In this section, we present a sample application of the CHIRPS, based on hydrologic simulations over East Africa, which has experienced exceptional drying during boreal spring31. The reasonably long period of record, low latency, high resolution and daily disaggregations of the CHIRPS make it suitable for hydrologic modeling. Focusing on East Africa, and especially southeastern (SE) Ethiopia, we demonstrate how CHIRPS can be used to support humanitarian relief efforts while guiding climate-smart development. In eastern East Africa, the interaction of declining boreal spring rains, population growth, cropland extensification, and land cover change and degradation may be enhancing food insecurity30–33. The CHIRPS dataset has been used to explore the Eastern Africa nexus of climate change, population growth, and vegetation declines34 and identify links between drought and low birth weights35. In this sample application of the CHIRPS, we i) use the dataset to drive a hydrologic model, the Variable Infiltration Capacity (VIC) model36,37, ii) analyze recent (post-1999) changes in soil moisture (SM), evapotranspiration (ET), rainfall, and air temperatures, and iii) consider how the near real-time CHIRPS can be used with sea surface temperature (SST) conditions to predict most of the recent severe droughts in southeastern Ethiopia.
High-quality, temporally consistent and near real-time precipitation datasets such as CHIRPS can help identify environmental changes, quantify the important role played by warming air temperatures and play an important role in seasonal drought prediction38–42. CHIRPS can be used to force hydrologic models and simulate near-real time initial hydrologic conditions. State of soil moisture, groundwater and snow pack43 have been shown to contribute to drought prediction skill44, which is valuable given that there is limited skill in climate forecasts (mainly for precipitation)41,42 beyond approximately 2 months. Furthermore, statistical models for drought prediction have also been shown to be sensitive to and derive their skill from the initial soil moisture state45,46. These type of models can predict extremes like the 2011 East Africa drought. Below we show that the near real-time CHIRPS can help capture this potentially valuable hydrologic information and support enhanced drought early warning.
While the VIC simulations were run across the Greater Horn of Africa, our analysis focuses on southeastern Ethiopia (38.5°E-44°E, 6°N-10°N) because this is a region with chronic food insecurity (Fig. 5a) and with a large population (Fig. 5b) that is growing rapidly34. The region exhibits declining precipitation, ET, SM, and runoff as seen in the data (Fig. 6a). Between 2010 and 2014, much of this region was classified by FEWS NET as facing food stress or crisis conditions, on average, based on Integrated Food Security Phase Classification (IPC) assessments (www.fews.net) (Fig. 5a). IPC values ranging between 2 and 3 indicate recurrent food crises associated with chronic malnutrition rates of 10–15%, acute dietary diversity deficits, and accelerated and critical depletion of livelihood assets; an IPC value of 3 or more denotes a humanitarian emergency—with acute malnutrition affecting more than 15% of the population, severe food access shortages, and near complete and irreversible depletion or loss of livelihood assets (IPC, 2008). Gridded Population of the World estimates for 2020 for this region are 32,630,400, a value that is expected to almost double by 2050 to 62,000,928 (ref. 47). Boreal spring rains in this region have exhibited substantial declines since the 1980s (refs 31,32). Given that the persistence of soil moisture conditions often presents opportunities for effective hydrologic forecasts43, we examine here whether CHIRPS can help provide effective soil moisture predictions for annual southeastern Ethiopia soil moisture. We also explore the relative contributions of rainfall declines and warming air temperatures.
