Projected changes in regional seasonal precipitation due to climate change are highly uncertain, with model disagreement on even the sign of change in many regions. Using a 20-member CMIP5 ensemble under the RCP8.5 scenario, the intermodel uncertainty of the spatial patterns of projected end-of-twenty-first-century change in precipitation is found not to be strongly influenced by uncertainty in global mean temperature change. In the tropics, both the ensemble mean and intermodel uncertainty of regional precipitation change are found to be predominantly related to spatial shifts in convection and convergence, associated with processes such as sea surface temperature (SST) pattern change and land–sea thermal contrast change. The authors hypothesize that the zonal-mean seasonal migration of these shifts is driven by 1) the nonlinear spatial response of convection to SST changes and 2) a general movement of convection from land to ocean in response to SST increases. Assessment of tropical precipitation model projections over East Africa highlights the complexity of regional rainfall changes. Thermodynamically driven moisture increases determine the magnitude of the long rains (March–May) ensemble mean precipitation change in this region, whereas model uncertainty in spatial shifts of convection accounts for almost all of the intermodel uncertainty. Moderate correlations are found across models between the long rains precipitation change and patterns of SST change in the Pacific and Indian Oceans. Further analysis of the capability of models to represent present-day SST–rainfall links, and any relationship with model projections, may contribute to constraining the uncertainty in projected East Africa long rains precipitation.
The tropical regions, home to approximately 40% of the world’s population (State of the Tropics 2014), a large proportion of developing countries, and some of the most biodiverse areas in the world, are highly reliant on precipitation. The timing, intensity, and frequency of precipitation in the tropics can have a significant impact on livelihoods, particularly impacting freshwater availability and agriculture. For example, agriculture in Africa is almost entirely rain-fed (FAO 2013), highlighting the link between precipitation and food security in the region. Furthermore, using a range of temperature and precipitation indicators Diffenbaugh and Giorgi (2012) identified almost all tropical regions as “hot spots” that show the strongest and fastest responses to climate change.
Under a warming climate the global water cycle will become intensified; global precipitation is projected to increase approximately 1%–3% per °C of warming (Collins et al. 2013). On a global scale this change leads to a general rich-get-richer pattern of change where wet regions become wetter and dry regions drier (Christensen et al. 2013; Chou et al. 2009; Held and Soden 2006; Liu and Allan 2013; Seager et al. 2010); this pattern of change has been observed over recent decades (Allan et al. 2010; Marvel and Bonfils 2013; Polson et al. 2013).
However, this rich-get-richer pattern of change is not seen on a regional scale in future precipitation projections within the tropics (Chadwick et al. 2013a), with other mechanisms such as “warmer-get-wetter” and “upped-ante” also being proposed (Xie et al. 2010; Huang et al. 2013; Chadwick et al. 2013a; Chou et al. 2009). Climate model agreement for precipitation change is lower in the tropics than for other regions, with large areas of little model consensus on the sign and magnitude of change (Knutti and Sedláček 2013; McSweeney and Jones 2013). This is despite strong model agreement for large-scale circulation changes such as weakening of the Hadley (meridional) and Walker (zonal) cells, and expansion of the Hadley cell poleward; these changes are robust across CMIP3 and CMIP5 (phases 3 and 5 of the Coupled Model Intercomparison Project) ensembles (Christensen et al. 2013; Ma and Xie 2013; DiNezio et al. 2013).
To reduce uncertainty in tropical precipitation projections, the mechanisms that drive this uncertainty must first be understood. Chadwick et al. (2013a) assessed the driving mechanisms behind mean tropical precipitation pattern changes in CMIP5 representative concentration pathway (RCP) 8.5 projections. They found that the spatial patterns of the weakening tropical circulation and the increase in atmospheric water vapor with temperature (i.e., wet-get-wetter) components cancel to a large extent, leaving the precipitation anomaly pattern to be dominated by shifts in the regions of convection and convergence, particularly over the oceans. Therefore the model precipitation anomaly does not follow a rich-get-richer pattern of change, unlike large-scale projected changes in precipitation minus evaporation (Seager et al. 2010). Further work by Chadwick et al. (2014) showed that while the direct atmospheric radiative effect of increased CO2 concentrations impacts the mean tropical circulation change (Bony et al. 2013), shifts in convergence zones, which dominate the total precipitation spatial pattern, appear to be primarily driven by surface temperature pattern changes (especially over the tropical oceans). This link between shifting convergence zones and SST changes is also described by Xie et al. (2010) and Ma and Xie (2013).
