Principal component analyses
The PCA of NCEP/NCAR reanalysis data revealed that surface temperature, zonal wind velocity, vertical velocity, and precipitation share two primary modes of variability in the tropical Indian and Pacific Ocean area during MAMJ. The first PC time series (PC1) accounts for 57% of the interannual variability among the original 12 PC time series (the first three PCs for each of the four climate variables). Figure 1d shows that this primary mode of variability is tightly coupled with surface warming in the tropical Indian Ocean (r > 0.8) and mildly coupled with warming in the Pacific Ocean (r ~ 0.6). PC1 also seems to be associated with a trend towards a zonally overturning Walker-like circulation, with an ascending branch centered over the rapidly warming Indian Ocean and a descending branch over the more slowly warming central tropical Pacific, between 180°W and 150°W (Fig. 1c). According to the NCEP/NCAR reanalysis, these ascending and descending branches have been accompanied by increased precipitation over the Indian Ocean and decreased precipitation over the central Pacific (Fig. 1e). Interestingly, PC1 correlates closely (r = 0.86) with the independently derived GISS record of global MAMJ mean surface temperature (Fig. 1a).
The second PC time series (PC2) accounts for 16% of the variability among the 12 PC time series. This second mode represents ENSO variability (Fig. 2). Positive PC anomalies are associated with cold phase (La Niña) conditions. Negative anomalies are associated with warm phase (El Niño) conditions. Among ENSO indices, correlation is strongest (r = −0.75) with mean SST anomalies within the Niño-4 region (5°S–5°N, 160°E–150°W), calculated from the independently derived NOAA extended SST dataset (Fig. 2a). Although Niño-4 temperatures have warmed slightly in recent decades, Fig. 2a shows a slight (statistically insignificant) positive trend in the second PC time series toward cold phase conditions. While the NCEP/NCAR reanalysis precipitation data are known to be problematic (Janowiak et al. 1998), the first two PC time series were not impacted significantly by whether the precipitation data were included in the PCA.
The results of the PCA were not unique to the NCEP/NCAR reanalysis dataset. The same variables within the ECMWF ERA-40 dataset also produced very similar representations of these first two PC time series (r = 0.88 and 0.9, respectively), despite coverage from only 1958 to 2002 (Fig. 3). Figure 3 shows that the first two PCs were also well represented within 1979–2009 radiosonde records of surface temperature and vertical profiles of zonal wind velocity. The shorter time scale of the radiosonde data, and thus diminished directional trends within records of temperature and wind velocity, caused the first and second PCs to be switched relative to the PC time series calculated from the longer reanalysis datasets. This was also the case when the PCA was replicated using only 1979–2009 NCEP/NCAR reanalysis surface temperature and zonal wind profile data (Fig. 3). A key difference between the ECMWF ERA-40 and NCEP/NCAR reanalyses is that ECMWF indicates trends toward upward vertical velocity anomalies above the Indian and central Pacific Oceans while the NCEP/NCAR reanalysis shows upward anomalies above the Indian Ocean and downward anomalies above the central Pacific. We therefore only conclude that, as global temperatures have increased, convection has increased over the Indian Ocean during MAMJ. The central Pacific is substantially less instrumented and we remain skeptical regarding trends in central Pacific vertical velocities.
The mechanism underlying the positive trend in the first PC time series appears to be an extension of the western edge of the tropical warm pool in the Indian Ocean. Regressing smoothed 1900–2009 SST data with global mean temperatures (Fig. 4a) suggests that as tropical SSTs have increased globally, they have increased in the Indian Ocean two to three times faster than SSTs in the tropical central Pacific (Fig. 4a). Figure 4b–c shows that the warm pool (here approximated as the area within the 28.75°C isotherm) has expanded by approximately 40° longitude on the west side of the warm pool in the past 50 years, from 100°E to 60°E. The majority of this expansion occurred dramatically in the 1960s and has been ongoing steadily since then. To the east, the 28.75°C isoline has expanded by approximately 20° (from 160°E to 180°E). The eastward expanse has been accompanied by much greater interannual and interdecadal variability due to ENSO. Over the past 5 years the eastern edge of the warm pool has had a similar position to that of the late 1980s and late 1990s. Figure 4d shows that, according to the NCEP/NCAR reanalysis, convective activity has increased over most of the warm pool region, especially on the western edge from 60°E to 90°E. In contrast, the NCEP/NCER record shows a decrease in convective activity on the eastern edge of the warm pool immediately east of the dateline, consistent with the patterns shown in Fig. 1c. Particularly striking in both 4c and 4d is the dramatic post-1998 increase in both SSTs and convection over both the core of the Pacific warm pool (~150°E) and over the central Indian Ocean (60°E–90°E).