Hydrologic simulation percentiles for 1999–2104
In this section we examine 1981–2014 hydrologic changes, based on annual October-September accumulations of runoff, ET, and 0 to 40 cm soil moisture from VIC model simulations (the top two VIC soil moisture layers). The VIC model36,37 has previously been used to analyze recent East African droughts43,48,49. The VIC model was run in the water balance mode using daily CHIRPS, temperature maxima and minima and wind speed. Daily temperature and wind speed data used to run model simulation from 1981–2010 were the same as those used by Princeton African Drought Monitor49. Post-2010 daily temperature data were generated by constraining the Global Ensemble Forecast System (GEFS) daily temperature minimum and maximum analysis fields with gridded monthly observations of minimum and maximum air temperatures produced by CRU9. Regressions between monthly 1901–2013 CRU temperature fields and Goddard Institute for Space Science (GISS) gridded air temperature anomalies50 were used with 2014 GISS anomalies to constrain the 2014 GEFS temperature anomalies. Wind speed data post-2010 were simply the daily climatological mean for the period 1981–2010.
In Fig. 6a,b we show the average of the annual (October-September) 1999–2014 CHIRPS precipitation and GISS air temperature values, respectively, expressed as expressed as percentiles. These maps were generated by i) calculating the average 1999–2014 rank at each grid cell based on the full 1981–2014 time series and ii) dividing this average rank by 33, the number of water years. An October-September water year was chosen to highlight the region’s sensitivity to back-to-back Indo-Pacific-forced boreal fall and spring droughts51. While recent drying trends have mostly arisen during spring, the worst recent food security crises in East Africa (like that occurring in 2011) have typically been associated with multi-season dry spells. We focus, therefore, on annual soil moisture and runoff extremes, and our capability to predict southeastern Ethiopia droughts at the beginning of the main April-September rains. We refer to the 1981–1982 October-September water year as ‘1982’. Since 1999, parts of eastern East Africa have experienced below normal rainfall (yellow and red areas in Fig. 6a). Air temperatures have also increased (Fig. 6b), especially over northeastern Kenya and Ethiopia. Spatial patterns of the VIC soil moisture, ET, and runoff 1999–2014 percentile values (Fig. 6c–e) follow closely the CHIRPS precipitation anomalies (Fig. 6a); we will show below, however, that warm temperatures have also contributed substantially to increased aridity in Ethiopia.
Focusing on our southeastern Ethiopia domain, we present time series of precipitation, ET, and runoff, expressed as percent anomalies of the 1982–2014 mean (Fig. 6f). These percentiles are calculated with regard to the each grid cell’s 33-year history. In general, the coefficient of variation of the runoff is much higher than that of the rainfall or ET time series. Since 1999, only 2 years have exhibited above normal ET, runoff, or rainfall. This increased frequency of droughts52 is probably due to tropical Pacific SST forcing53. The most recent FEWS NET research links these droughts to an exceptionally strong tropical Pacific SST gradient32,43,48,54,55, caused largely by anthropogenic warming of the western Pacific56 and a natural decadal cooling of the eastern Pacific57.
We can examine the relative contributions of annual rainfall and temperature to annual southeastern Ethiopia runoff using leave-one-out cross-validated (CV) regressions (CV R2=0.63). Figure 6f shows the corresponding runoff estimates based on i) precipitation and GISS temperatures, and ii) GISS temperatures alone. The southeastern Ethiopia VIC runoff declined by ~20 percent between 1981–1998 and 1999–2014. Regression estimates suggest that about half (10 percent) of this decline may be attributed to lower rainfall and that the other half may be attributed to warming air temperatures.
We can apply a similar analysis to southeastern Ethiopia soil moisture anomalies (Fig. 6g). We have expressed the soil moisture time series using standardized anomalies to facilitate their interpretation in a drought monitoring context. Since 1999, only 2 years have exhibited above normal soil moisture, and the change between the 1981–1998 and 1999–2014 standardized anomalies was −0.9. A cross-validated regression predicting annual soil moisture based on rainfall and air temperatures fits almost perfectly (CV R2 0.88). These estimates (Fig. 6g), suggested that most (70%) of the decline in soil moisture was caused by reductions in precipitation, while air temperature increases may have accounted for about 30% of the decline.