Previous work has mainly focused on understanding multimodel mean projected rainfall changes (with some exceptions; e.g., Rowell 2012; Ma and Xie 2013; Endo and Kitoh 2014). In this study we build on this by analyzing which processes drive the intermodel uncertainty in CMIP5 regional tropical rainfall projections. Recent work (Grise and Polvani 2014) has shown that uncertainty in some future dynamical atmospheric changes may be independent of uncertainty in estimates of global mean temperature change, and so we assess here how dependent regional rainfall projection uncertainty is on uncertainty in global mean temperature change.
Huang et al. (2013) highlight the need to assess seasonal precipitation anomalies instead of relying on the annual mean. In this paper we apply the annual mean methodology developed by Chadwick et al. (2013a) to seasonal and monthly time scales to assess if its assumptions are still valid, and if so, whether it can be used to better understand which components of the precipitation anomaly drive the large model disagreement in the tropics.
In section 2 the decomposition methodology of Chadwick et al. (2013a) and the climate model data sources are detailed, while in section 3 we assess the dependence of regional precipitation change on global mean temperature and consider the signal-to-noise resulting from low-frequency natural variability. Section 4 shows the results of decomposing the precipitation anomaly at subannual time scales and we assess which mechanisms are driving the intermodel uncertainty in the tropical precipitation projections. In section 5 we demonstrate the application of this paper’s methodology, and its use in understanding the mechanisms driving local intermodel uncertainty, for the East African region.
2. Data and methods
a. Model data
Monthly mean climate model output from a 20-member CMIP5 (Taylor et al. 2012) ensemble (Table 1) was retrieved, covering the periods 1971–2000 and 2071–2100 from the historical and RCP8.5 scenario experiments respectively. Parameters were averaged across these 30-yr time periods to derive climatological values for each month. Following Chadwick et al. (2013a), the model data were interpolated to a common 2.5° × 2.5° grid and only the first ensemble member for each model was utilized.
b. Decomposition framework
The framework utilized in this study was originally developed in Chadwick et al. (2013a, 2014). A discussion on the comparison of the decomposition’s construction in relation to other studies is provided within Chadwick et al. (2013a). It is built upon the assumption that precipitation, P, can be defined as a function of the mass flux from the boundary layer to the free troposphere, M, and the boundary layer specific humidity, q (Held and Soden 2006):
As it is not possible to obtain a reasonable estimate of M (as defined above) directly from climate model data, a proxy for the mass flux (M*) was used instead:
Chadwick et al. (2013a) demonstrated that this is a suitable proxy for the convective mass flux within GCMs at the grid point scale, and that it shows consistent behavior with GCM column-integrated convective mass flux (Mint).
Therefore the projected precipitation anomaly due to climate change, ΔP, can be written as
where M* and q are the climatological values (1971–2000) of proxy mass flux and 2-m specific humidity respectively. Also, ΔP can be decomposed into its individual components as
where Δqcc is the Clausius–Clapeyron change in surface q (expected under fixed relative humidity conditions) and Δqrh is the residual of Δq − Δqcc due to changes in surface relative humidity. Also, ΔM*weak represents the tropics-wide weakening circulation and is proportional to the climatological M*; ΔM*shift = ΔM* − ΔM*weak is the spatial shift component of ΔM. This can be reformulated to describe the total precipitation anomaly in terms of the decomposed precipitation mechanisms:
where ΔPt = M*Δqcc is the thermodynamic change due to Clausius–Clapeyron-driven increases in specific humidity, ΔPrh = M*Δqrh is the change due to near-surface relative humidity changes, ΔPweak = qΔM*weak is the change due to the weakening tropical circulation, ΔPshift = qΔM*shift is the change due to spatial shifts in the pattern of mass flux, and ΔPcross = ΔqΔM* is the cross component of precipitation change, associated with interactions between the other components.
