previous studies of assimilation and correction of radar precipitation, satellite remote sensing of soil moisture, and hydrologic forecasts (Anagnostou et al. 1999; Atlas et al. 1990; Gao et al. 2007; Reichle and Koster 2004; Wood et al. 2002).
For discontinuous fields such as precipitation, we take extra care to avoid errors induced by the disparity in spatial coverage between the gridded values and the point station data. For example, if the grid cell has a higher mean number of rain days in a year than the station data, directly matching the distribution leads to an overestimation in the annual mean. Therefore, prior to cdf matching, the number of rain days of the collocated grid cell is adjusted to ensure the climatology of the mean number of rain days per month matches that of the station. Rain days are added or removed to account for the spatial correlation of daily rain events by selecting days for which the surrounding region (radius equal to 1.0°) is also a rain or dry event. For example, if a rain day is to be added, we search for a dry day in the month that has the maximum fraction of wet grid cells (e.g., 90%). When possible, the choice of when to add/remove a rain day is made to improve the grid cell’s climatology of first-order transition probabilities compared to those of the station. However, preference is always given to the spatial properties. When neither the spatial properties can be maintained nor the transition probabilities can be improved, the day is chosen at random. This is repeated until the number of rain days matches those of the station. When a rain day is added, the rain amount is sampled from a gamma distribution that is previously fitted to the collocated grid cell’s rain day values. Finally, the distribution of the collocated grid cell is matched to the available station data.
REFERENCES
Adler, R. F., and Coauthors, 2003: The Version 2 Global Precipitation Climatology Project (GPCP) monthly precipitation analysis (1979–present). J. Hydrometeor., 4, 1147–1167, doi:10.1175/1525-7541(2003)004<1147:TVGPCP>2.0.CO;2.
Aguilar, E., and Coauthors, 2009: Changes in temperature and precipitation extremes in western central Africa, Guinea Conakry, and Zimbabwe, 1955–2006. J. Geophys. Res., 114, D02115, doi:10.1029/2008JD011010.
Alexander, L. V., and Coauthors, 2006: Global observed changes in daily climate extremes of temperature and precipitation. J. Geophys. Res., 111, D05109, doi:10.1029/2005JD006290.
Anagnostou, E. N., A. J. Negri, and R. F. Adler, 1999: Statistical adjustment of satellite microwave monthly rainfall estimates over Amazonia. J. Appl. Meteor., 38, 1590–1598, doi:10.1175/1520-0450(1999)038<1590:SAOSMM>2.0.CO;2.
Athar, H., 2014: Trends in observed extreme climate indices in Saudi Arabia during 1979–2008. Int. J. Climatol.,34, 1561–1574, doi:10.1002/joc.3783.
Atlas, D., D. Rosenfeld, and D. B. Wolfd, 1990: Climatologically tuned reflectivity-rain rate relations and links to area-time integrals. J. Appl. Meteor., 29, 1120–1135, doi:10.1175/1520-0450(1990)029<1120:CTRRRR>2.0.CO;2.
Barnett, T. P., J. C. Adam, and D. P. Lettenmaier, 2005: Potential impacts of a warming climate on water availability in snow-dominated regions. Nature, 438, 303–309, doi:10.1038/nature04141.
Beek, E. G., A. Stein, and L. L. F. Janssen, 1992: Spatial variability and interpolation of daily precipitation amount. Stochastic Hydrol. Hydraul., 6, 304–320, doi:10.1007/BF01581623.
Busuioc, A., and H. von Storch, 1996: Changes in the winter precipitation in Romania and its relation to the large-scale circulation. Tellus,48, 538–552, doi:10.1034/j.1600-0870.1996.t01-3-00004.x.
Chiles, J.-P., and P. Delfiner, 1999: Geostatistics: Modeling Spatial Uncertainty.Wiley, 695 pp.
Chirlin, G. R., and E. F. Wood, 1982: On the relationship between kriging and state estimation. Water Resour. Res., 18, 432–438, doi:10.1029/WR018i002p00432.
Demaria, E. M. C., E. P. Maurer, J. Sheffield, E. Bustos, D. Poblete, S. Vicuña, and F. Meza, 2013: Using a gridded global dataset to characterize regional hydroclimate in central Chile. J. Hydrometeor., 14, 251–265, doi:10.1175/JHM-D-12-047.1.
Donat, M. G., and L. V. Alexander, 2012: The shifting probability distribution of global daytime and night-time temperatures. Geophys. Res. Lett., 39, L14707, doi:10.1029/2012GL052459.
Donat, M. G., L. V. Alexander, L. Yang, I. Durre, R. Vose, and J. Caesar, 2013a: Global land-based datasets for monitoring climatic extremes. Bull. Amer. Meteor. Soc., 94, 997–1006, doi:10.1175/BAMS-D-12-00109.1.