Comparison of reanalysis to simulated and observed climate data
The disparity in SST warming rates between the tropical Indian and central Pacific Oceans is not accurately represented within the state-of-the-art GCMs used by the IPCC AR4. Six observational datasets concur that for every degree increase in global temperatures, the tropical central Pacific increased by ~0.5°C and the tropical Indian Ocean increased by ~1.5°C (Fig. 5a). Comparing these results to those modeled by 21 GCMs in 84 model runs within the CMIP3 20c3m experiment, most models over predicted warming rates in the central Pacific Ocean by 200–300%. GCMs also generally over predicted the warming rate in the Indian Ocean, but not as severely (Fig. 5a). The tendency for GCMs to greatly over-estimate central Pacific surface warming relative to Indian Ocean warming may explain their tendency to slow the Pacific trade wind circulation, ultimately reducing transports of energy into the extra-tropics. The disparity between observations and CMIP3 results shown in Fig. 5a corresponds well with Hoerling et al. (2010), who show a similar discrepancy between the CMIP3 multi-model ensembles driven with greenhouse gas and aerosols, and 20th century multi-model ensembles driven with observed SST patterns. Models driven with observed SSTs exhibit a strengthened Walker circulation, and Hoerling et al. note that this overturning circulation has likely enhanced moisture transports across the Pacific, leading to moisture convergence and rainfall increases across parts of the warm pool.
Among the other variables evaluated, there is little GCM consensus on how conditions might have been impacted as global temperatures rose in the 20th century (Fig. 5b–d). Interestingly, the NCEP/NCAR and ECMWF reanalyses also disagree as to how zonal wind, vertical velocity, and precipitation have varied over the past half century over the sparsely instrumented central Pacific Ocean (Fig. 5b–d). Over the Indian Ocean, however, these reanalyses agree that convection and rainfall have increased in step with SSTs since the mid-1900s. While the NCEP/NCAR reanalysis suggest intensification of upper-level westerly anomalies above the eastern Indian Ocean and western Pacific Ocean, the ECMWF does not indicate a significant change in upper-level (or lower) zonal wind velocities in this region (Fig. 5b).
Given the disagreement among modeled and reanalysis datasets in how global temperatures relate to the overturning Walker circulation, it was important to validate that the PC time series presented in this paper represent true climate processes and not artifacts of the reanalysis parameterization and/or inputs. In particular, we were interested in whether other precipitation datasets reflect the positive relationship between PC1 and precipitation over the Indian Ocean that is apparent in the NCEP/NCAR and ECMWF reanalyses. This is important because we believe trends in precipitation over the tropical Indian Ocean should impact trends in east African long rains. To examine this question we examined the correlation between PC1 and Indian Ocean Precipitation (0°S–15°S, 60°E–90°E) for each of the of the five validation precipitation datasets listed in Sect. 2.1. Considering 1979 through 2008 (the satellite era, during which we have relatively high confidence in ocean precipitation retrievals), PC1 correlated positively (0.41 < r < 0.66) and significantly (10−6 < p < 0.03) with all five data sets. In contrast, there was disagreement among the validation datasets regarding relationships with PC1 in the central Pacific Ocean, as was the case for NCEP/NCAR and ECMWF. PC2 was very well represented among the five validation datasets, with clear correlation patterns associated with ENSO (positive correlations over the Maritime Continent and negative correlations in the central Pacific).
It should be noted that the magnitude of the increases in latent heating shown by the various Indian Ocean rainfall observations (~30–50 cm °C−1) is thermodynamically substantial. A 40 cm change in seasonal precipitation is equivalent to a 97 W m−2 flux, associated with a total of 3.58 × 1014 W across the associated region (0°S–15°S, 60°E–90°E). This is roughly 10% of the total latent heating across the entire tropical Indian Pacific (15°S–15°NS, 60°E–80°W). It seems likely, therefore, that the ~40° geographic extension of the western border of the warm pool (Fig. 4b, c) has also influenced the warm pool thermodynamic circulation.