One key assumption of this decomposition methodology is that the tropical circulation weakens through a constant fractional decrease, α, in mass flux across the tropics and thus ΔM*weak = −αM*. This appears reasonable from GCM projections (Chadwick et al. 2013a) and is consistent with a dynamical argument for how the circulation weakens (Ma et al. 2012), but if future work does suggest that the circulation weakens in a nonfractionally uniform way, the current decomposition framework could still be applied. In this case any nonuniform weakening would be included in the shift component. As nonuniform circulation weakening would manifest itself through more convection in some regions, and less in others, it would appear reasonable to view this as shifts, although the interpretation of mechanisms would of course need to be reviewed.
The ΔP intermodel variance, σ2ΔP, can be written as the sum of the covariance matrix for all components:
This allows quantitative assessment of each component’s intermodel variance contribution to σ2ΔP, at both the grid box and regional spatial scales. Furthermore, the impact of correlation between components can be assessed as the residual of the σ2ΔP and the sum of all component variances:
c. Regional definitions
The tropics are defined as the region between 30°S and 30°N. A number of climatological subregions within the tropics have been previously defined (Giorgi and Francisco 2000; Rowell 2013; Lyon and DeWitt 2012) and are utilized in this study (see Fig. 1): East Africa (EAF), the Greater Horn of Africa (GHA), Niño-3.4 (ENSO), west Indian Ocean (WID), east Indian Ocean (EID), and the Indian Ocean dipole (IOD; WID − EID). Here, we also define the coastal waters along the GHA (GHC). For all subregions a land–sea mask is applied to remove grid points of unwanted surface type.
3. Causes of regional precipitation uncertainty across the globe
a. Dependence of precipitation uncertainty on global mean temperature change
The range of responses in the global mean surface temperature across CMIP5 models spans several degrees (Andrews et al. 2012). This is due to a combination of uncertainty in climate sensitivity, radiative forcing, and ocean heat uptake across models, but for the end of the twenty-first-century RCP8.5 scenario the largest contribution is likely to be from uncertainty in climate sensitivity (Forster et al. 2013). To assess whether global climate sensitivity is a driver of or a constraint on local precipitation change, two normalization methods have been compared.
In method 1, each model’s precipitation change is scaled by its own global mean temperature change. In method 2 all models’ precipitation changes are scaled by the ensemble mean global mean surface temperature change. Methods 1 and 2 give approximately the same magnitude of ensemble mean precipitation change, but 1 removes the dependence of the intermodel uncertainty on global mean temperature change (as in Chadwick et al. 2013a), whereas 2 does not.
The interpretation of normalization method 1 requires the assumption that precipitation anomaly patterns scale linearly for each model with global mean temperature. This assumption is only approximately true (e.g., Held et al. 2010; Chadwick et al. 2013b; Chadwick and Good 2013), but it holds well enough for the purposes of this study.
The intermodel ΔP standard deviations, a measure of the model uncertainty, are shown in Fig. 2 for each normalization method in December–February (DJF) and June–August (JJA). In both seasons the two normalization methods result in very similar patterns of model uncertainty. This similarity is also found in the ensemble mean ΔP values, indicating that the spatial pattern of ΔP is not strongly influenced by global mean temperature change.
The percentage difference in the ΔP intermodel spread (standard deviation) between the two normalization methods is shown for DJF in Fig. 2c and JJA in Fig. 2f. Across much of the globe the differences in model uncertainty are less than 10%, particularly within tropical regions. Slightly larger differences in the region of 20%–30% can be found in some mid to high latitudes, although this is where the intermodel standard deviations (in mm day−1) are small (see Figs. 7 and 8). In the Arctic during DJF, the relatively high dependence of precipitation uncertainty on global mean temperature change is likely to be due to the direct connection between winter sea ice retreat and evaporation and precipitation increases in this region (Bintanja and Selten 2014; Rowell 2012). Similarly, a poleward shift of the Southern Hemisphere midlatitude jet has been shown to be related to model climate sensitivity in summer (Grise and Polvani 2014), so this may explain the relatively strong relationship in the Southern Ocean storm-track region during DJF.
For most regions these values indicate a very weak relationship between uncertainty in global mean temperature change and ΔP intermodel spread. This suggests that dynamical uncertainty unrelated to climate sensitivity dominates precipitation change uncertainty across the globe, and this is investigated further for the tropics in section 4.