Donat, M. G., and Coauthors, 2013b: Changes in extreme temperature and precipitation in the Arab region: Long-term trends and variability related to ENSO and NAO. Int. J. Climatol., 34, 581–592, doi:10.1002/joc.3707.
Donat, M. G., and Coauthors, 2013c: Updated analyses of temperature and precipitation extreme indices since the beginning of the twentieth century: The HadEX2 dataset. J. Geophys. Res., 118, 2098–2118, doi:10.1002/jgrd.50150.
Dunn, R. J. H., K. M. Willett, P. W. Thorne, E. V. Woolley, I. Durre, A. Dai, D. E. Parker, and R. S. Vose, 2012: HadISD: A quality-controlled global synoptic report database for selected variables at long-term stations from 1973–2011. Climate Past, 8, 1649–1679, doi:10.5194/cp-8-1649-2012.
Durre, I., M. J. Menne, B. E. Gleason, T. G. Houston, and R. S. Vose, 2010: Comprehensive automated quality assurance of daily surface observations. J. Appl. Meteor. Climatol., 49, 1615–1633, doi:10.1175/2010JAMC2375.1.
Ferguson, C. R., and G. Villarini, 2012: Detecting inhomogeneities in the Twentieth Century Reanalysis over the central United States. J. Geophys. Res., 117, D05123, doi:10.1029/2011JD016988.
Frich, P., L. V. Alexander, P. Della-Marta, B. Gleason, M. Haylock, A. M. G. Klein Tank, and T. Peterson, 2002: Observed coherent changes in climatic extremes during the second half of the twentieth century. Climate Res., 19, 193–212, doi:10.3354/cr019193.
Gao, H., E. F. Wood, M. Drusch, and M. F. McCabe, 2007: Copula-derived observation operators for assimilating TMI and soil moisture into land surface models. J. Hydrometeor., 8, 413–429, doi:10.1175/JHM570.1.
Gupta, V. H., H. Kling, K. K. Yilmaz, and G. F. Martinez, 2009: Decomposition of the mean squared error and NSE performance criteria: Implications for improving hydrological modelling. J. Hydrol., 377, 80–91, doi:10.1016/j.jhydrol.2009.08.003.
Hansen, J., M. Sato, and R. Ruedy, 2012: Perception of climate change. Proc. Natl. Acad. Sci. USA, 109, E2415–E2423, doi:10.1073/pnas.1205276109.
Harris, I., P. D. Jones, T. J. Osborn, and D. H. Lister, 2013: Updated high-resolution grids of monthly climatic observations—The CRU TS3.10 dataset. Int. J. Climatol., 34, 623–642, doi:10.1002/joc.3711.
Hegerl, G. C., F. W. Zwiers, P. A. Scott, and V. V. Kharin, 2004: Detectability of anthropogenic changes in annual temperature and precipitation extremes. J. Climate, 17, 3683–3700, doi:10.1175/1520-0442(2004)017<3683:DOACIA>2.0.CO;2.
Held, I. M., T. L. Delworth, J. Lu, K. L. Findell, and T. R. Knutson, 2005: Simulation of Sahel drought in the 20th and 21st centuries. Proc. Natl. Acad. Sci. USA, 102, 17 891–17 896, doi:10.1073/pnas.0509057102.
Hisdal, H., K. Stahl, L. M. Tallaksen, and S. Demuth, 2001: Have streamflow droughts in Europe become more severe or frequent? Int. J. Climatol., 21, 317–333, doi:10.1002/joc.619.
Hollingsworth, A., D. B. Shaw, P. Lonnberg, L. Illari, K. Arpe, and A. J. Simmons, 1985: Monitoring of observation and analysis quality by a data assimilation system. Mon. Wea. Rev., 114, 861–879, doi:10.1175/1520-0493(1986)114<0861:MOOAAQ>2.0.CO;2
Huffman, G. J., and Coauthors, 2007: The TRMM Multisatellite Precipitation Analysis (TMPA): Quasi-global, multiyear, combined-sensor precipitation estimates at fine scales. J. Hydrometeor., 8, 38–55, doi:10.1175/JHM560.1.
Joyce, R. J., J. E. Janowiak, P. A. Arkin, and P. Xie, 2004: CMORPH: A method that produces global precipitation estimates from passive microwave and infrared data at high spatial and temporal resolution. J. Hydrometeor., 5, 487–503, doi:10.1175/1525-7541(2004)005<0487:CAMTPG>2.0.CO;2.
Kalnay, E., and Coauthors, 1996: The NCEP/NCAR 40-Year Reanalysis Project. Bull. Amer. Meteor. Soc., 77, 437–471, doi:10.1175/1520-0477(1996)077<0437:TNYRP>2.0.CO;2.