The result that temperature change potentially accounted for 50 and 30% of runoff and soil moisture declines, respectively, may have important implications for future impacts of climate change. These estimated temperature influences are shown with purple bars on Fig. 6f,g. Based on the runoff results, warming might have prevented any change in runoff even if the rainfall trend had been in the opposite direction (a wetting trend). Note that such results can be highly model sensitive58, and these findings will need to be verified with more models. We plan future analyses using multiple hydrologic models in conjunction with numerical experiments, isolating precipitation and temperatures impacts through multiple suites of simulations.
Predictions of southeastern Ethiopia hydrologic extremes
Can CHIRPS help us predict southeastern Ethiopia droughts? As context, consider early April of 2011. The fall 2010 rains had been poor across the Horn of Africa, and millions of pastoralists and farmers faced severe food shortages unless conditions improved. What was the chance that drought would persist? Datasets like CHIRPS can help us answer this question by providing a basis for hydrologic simulations in near real-time and by giving us a firm foundation for exploring teleconnections and prediction. This requires precipitation estimates with both reasonably long periods of record and reasonably low latencies. Present soil conditions are often a good indicator of future hydrologic outcomes. Average October-March southeastern Ethiopia precipitation comprises on the order of 25% of the annual total (145 mm out of 622 mm). The main rains for this region come between April and September. However, October-March southeastern Ethiopia soil moisture conditions are a good indicator of the overall performance of the mean October-September soil moisture conditions (CV R2 0.47). While persistence of soil moisture anomalies certainly accounts for some of this predictability43, we find a similar persistence between October-March and April-September CHIRPS rainfall anomalies as well (CV R2 0.46), perhaps because both the October-March and April-September rainy seasons are suppressed by warm SSTs in the Indo-Pacific warm pool and stronger west to east tropical Pacific SST gradients54,59. This persistence between seasons has also been noted in another recent study60. We show how climate and land surface persistence can be combined to make effective hydrologic forecasts at a coarse spatial and temporal scale for southeastern Ethiopia.
Figure 7 shows January-March composites of the large-scale circulation based on the difference between the dry and wet years noted above. The January-March time period shows climate conditions just as rains commence in earnest over southeastern Ethiopia. The purpose of this plot is to show that just before the onset of the rains, the large-scale Indo-Pacific climate system exhibits large coherent anomalies consistent with an intensification of the Walker circulation. Conditions like these can be used by East African climate experts to predict some boreal spring droughts32,43,61. Coupled and uncoupled climate model simulations indicate that when there is a strong tropical SST gradient during boreal winter, with warm SST in the western Pacific and cool SST in the central Pacific, predictable rainfall deficits often follow during East Africa boreal spring rains.
In Fig. 7a we see a large difference between the Indo-Pacific climate system in the months preceding the driest and wettest years in southeastern Ethiopia. These years are indicated in Fig. 6f. Using 700 hPa winds and geopotential height anomalies from the Modern-Era Retrospective Analysis for Research and Applications (MERRA) reanalysis62, we can see that the difference between dry and wet years is characterized by strong anomalous westerly low winds extending from East Africa to the maritime continent. These anomalous winds are forced by a stronger equatorial height gradient between East Africa and the eastern Indian Ocean, and cells of sub-tropical low pressure to the north and south of the maritime continent. This type of circulation change probably reduces moisture transports into Ethiopia during both boreal spring63 and summer64, and appears related to warmer West Pacific SSTs51. While Ethiopia does not appear to be experiencing precipitation declines at a national scale65, the western part of country has been getting wetter while the eastern part has seen rainfall decreases66.
The SST67 and GPCP68 precipitation responses shown in Fig. 7b,c are characteristic of conditions associated with recent East Africa precipitation declines52,53 and drought predictions32,43,48,55. Warming in the western Pacific and cooling in the eastern Pacific intensify the climatological SST gradient supporting an intensification of the Walker Circulation with more (less) precipitation over the western (central) Pacific (Fig. 7c). The diabatic forcing from the increased Indo-Pacific precipitation helps force an equatorial Rossby wave response over the Indian Ocean54,69,70, producing westerly wind anomalies (Fig. 7a) and drying over eastern East Africa (Fig. 7c). Presumably, the January-March GPCP rainfall deficits across Tanzania, Kenya, and southern Somalia in Fig. 7c presage hotter drier conditions in southeastern Ethiopia in April-September.