Assessment of the two normalization methods has highlighted only minor differences in the ensemble mean ΔP and intermodel spread that will not affect the analysis performed in this study. The methods were also tested on the CMIP5 ensemble regridded to a 10° × 10° grid to assess the sensitivity of the results to resolution; this was not found to impact the results. Therefore in order to include the full range of local precipitation uncertainty, all precipitation change components were scaled only by the ensemble mean global surface temperature change.
b. The relative contributions of model uncertainty and internal variability to total uncertainty in regional precipitation projections
The relative proportions of regional precipitation change uncertainty due to modeling uncertainties and natural variability are now estimated. The methodology follows that of Rowell (2012), applied to the difference between 2071 and 2100 under the RCP8.5 scenario and the preindustrial control mean for the 20 CMIP5 models analyzed in this study. The uncertainty due to natural (unforced) variations of 30-yr mean anomalies, σ2ΔP_Nat, is computed from the preindustrial control runs of each model [for details, see the AO-PPE-A section of appendix 2 of Rowell (2012)], and the uncertainty due to model formulation is then computed as a residual from the total uncertainty computed above: σ2ΔP_Model = σ2ΔP − σ2ΔP_Nat. As with other analysis-of-variance calculations, small negative estimates occasionally occur by chance if the population value of σ2ΔP_Model is close to zero, and so these are reset to zero.
Figure 3 then shows the percentage of the total uncertainty that is due to modeling uncertainty [i.e., 100(σ2ΔP_Model/σ2ΔP)], appropriate to 30-yr mean changes. This percentage varies substantially across different parts of the globe, with a substantial contribution from natural variability in many subtropical and extratropical areas, but model uncertainty is dominant in the tropical and polar regions.
For the changes in the tropical rainy seasons and regions that form the focus of the remainder of this study, the contribution from modeling uncertainty is generally above 80%, and above 90% in many land regions. Therefore, analysis of the intermodel standard deviation can be simplified by viewing it primarily as a property of modeling uncertainty, and neglecting the smaller contribution from natural variability in these regions.
4. Seasonal and monthly analysis of the tropical precipitation anomaly
a. Seasonal and monthly precipitation anomalies
The ensemble mean tropical precipitation anomaly, ΔP, from 1971–2000 to 2071–2100 for RCP8.5 is shown for DJF in Fig. 4a and JJA in Fig. 5a [see appendix B for March–May (MAM) and September–November (SON)]. Figures 4b–f and 5b–f show the corresponding precipitation anomaly components. Throughout the annual cycle, ΔPt, the thermodynamic component, due to increases in specific humidity, shows large increases across the tropics (Figs. 4c and 5c). However, as shown in the annual mean analysis of Chadwick et al. (2013a), the pattern of the total ΔP is not similar to this thermodynamic wet-get-wetter pattern and this result is shown here to hold on a climatological monthly time scale throughout the annual cycle (not shown; see appendix A for an assessment of the framework validity at monthly and seasonal time scales).
The spatial patterns of ΔPt and ΔPweak are both strongly linked to the climatological proxy convective mass flux, M* (Chadwick et al. 2013a). Combining these components, ΔPtw = ΔPt + ΔPweak (Figs. 4g and 5g), provides a residual that is relatively small across the oceans but that can be larger over land where surface warming (and therefore ΔPt) is enhanced. In this way the tropics-wide weakening tropical circulation acts to counteract increases in convective precipitation due to Clausius–Clapeyron specific humidity increases. The surface warming component, ΔPt, is however slightly larger in magnitude than the weakening circulation, and this ΔPtw shows a weak wet-get-wetter spatial pattern of change across all months, unlike ΔP.
The ΔPrh component (Figs. 4e and 5e), associated with the difference between the actual change in specific humidity and the expected change due to surface warming (under Clausius–Clapeyron), shows a seasonal pattern of negative anomalies over land, with the largest magnitudes found in the summer hemisphere. This acts to oppose precipitation increases driven by ΔPt in these regions. There is also a small, year-round positive ΔPrh signal north and south of the equatorial Pacific. The term ΔPtwrh = ΔPt + ΔPweak + ΔPrh (Figs. 4h and 5h) represents the precipitation change due to changes in low-level specific humidity, combined with a weakened circulation. This combination of ΔPt and ΔPrh acts to reduce the precipitation anomaly over land regions in ΔPtwrh compared to ΔPtw.