Karl, T. R., and R. W. Knight, 1997: The 1995 Chicago heat wave: How likely is a recurrence? Bull. Amer. Meteor. Soc., 78, 1107–1119, doi:10.1175/1520-0477(1997)078<1107:TCHWHL>2.0.CO;2.
Katz, R. W., P. F. Craigmile, P. Guttorp, M. Haran, B. Sanso, and M. L. Stein, 2013: Uncertainty analysis in climate change assessments. Nat. Climate Change, 3, 769–771, doi:10.1038/nclimate1980.
Kendall, M. G., 1975: Rank Correlation Methods. Charles Griffin, 202 pp.
Klein Tank, A. M. G., F. W. Zwiers, and X. Zhang, 2009: Guidelines on analysis of extremes in a changing climate in support of informed decisions for adaptation. WMO Climate Data and Monitoring Rep. 72 and WMO Tech. Doc. 1500, 55 pp.
Kruger, A., 2006: Observed trends in daily precipitation indices in South Africa: 1910–2004. Int. J. Climatol., 26, 2275–2285, doi:10.1002/joc.1368.
Kubota, T., and Coauthors, 2007: Global precipitation map using satellite-borne microwave radiometers by the GSMaP project: Production and validation. IEEE Trans. Geosci. Remote Sens., 45, 2259–2275, doi:10.1109/TGRS.2007.895337.
Lettenmaier, D. P., E. F. Wood, and J. R. Wallis, 1994: Hydro-climatological trends in the continental United States, 1948–88. J. Climate, 7, 586–607, doi:10.1175/1520-0442(1994)007<0586:HCTITC>2.0.CO;2.
Lins, H. F., and J. R. Slack, 1999: Streamflow trends in the United States. Geophys. Res. Lett., 26, 227–230, doi:10.1029/1998GL900291.
Mann, H. B., 1945: Nonparametric test against trend. Econometrica, 13, 245–259, doi:10.2307/1907187.
McPherson, R. A., and Coauthors, 2007: Statewide monitoring of the mesoscale environment: A technical update on the Oklahoma Mesonet. J. Atmos. Oceanic Technol., 24, 301–321, doi:10.1175/JTECH1976.1.
Mitchell, K. E., D. Lohmann, P. R. Houser, E. F. Wood, and J. C. Schaake, 2004: The multi-institution North American Land Data Assimilation System (NLDAS): Utilizing multiple GCIP products and partners in a continental distributed hydrological modeling system. J. Geophys. Res., 109, D07S90, doi:10.1029/2003JD003823.
Mitosek, H. T., 1995: Climate variability and change within the discharge time series: A statistical approach. Climatic Change, 29, 101–116, doi:10.1007/BF01091641.
Mood, A. M., F. A. Graybill, and D. C. Boes, 1974: Introduction to the Theory of Statistics. 3rd ed. McGraw-Hill, 564 pp.
New, M., M. Hulme, and P. Jones, 2000: Representing twentieth-century space–time climate variability. Part II: Development of 1901–96 monthly grids of terrestrial surface climate. J. Climate, 13, 2217–2238, doi:10.1175/1520-0442(2000)013<2217:RTCSTC>2.0.CO;2.
New, M., and Coauthors, 2006: Evidence of trends in daily climate extremes over southern and west Africa. J. Geophys. Res., 111, D14102, doi:10.1029/2005JD006289.
Peterson, T. C., and Coauthors, 1998: Homogeneity adjustments of in situ atmospheric climate data: A review. Int. J. Climatol., 18, 1493–1517, doi:10.1002/(SICI)1097-0088(19981115)18:13<1493::AID-JOC329>3.0.CO;2-T.
Peterson, T. C., X. Zhang, M. Brunet-India, and J. L. Vazquez-Aguirre, 2008: Changes in North American extremes derived from daily weather data. J. Geophys. Res., 113, D07113, doi:10.1029/2007JD009453.
Pettitt, A. N., 1979: A non-parametric approach to the change-point problem. J. Roy. Stat. Soc., 28C, 126–135.
Reeves, J., J. Chen, X. L. Wang, R. Lund, and Q. Lu, 2007: A review and comparison of changepoint detection techniques for climate data. J. Appl. Meteor. Climatol., 46, 900–915, doi:10.1175/JAM2493.1.
Reichle, R. H., and R. D. Koster, 2004: Bias reduction in short records of satellite soil moisture. Geophys. Res. Lett., 31, L19501, doi:10.1029/2004GL020938.
Schar, C., and G. Jendritzky, 2004: Climate change: Hot news from summer 2003. Nature, 432, 559–560, doi:10.1038/432559a.