Western Pacific and central Pacific SSTs exhibit substantial persistence, and partial correlations, controlling for antecedent October-March soil moisture conditions, show that January-March SST averaged over the ‘western V’56 boxes shown in Fig. 7b are reasonably well correlated with April-September southeastern Ethiopia soil moisture conditions (R=0.5). Exceptionally warm SSTs in this region have contributed to the 2014 East African spring drought32,48 and the long term decline in rainfall observed in the Centennial Trends precipitation product31.
Combining these January-March western V SSTs with the observed October-March southeastern Ethiopia soil moisture anomalies supports effective forecasts of most of this region’s hydrologic extremes (Fig. 8a). The overall cross-validated correlation of this simple forecast model is 0.78. Most (7 out of 8) of the targeted dry anomalies were predicted to be below normal, and most (4 out of 5) of the wet anomalies were predicted reasonably well. Thus, while southeastern Ethiopia has been experiencing declines in soil moisture, a majority of the individual drought events producing this decline are predictable.
Projections of warming air temperature impacts on southeastern Ethiopia
We conclude our analysis by taking a longer (1950–2014) look at southeastern Ethiopia climate conditions, combining gridded precipitation estimates and air temperatures values to approximate changes in soil moisture over longer periods than those covered by the CHIRPS. The basis of this approximation was the regression between annual soil moisture, rainfall, and air temperatures, which indicated strong relationships (CV R2 0.88). We re-iterate that this topic will be revisited using multiple hydrologic models and a more detailed experimental design. Our point here, however, is to suggest i) that the recent hydrologic changes shown in Fig. 6 appear to be quite severe in the context of the observed changes in Ethiopia rainfall and air temperatures and ii) that warming air temperatures may already be contributing to substantial soil moisture reductions.
To represent rainfall on longer time scales we make use of a new FEWS NET resource, the Centennial Trends (CenTrends) dataset31; CenTrends covers Eastern Africa and benefits from 217 rainfall stations provided by the Ethiopian national meteorological agency. Standard error estimates from the CenTrends estimation process (kriging) indicates reasonable accuracies back to the 1950s.
The same station data have been used in the 1981–2014 CHIRPS version 2 and CenTrends version 1.0 datasets; the correlation between the October-September southeastern Ethiopia rainfall time series is 0.95. Fig. 8b shows 15-yr running averages of 1950–2014 standardized October-September CenTrends rainfall, standardized March-June CenTrends rainfall, and unstandardized GISS air temperatures. Both October-September and March-June rainfall decline substantially between the 1970s and 2000s. Air temperature increases have been slowly accelerating since the 1960s, in line with a 73 member/39 model ensemble of climate change simulations obtained from the Royal Netherlands Meteorological Institute (http://climexp.knmi.nl/). These air temperature anomalies (Fig. 8b) were adjusted (multiplied by 1.5), based on a regression between the observed time series and the mean of the climate change simulations.
The 1950–2014 correlation between the 15-yr observed and modeled air temperatures (red and dark red lines, Fig. 8b) is 0.96. The climate change simulations provided projections based on observed 1950–2005 greenhouse gases, aerosols and greenhouse gasses (the historic experiment71) (Taylor et al.71), and 2006–2038 simulations from the 8.5 Wm−2 Representative Concentration Pathway experiment. This rapid warming scenario tracks closely with recent global emissions72. In both the model simulations and observations, the amount of warming since 1960 is quite large compared to the interannual standard deviation (~1 °C or 1.95 standardized deviations). By 2030, mean air temperatures are likely to rise another 0.4 °C (+~0.8 standardized deviation).