The cross term, ΔPcross (Figs. 4f and 5f), has a relatively small magnitude both spatially across the tropics and temporally throughout the annual cycle. It is found to generally oppose the climatological precipitation and thus the surface warming–driven (wet-get-wetter) pattern of change. The ΔPweak and ΔPcross terms together almost completely cancel out the ΔPt component over the ocean throughout the annual cycle, ΔPtwrc = ΔPt + ΔPweak + ΔPrh + ΔPcross (Figs. 4i and 5i).
With many of the other components cancelling each other out, the spatial pattern of ΔP is strongly linked to the ΔPshift component (Figs. 4a,b and 5a,b). This component represents spatial shifts in the pattern of convective mass flux and was found to largely dominate the annual mean ΔP pattern (Chadwick et al. 2013a). This strong relationship is found here to also hold at monthly and seasonal time scales; for the ensemble mean, the spatial correlation between ΔPshift and ΔP does not fall below 0.97 for any month. Furthermore, the result is robust across models, with the correlation score between the components always greater than 0.95 and generally above 0.99. This is in contrast to the low spatial correlation between ΔP and ΔPt (Chadwick et al. 2013a).
In DJF and JJA, as well as throughout the annual cycle, it can be seen that the ΔPshift component over the oceans appears to follow the pattern of developing sea surface temperature anomalies (Figs. 4h and 5h). This is known as the warmer-get-wetter mechanism (Xie et al. 2010; Chadwick et al. 2013a; Huang et al. 2013; Ma and Xie 2013). Large increases in ΔPshift track the seasonal zonal migration of the equatorial Pacific SST warming, and also follow patterns of SST change in the Indian Ocean.
Ma and Xie (2013) and Chadwick et al. (2014) demonstrate the causality of the SST pattern change influence on oceanic precipitation pattern change in the tropics using a number of atmosphere-only GCM experiments. In comparison, ΔPshift anomalies over tropical land appear not to be strongly associated with SST pattern change in many regions (Rowell 2012; He et al. 2014). Driving mechanisms behind land convective shifts are likely to include changes in land–sea temperature contrasts (Bayr and Dommenget 2013; Chadwick et al. 2014), local versus remote atmospheric stability response to land warming and SST warming (Giannini 2010; Chadwick et al. 2014), and the response of convection to a reduction in RH and consequent increase of the lifting condensation level (Fasullo 2012).
Over land the magnitudes of ΔPshift and ΔPtwrhc are often more comparable than over the oceans, so the balance of different components that make up the ensemble mean precipitation change can be more intricate. This is examined in detail in section 5 for the regional example of East Africa.
b. Seasonal cycle of precipitation anomalies
The seasonal cycle of the zonal mean precipitation anomaly and the individual components is shown in Fig. 6. The main feature is a large near-equatorial positive anomaly that progresses meridionally over the year, favoring the summer hemisphere. Huang et al. (2013) show a similar seasonal cycle of precipitation change under warming, and attribute the seasonal meridional progression of precipitation anomalies to a thermodynamically driven wet-get-wetter effect. In contrast, by here separating out dynamical changes into two different processes, and accounting for the cancellation between ΔPt and ΔPweak, ΔPshift is found to be slightly more influential than ΔPtw in driving the seasonal progression, although both contribute. Overall, ΔPtw is relatively more important in the zonal mean than at regional scales as many of the shifts in convection have been averaged out, reducing the magnitude of ΔPshift.
From zonal mean anomalies alone, both the Huang et al. (2013) interpretation (where ΔPweak and ΔPshift largely cancel each other out in a single dynamical term leaving ΔPt to dominate the seasonal progression) and the alternative described here (where ΔPt and ΔPweak cancel, leaving ΔPshift to be more influential) appear equally viable. However, from examination of the full two-dimensional spatial pattern of precipitation component anomalies (Figs. 4 and 5) it is clear that the anticorrelation and cancellation between ΔPt and ΔPweak, which spatially are both strongly driven by the climatological precipitation (Chadwick et al. 2013a), is far more comprehensive than that between ΔPweak and ΔPshift. This suggests that the new interpretation described here may be the more physically justifiable.