In summary, observational records and reanalysis products all show a dominant mode of climate variability in the tropical warm-pool area that is tightly coupled to global temperatures. Observational and reanalysis data also agree that the Indian Ocean’s temperature response to global warming appears to be roughly three times that of the central Pacific’s (Fig. 5a). The GCMs also show a greater rate of warming in the Indian versus Pacific, but overestimated warming in the central and eastern Pacific, causing the simulated gradients between Indian and Pacific SSTs to grow more slowly. Despite the fact that the observed gradients increased faster than simulated gradients, both observed and simulated records concur that the Indian Ocean has warmed faster than the central Pacific. This suggests, in turn, that the differential heating rates may continue for some time to come. Another consensus among the observations and most models is the increase in rainfall across the Indian Ocean (Fig. 5d). The ECMWF agrees with the NCAR/NCEP reanalysis, suggesting that 1°C of global warming will be associated with increased convection and an increase of more than 50 cm of rainfall over the Indian Ocean during MAMJ. It seems safe to conclude that models and observations all concur on a positive correlation among global temperatures, Indian Ocean SSTs, and Indian Ocean rainfall.
What is uncertain is the atmospheric response to the warming signal over the central tropical Pacific. The NCEP/NCAR reanalysis shows a distinct overturning circulation, with increased ascent over the Indian Ocean coupled with subsidence across the central Pacific. This response, however, is not shown in the ECMWF reanalysis. Further, the CMIP3 models simulate a weak association between global temperatures and increased convection over both the central Pacific and Indian Oceans (Fig. 5c). The tendency for models to simulate an increase in convection over the central Pacific contributes to the popular interpretation that continued warming will lead to a more ‘El Nino-like’ climate. The strength of the correlations between modeled central Pacific vertical velocities and global temperatures, however, are weak (less than 0.4), so it is unclear how atmospheric circulation over the central Pacific will respond to continued warming and how these circulation changes will impact precipitation in regions where precipitation has been historically correlated with variability in ENSO indices.
Precipitation and circulation impacts of PC1 and PC2
As PC1 and global temperature have grown more positive and convection and rainfall have increased over the Indian Ocean in recent decades, precipitation totals during the long-rains season in eastern Africa have declined. In Ethiopia and Kenya, which are reasonably well instrumented with meteorological stations, our CHG-CLIM precipitation data indicate that declines have been most severe in the central and eastern regions (Fig. 6). In the last 30 years, MAMJ rainfall totals have declined by 35–45% of the 1950–1979 mean throughout much of the area to the south and east of the Ethiopian Highlands (Fig. 6b). The spatial characteristics and magnitude of these precipitation declines are corroborated by GPCC and GPCP data (Fig. 6a, c).
Figure 7 shows how long-season precipitation in eastern Africa and the leading modes of climate variability relate to moisture transports across the broad geographic region studied in this paper. On average, large transports of moisture flow across the southern Indian Ocean and enter eastern Africa between 20°S and 20°N. Figure 7a indicates that the driest 20% of long-rains seasons have been associated with anomalously low specific humidity across equatorial east Africa, anti-cyclonic moisture transport anomalies across Ethiopia and northern Kenya, and divergence anomalies of moisture over the east coast of equatorial Africa and the western Indian Ocean. Dry seasons have also been associated with westerly moisture transport anomalies across the equatorial Indian Ocean and high specific humidity and convergence anomalies over much of the Maritime continent and southeastern Indian Ocean. In contrast, wet seasons have been characterized by approximately the opposite anomaly patterns (data not shown).
While the mean wind velocity and humidity data provided by the NCEP/NCAR reanalysis are too low in spatial resolution to accurately discern the complex paths of air parcels before reaching eastern Africa (Gatebe et al. 1999), the anomaly map in Fig. 7a suggests that the tropical Atlantic Ocean and Congo Basin may be important sources of water vapor for eastern Africa during MAMJ, consistent with Segele et al. (2009). During dry seasons, Walker circulation over the tropical Indian Ocean is particularly strong and the descending western branch of the cell appears to suppress moisture transports from the Atlantic, interior of Africa, and Indian Ocean (Fig. 7a). During wet seasons, moisture more freely converges upon eastern Africa and a weakened or eliminated descending branch of the Indian Ocean Walker circulation allows for increased convection and precipitation (data not shown).