We can use our regression to estimate what this potential warming might entail (Fig. 8c). Between ~1970 and the 1990s, the influences of warming air temperatures may have been offset by increases in rainfall, resulting in little change in soil moisture. As rainfall decreased abruptly in the late 1990s and 2000s, soil moisture decreased as well, abetted by rising air temperatures. Overall, between the mid-1980s and late 2000s, 15-yr soil moisture anomalies decreased by ~1 standardized deviation—a substantial increase in aridity. About half of this decrease may be attributable to precipitation declines, and the VIC simulations suggest a similar magnitude of influence from rising air temperatures. If warming continues at the observed rate, as predicted by the climate change ensemble examined here, the mid-1980s to 2030 decrease in typical soil moisture conditions due to temperature influences alone might be ~−0.7 standardized deviation. A −0.7 sigma soil moisture anomaly is characteristic of the recent southeastern Ethiopia drought years (Fig. 6f).
Summary and discussion
The CHIRPS dataset presented here has been designed for drought monitoring in places like Ethiopia—regions with complex topography, changing observation networks and deep convective precipitation systems that correspond reasonably well with CCD estimates. While broadly equivalent at seasonal time scales to gridded precipitation products like the GPCC (Figs 3 and 4, Table 2), the CHIRPS is updated frequently, and provides data at higher spatial and temporal resolutions. The CHIRPS incorporates satellite information in three ways: by using satellite means to produce high resolution precipitation climatologies, by using CCD fields to estimate monthly and pentadal precipitation anomalies, and by using satellite precipitation fields to estimate local distance decay functions (Fig. 2), guiding the interpolation process. Validation studies suggest that satellite-enhanced CHPclim compares favorably (Table 1) to the CRU and WordClim climatologies19. Constraining the CHIRP by the CHPclim reduces systematic estimation errors (Fig. 3), producing low MAE and bias statistics (Table 2). This fidelity, in terms of low frequency performance, may come at the cost of under-representing extremes, as suggested by recent CHIRPS evaluations in Mozambique73 and South America74. Future versions of the CHIRPS will explore less prescriptive estimation procedures.
Another issue identified in this study is the tendency for the CHIRPS to underestimate the variance in some places, like southwest North America (Fig. 4e), where the satellite-only estimates perform well, in terms of correlation (R=0.70), but heavily underestimate the variance. This may relate to specifics of the frontal precipitation systems characteristic of this region, as well as the fact that the TMPA 3B42 training data used in the CHIRP development process also performs relatively poorly over this region (Fig. 4f). On the other hand, the consistency of the CHIRP inputs and the anomaly-based interpolation process used to incorporate stations helps limit spurious excursions caused by changes in the satellite observation network or precipitation gauge networks.
While improvements to the CHIRPS process are already being developed, the current version 2.0 seems well suited to monitoring droughts in regions where CCD estimates relate reasonably well to observed precipitation systems. Our Ethiopia study provides a good example of CHIRPS utility. Because CHIRPS is produced in near real-time, and October-March southeastern Ethiopia soil moisture exhibits a strong lagged correlation with April-September conditions, October-March CHIRPS data can be used in conjunction with hydrologic models like the VIC to monitor conditions and make successful forecasts (Fig. 8a) that capture most of the recent extreme drought events.
These forecasts combine skill from the inherent persistence of soil moisture states with skill accruing from persistent La Niña-like climate disruptions (Fig. 7). These patterns associate warming in the western Pacific and cooling in the eastern Pacific with increased drought over eastern East Africa, and the exceptional strength of this SST gradient is thought to contribute to the declining boreal spring rains in that region32,48. The SST warming patterns associated with southeastern Ethiopia drying (Fig. 7b) highlight the same warming regions identified by ENSO-residual trend analyses75,76 or ENSO-residual empirical orthogonal function analysis77. The regions associated with dry southeastern Ethiopia weather have warmed rapidly, in line with climate change projections77, especially when paired with La Niña-like cool central Pacific SST, can produce southeastern Ethiopia drying, and may continue to do so, unless eastern Pacific SST begins warming in accordance with most climate change simulations, which predict a wetter East Africa78. Climate change projections of precipitation for Ethiopia are quite diverse, but on average predict little change at the national level65. New paleo-climate evidence, however, based on sediment records from the Gulf of Aden, indicate long-term links between global warming and hotter, drier conditions in East Africa79, in line with observations of boreal spring rainfall31, the west Pacific SST gradient32,48, and the West Pacific Warming Mode56.