It is somewhat surprising that zonal mean ΔPshift exhibits a pronounced seasonal meridional progression, as the zonal mean warm SST anomalies remain on the equator throughout the year (Fig. 6f). We hypothesize that the seasonal progression of ΔPshift (Fig. 6b) is driven by two factors. First, the spatial response of convection to SST changes is nonlinear (e.g., Johnson and Xie 2010; Power et al. 2013; Huang 2014; Watanabe et al. 2014), causing ΔPshift to track the summer hemisphere (climatologically warmer) side of the SST anomaly; this forms part of the “modified warmer-get-wetter” mechanism in Huang (2014). Second, a general movement of convection from land to ocean in response to SST increases (Chadwick et al. 2014; He et al. 2014). As specific humidity is generally larger over oceans than over land, this results in a mean increase in qΔM*shift (i.e., ΔPshift). This land–sea effect happens much more in the summer hemisphere (where there is more climatological land precipitation to shift), and therefore also tracks the seasonal cycle. The evidence on which we base our hypothesis for these two effects is shown by use of idealized atmosphere-only GCM experiments in appendix C, but it should be noted that the hypothesized land–sea contrast effect in particular is in need of further investigation.
There is found to be some seasonal variation in the importance of the influencing drivers on the zonally averaged values, with ΔPt becoming relatively more influential from July to October (Fig. 6c). This is likely to be due to weaker equatorial SST increases driving smaller ΔPshift anomalies in these months (Fig. 6f).
c. Causes of intermodel uncertainty
The intermodel standard deviation for each component is shown in Figs. 7 and 8. As seen for the ensemble mean precipitation anomalies (Figs. 4 and 5), the ΔPt and ΔPweak terms (Figs. 7c,d and 8c,d) strongly oppose each other across models, reducing their combined variance to almost zero over the oceans (ΔPtw; not shown). Furthermore, when combined with the ΔPrh component, to form ΔPtwrh (Figs. 7g and 8g), the intermodel spread is reduced even more so over land regions (due to negative covariance across models between ΔPt and ΔPrh over tropical land).
The ΔPt, ΔPweak, and ΔPrh terms do not contribute significantly to the total ΔP intermodel spread. Instead, this appears to be strongly linked across the tropics to the ΔPshift model spread; for the ensemble mean the spatial correlation with ΔP does not fall below 0.99 throughout the annual cycle. Combining ΔPcross with ΔPtwrh makes only a very small impact on the spread compared with ΔPtwrh.
The ΔPt and ΔPweak components scale strongly with global mean warming and have reasonably consistent patterns of change across the ensemble of models. Therefore, if they were a major cause of intermodel spread then the ΔP uncertainty would be strongly related to uncertainty in global mean temperature change. This result is not seen in the CMIP5 ensemble (see section 2). For each individual GCM, the magnitude of ΔPshift does approximately scale with global mean warming as temperature increases in a given emissions scenario, or across different scenarios. However the ΔPshift spread across models is found to be driven by different patterns of convergence zone shifts, presumably not directly related to the processes that cause uncertainty in climate sensitivity. The intermodel spread in ΔPshift is therefore uncoupled from uncertainty in global mean temperature change.
Despite evidence of a strong SST pattern-change influence on precipitation pattern change over the tropical oceans (Ma and Xie 2013; Chadwick et al. 2014), at the tropics-wide scale there is only weak spatial correlation between the intermodel variances in relative SST changes (where the tropical mean SST for each model is removed; Figs. 7h and 8h) and ΔPshift (Figs. 7b and 8b). This is likely to be due to one or more of the following causes. 1) The first is the nonlinear response of convection to SST changes (e.g., Johnson and Xie 2010), which leads to a strongly reduced influence of SST changes on the precipitation anomaly in regions far from the SST convective threshold (e.g., southeastern tropical Pacific). In regions close to the SST convection threshold, such as the western equatorial Pacific, the influence of ΔSST on precipitation becomes much larger, and therefore uncertainty in relative SST changes is collocated with the model uncertainty in ΔPshift in these regions. 2) Also, precipitation pattern change uncertainty can arise through intermodel uncertainty in the atmospheric response to SST pattern change, and the location of this uncertainty will not necessarily correspond to the regions of largest uncertainty in the SST change patterns themselves. 3) Finally, SST patterns are not the only influence on precipitation change over the oceans, and other mechanisms such as land–sea temperature contrasts (e.g., Bayr and Dommenget 2013) or the upped-ante mechanism (Chou et al. 2009) may also be important.