The strength of the Indian Ocean Walker circulation is impacted by the ENSO cycle. The warm ENSO phase leads to a weaker Walker circulation and is weakly associated with enhanced long-rains precipitation in Kenya and Ethiopia. While correlations between long-rains precipitation and ENSO indices are weak, there is a strong association between low-level moisture convergence over the tropical warm pool and low-level divergence and dry conditions over eastern Africa (Fig. 7a). This contributes to the La Niña-like pattern in the map of dry anomalies in Fig. 7a, as can be seen by comparing Fig. 7a to a map of correlation between PC2 and vapor-transports, where the length of arrows depict the strength of correlation (Fig. 7c). Interestingly, ENSO indices have slightly trended toward a warmer phase in recent decades, but east African vapor transports and rainfall have declined. Figure 7b suggests that the decline of vapor transports into eastern Africa, despite the slight tilt toward warm-phase conditions, may be associated with rapid warming of the tropical Indian Ocean and a westward extension of the ascending branch of the Walker circulation (increasing PC1).
Viewing these relationships from a rotated perspective, Fig. 8a shows how PC1 correlates with vertical and zonal winds between 10°S and 10°N. Increased convergence and convection over the tropical Indian Ocean are associated with increased mid-level (~400–700 hPa) divergence aloft. Correlation arrows to the east of this divergence zone generally point upward and assimilate into a westerly correlation pattern associated with slowing of the tropical easterly jet. Correlation arrows to the west continue west within the mid-level troposphere and descend toward the surface over the Sahara Desert. This subsidence pattern over the Sahara manifests itself in Fig. 7b as an anti-cyclonic flow pattern of dry air below 700 hPa. The eastern reach of this dry, anti-cyclonic correlation pattern comes across Sudan, Ethiopia, and Kenya, suggesting reduced vapor transports from the Atlantic Ocean and Congo Basin region to the west. Figures 7b and 8a also indicate reduced vapor transports from the Indian Ocean, as low-level winds over the western Indian Ocean have trended toward a low-level convergence pattern over the central Indian Ocean.
Relationship between PC1 and interannual variability
As was shown in Figs. 7b and 8a, PC1 manifests itself over tropical Africa and the Indian Ocean in a way that should impact the long rains. Using the correlation fields displayed in Fig. 8 as PC loadings, PC1 can be recalculated using zonal and vertical velocities over equatorial Africa and the Indian Ocean out to 120°E. The recalculated PC1 reflects how the circulation trends associated with the original PC1 impact climate across tropical Africa and the Indian Ocean. A 31-years running correlation between the recalculated PC1 and long-rains precipitation reveals a progressively more negative relationship over the past six decades, growing from r ≈ 0 to r ≈ −0.45 (not shown). This suggests that the manifestations of PC1 over Africa and the Indian Ocean are becoming increasingly influential on long-rains precipitation. Because these processes appear to have only a secondary influence on long rains totals, we were interested in identifying the primary atmospheric circulation factors related to long-rains precipitation and then determining whether the multi-decadal processes associated with PC1 have influenced these interannual relationships.
As discussed in Sect. 2.4, exploratory data analysis revealed a fairly simple circulation index (the MHG), based on 500 hPa geopotential heights (GPHs), that correlated well with East African rainfall at seasonal (r = 0.61) and multi-seasonal time-scales (r = 0.74 for 5-years running means). Wet years across eastern Africa are associated with decreased easterly flows aloft, and this pattern corresponds, in turn, to higher 500 hPa GPHs along the equator and lower 500 hPa GPHs across extra topics (see Fig. 9a for boundaries of these regions). This pattern corresponds with the anomaly association associated with MJO activity (Pohl and Camberlin 2006a), but correlates better with east African rainfall throughout the 1948–2009 period than does MJO. Figure 9b plots seasonal MHG, MJO, and long-rains precipitation values. Figure 9c, d shows how vapor transports, zonal wind velocities, and vertical velocities correlate with MHG. High MHG years are associated with decreased convection over the Maritime continent, a weak tropical easterly jet across the Indian Ocean and Africa, increased moisture convergence and convection over eastern Africa, and increased moisture transports from the tropical Atlantic across central Africa. These are generally the same circulation characteristics associated with the MJO amplitude time series, but the MHG more explicitly represents the processes associated with long-rains precipitation.
We evaluated the relative impacts of PC1 and MHG by considering composites of MAMJ precipitation during each of four classes of seasons: (1) Low PC1 High MHG, (2) High PC1 High MHG, (3) Low PC1 Low MHG, (4) High PC1 High MHG (Fig. 10). Composites calculated with GPCC and CHG-CLIM suggest that, which MHG has the primary impact on precipitation throughout the Horn of Africa, increases in PC1 have been associated with a tendency towards decreased precipitation in this region during both low and high MHG seasons. This tendency is only moderate within Kenya’s coastal region (one of the two regions used to calculate the long-rains precipitation record used in this study) and much more pronounced inland.