While temperature data are sparse, substantial changes in temperature, on the other hand, have been observed in Ethiopia, with estimates of warming rates around +0.3 °C per decade66,80,81. The VIC modeling results presented suggest that at annual time scales over the large southeastern Ethiopia domain, the aggregate effect of this warming may be affecting water availability, substantially reducing runoff (Fig. 6f) and soil moisture (Figs 6g and 8c) in a chronically food-insecure region (Fig. 5a) likely to have 32 million people by 2020 and 62 million people by 2050 (ref. 47). Building the close correspondence between annual simulated southeastern Ethiopia VIC soil moisture values and regression estimates based on annual air temperatures and rainfall, we suggest that a ~+2 °C warming of air temperatures may soon (~2030) reduce average southeastern Ethiopia soil moisture conditions by more than −0.7 of a standardized deviation. These results need to be corroborated with more hydrologic models.
In line with recent global assessments82, we have potentially identified substantial temperature contributions that have enhanced the effects of drought in southeastern Ethiopia, contributing perhaps 50 and 30% of the 1999–2014 reductions in runoff and soil moisture. As shown in Trenberth et al.82, precipitation must be taken into account when evaluating the drought impacts of temperature, and use of datasets like the CHIRPS should aid in this task. ‘Routine’ monitoring, like driving the VIC with real-time precipitation, can produce soil moisture estimates indicative of future outcomes (Fig. 8a)40,43. These types of powerful yet relatively simple forecasts may help us better manage 21st century climate variability, thereby better adapting to our changing climate.
While more work needs to be done to verify the simulation results presented here, and the CHIRPS and CHPclim algorithms will continue to be improved, this paper has described the CHIRPS algorithm, presented some promising validation results, and provided a motivating example application. One substantial weakness of the current CHIRPS algorithm is the lack of uncertainty information provided by the inverse distance weighting algorithm used to blend the CHIRP data and station data. We are currently exploring more rigorous geostatistical models (related to kriging83,84) and plan to use such frameworks to provide standard error fields in future CHIRPS releases. Ongoing research is also exploring mean bias errors with independent validation datasets. This work should lead to improved CHPclim fields, which would improve skill by better capturing the geographic variability of rainfall.
The U.S. Geological Survey (USGS)/Climate Hazards Group science team is also developing tools that can work with the CHIRPS, or other gridded precipitation datasets, to calculate the crop Water Requirement Satisfaction Index85,86, blend in situ observations with the satellite precipitation fields, and analyze climate trends and anomalies. These tools (GeoWRSI and GeoCLIM) and the CHIRPS dataset are available at http://chg.ucsb.edu/tools. The CHIRPS can also be viewed dynamically using the suite of visualizations available at earlywarning.usgs.gov. Climate variability can have a profound impact on the economies of countries like South Africa87 or Ethiopia65, and developing national capacities to deal with climate variability will be a critical component of climate adaptation88. National meteorological agencies can begin with precipitation estimates like CHIRPS, TARCAT, RFE2, or ARC gridded precipitation products, add their station data using tools like GeoCLIM89 or the International Research Institute’s Enhancing National Climate Time Series (ENACTS) software90, and develop information products that may improve disaster mitigation and climate adaption. By taking advantage of the processing involved in producing the CHIRPS, and adding value to the dataset by incorporating more in situ observations and local knowledge regarding risks and exposure, we hope that more scientists will be able to provide valuable climate services and environmental data, leading to improved adaptation and disaster mitigation.