The signal (ensemble mean) to noise (intermodel standard deviation) ratio (SNR) for ΔP is shown in Fig. 9. Low magnitude values (generally less than or around 1) are seen across the tropics, indicating low confidence in the magnitude of precipitation change even in regions where most models agree on the sign of the change. The regions with smaller values indicate regions of greatest intermodel uncertainty and closely match the CMIP5 model agreement results (Knutti and Sedláček 2013; McSweeney and Jones 2013). Smaller SNR values are found near the edges of the larger-scale ΔP patterns; this is to be expected as models are likely to disagree on the detailed geographic location of changes. For all months the SNR is spatially similar to ensemble mean ΔP, and therefore ΔPshift, further highlighting the importance of ΔPshift with regard to tropical precipitation projection uncertainties.
Strong spatial correlations have been found between ΔP and ΔPshift for both the ensemble mean change and the intermodel uncertainty. However, a number of the components are intercorrelated and this could influence the corresponding intermodel ΔP variance.
The ΔP intermodel variance can be interpreted as the sum of the covariance matrix between all components (see section 2). The covariance between a component and itself is the variance of that component (as shown in Figs. 7 and 8), while the covariance between different components will identify the contribution of variance due to correlation. The contributions from each component to the ΔP intermodel variance are shown in Figs. 10 and 11 for DJF and JJA, respectively. Throughout the annual cycle the ΔPshift model variance comprises between ~60% and 80% of the total ΔP variance, and is in general slightly greater over the oceans than the land (often ~60%). The ΔPcross term is very small in all months, as is the variance contribution due to the ΔPtwrh term, although this can become more significant over some land regions.
Over the oceans, the remaining ~20%–40% of the ΔP intermodel variance arises primarily from the covariance residual term (Figs. 10e and 11e); this is the sum of the covariances between all components and is almost entirely driven by the covariance between ΔPshift and ΔPcross and between ΔPshift and ΔPtwrh (Figs. 10d and 11d). Spatially, these components are strongly correlated; therefore the influence of the large ΔPshift uncertainty provides additional variance to the total ΔP variance. Taking this into account, a very large proportion, ~>90% (Figs. 10f and 11f), of the ΔP intermodel variance over oceans is associated with the ΔPshift component (and its covariance with ΔPcross and ΔPtwrh). Over land, ΔPshift still dominates the intermodel variance but to a lesser extent in some regions, primarily due to an increased contribution from the ΔPrh component. Furthermore, none of the other component variance contributions to the ΔP variance are spatially coherent with the total ΔP variance; the relative contributions from the components are relatively constant and the total ΔP variance is predominantly driven by ΔPshift.
The ensemble of models utilized in this report were also regridded to a 10° × 10° grid to assess the sensitivity of the results to resolution. For all seasons, as with the assessment performed using the 2.5° gridded data, the ΔPshift component was found to be the primary driver of ΔP intermodel variance, indicating that the uncertainty stems from larger-scale reorganization of convection instead of smaller-scale local shifts.
These results indicate that in order to understand and attempt to reduce the uncertainty in tropical precipitation projections, it is vital to understand the causes of intermodel uncertainty in convective shifts. Over the oceans much of this may be associated with SST pattern change (Ma and Xie 2013). Over land the drivers are less clear (He et al. 2014), and land–sea warming contrasts are likely to be important (Giannini 2010; Bayr and Dommenget 2013; Chadwick et al. 2014).
5. Regional application: East Africa
The ΔPshift component has been identified as the main driver of the ΔP spatial pattern and intermodel uncertainty across the tropics. This decomposition framework can also be utilized to assess precipitation responses within models in more detail at regional scales, and in this section will be used to analyze drivers of ΔP change and uncertainty over East Africa (see Fig. 1).