Sen, P. K., 1968: Estimates of the regression coefficient based on Kendall’s tau. J. Amer. Stat. Assoc., 63, 1379–1389, doi:10.1080/01621459.1968.10480934.
Sheffield, J., and E. F. Wood, 2008: Global trends and variability in soil moisture and drought characteristics, 1950–2000, from observation-driven simulations of the terrestrial hydrologic cycle. J. Climate, 21, 432–458, doi:10.1175/2007JCLI1822.1.
Sheffield, J., G. Goteti, and E. F. Wood, 2006: Development of a 50-year high-resolution global dataset of meteorological forcings for land surface modeling. J. Climate, 19, 3088–3111, doi:10.1175/JCLI3790.1.
Sheffield, J., and Coauthors, 2014: A drought monitoring and forecasting system for sub-Sahara African water resources and food security. Bull. Amer. Meteor. Soc., doi:10.1175/BAMS-D-12-00124.1, in press.
Smith, A., L. Neal, and V. Russ, 2011: The integrated surface database: Recent developments and partnerships. Bull. Amer. Meteor. Soc., 92, 704–708, doi:10.1175/2011BAMS3015.1.
Sorooshian, S., K.-L. Hsu, X. Gao, H. V. Gupta, B. Imam, and D. Braithwaite, 2000: Evaluation of PERSIANN system satellite–based estimates of tropical rainfall. Bull. Amer. Meteor. Soc., 81, 2035–2046, doi:10.1175/1520-0477(2000)081<2035:EOPSSE>2.3.CO;2.
Stackhouse, P. W., Jr., S. K. Gupta, S. J. Cox, J. C. Mikovitz, T. Zhang, and M. Chiacchio, 2004: 12-year surface radiation budget data set. GEWEX News, Vol. 14, No. 4, International GEWEX Project Office, Silver Spring, MD, 10–12.
Theil, H., 1950: A rank-invariant method of linear polynomial regression analysis. Indagationes Math., 12, 85–91.
Troy, T. J., J. Sheffield, and E. F. Wood, 2011: Estimation of the terrestrial water budget over northern Eurasia through the use of multiple data sources. J. Climate, 24, 3272–3293, doi:10.1175/2011JCLI3936.1.
Van de Beek, C. Z., H. Leijnse, P. J. J. F. Torfs, and R. Uijlenhoet, 2011: Climatology of daily rainfall semi-variance in the Netherlands. Hydrol. Earth Syst. Sci., 15, 171–183, doi:10.5194/hess-15-171-2011.
Venema, V., and Coauthors, 2012: Benchmarking homogenization algorithms for monthly data. Climate Past, 8, 89–115, doi:10.5194/cp-8-89-2012.
Villarini, G., F. Serinaldi, J. A. Smith, and W. F. Krajewski, 2009: On the stationarity of annual flood peaks in the continental United States during the 20th century. Water Resour. Res., 45, W08417, doi:10.1029/2008WR007645.
Wang, A., D. P. Lettenmaier, and J. Sheffield, 2011: Soil moisture drought in China. J. Climate, 24, 3257–3271, doi:10.1175/2011JCLI3733.1.
Webster, R., and M. A. Oliver, 2007: Geostatistics for Environmental Scientists. 2nd ed. John Wiley and Sons, 330 pp.
Williams, A. P., and Coauthors, 2012: Recent summer precipitation trends in the Greater Horn of Africa and the emerging role of Indian Ocean sea surface temperature. Climate Dyn., 39, 2307–2328, doi:10.1007/s00382-011-1222-y.
Wood, A. W., E. P. Maurer, A. Kumar, and D. P. Lettenmaier, 2002: Long range experimental hydrologic forecasting for the eastern United States. J. Geophys. Res., 107, 4429, doi:10.1029/2001JD000659.
Zhang, X., and Coauthors, 2005: Trends in Middle East climate extreme indices from 1950 to 2003. J. Geophys. Res., 110, D22104, doi:10.1029/2005JD006181.
Zhang, X., L. Alexander, G. C. Hegerl, P. Jones, A. Klein Tank, T. C. Peterson, B. Trewin, and F. W. Zwiers, 2011: Indices for monitoring changes in extremes based on daily temperature and precipitation data. Wiley Interdiscip. Rev.: Climate Change, 2, 851–870, doi:10.1002/wcc.147.
Ziegler, A. D., J. Sheffield, E. P. Maurer, B. Nijssen, E. F. Wood, and D. P. Lettenmaier, 2003: Detection of intensification in global and continental-scale hydrological cycles: Temporal scale of evolution. J. Climate, 16, 535–547, doi:10.1175/1520-0442(2003)016<0535:DOIIGA>2.0.CO;2.