The East Africa region experiences a bimodal distribution of precipitation within the annual cycle; these are the long rains of March–May and the short rains of October–December. The region experiences large climate variability with frequent occurrences of excess and deficit rainy seasons impacting livelihoods and food security in the region, the most recent being failure of the rainy seasons in 2010/11. Projected changes in the East African rainy seasons due to climate change are uncertain. There is a tendency toward increased precipitation in both the long and short rains, but with lower confidence in both sign and magnitude in the long rains (Shongwe et al. 2011; Christensen et al. 2013). The decomposition framework is used here to assess which precipitation components are driving the large ΔP intermodel spread for both rainy seasons.
The ensemble ranges for ΔP and each component, for both the short (OND) and long (MAM) rainy seasons over East Africa, are shown in Fig. 12 as a percentage of the climatological precipitation. The full ΔP ensemble range is found to be large in both seasons and can reach up to 20% of the climatological precipitation (note that this is change per degree K of global mean warming, so would represent an uncertainty range of up to 80% of climatological precipitation under global warming of 4 K). In MAM the mean and median ΔP values are small in magnitude but differ substantially due to the skewed ΔP distribution. The interquartile range (IQR) crosses zero, highlighting the model uncertainty and low confidence in the precipitation projections for this season (Christensen et al. 2013). For the short rains (OND), however, the ensemble average is larger and only the full ensemble range, not the IQR, crosses zero, reflecting the higher confidence but still uncertain projections. The ΔPtwrh component is positive for all models in both rainy seasons, and relatively large compared to many other land regions (during their respective wet seasons) due to a small ΔPrh component, and the ΔPcross term is small with a tendency to slightly negative (drying) values (Figs. B1 and B2). The ΔPshift component spans the zero line in MAM, and to a lesser extent OND, and has the largest ensemble spread across the components; it has both the largest IQR and absolute ranges. The magnitude and range of the ΔPtwrh and ΔPcross components are consistent across both rainy seasons; therefore, the better model agreement on the sign of ΔP during OND is driven by seasonal variation in ΔPshift.
The ability of climate models to reproduce the observed annual cycle of rainfall in East Africa is limited (Yang et al. 2014) and warrants further research. For both the short and long rains there is minimal correlation between the climatological precipitation and projected precipitation anomaly across models indicating that the main mechanisms that drive future change in models are different to the mechanisms that lead to present-day model climatology errors.
A benefit of utilizing the decomposition framework for regional analysis is that it highlights the relative importance of the different components depending on which aspect of ΔP is being assessed. In MAM (Fig. 12a) over East Africa it is the ΔPtwrh (thermodynamic residual) component that predominantly drives the magnitude of the ensemble mean ΔP, while the ΔPshift component, together with the variance associated with it, mainly determines the intermodel spread. In OND (Fig. 12b), however, the magnitude of the ensemble mean appears to be equally driven by ΔPshift and ΔPtwrh while the ΔP intermodel spread is once again dominated by the ΔPshift component. In both seasons, ΔPtwrh is relatively large compared to other land regions due to only a small ΔPrh component (see also Figs. 4 and 5).
The contribution of each component’s intermodel variance, and the covariance between components, to the total ΔP variance throughout the annual cycle is shown in Fig. 13. As seen for the tropics-wide analysis (section 4), out of all the components it is ΔPshift (Fig. 13, red line) that provides the largest single contribution to the ΔP model variance, singly accounting for approximately 40%–50%. The ΔPtwrh and ΔPcross components, and the covariance between them, are found to provide only a minor contribution, often less than 5% of the ΔP model variance. However, the covariance between ΔPshift and ΔPtwrh and ΔPcross is not negligible, each accounting for approximately 10%–20% of the total variance over East Africa. Figure 13 (black line) shows that combining the total variance associated with the ΔPshift component, which represents the variance linked to dynamical processes, comprises 80%–90% of the ΔP model variance throughout the annual cycle.
Previous studies of interannual precipitation variability over the East Africa region have highlighted strong relationships with a number of SST anomalies across the equatorial Pacific and Indian Oceans, for the short rains (OND; Shongwe et al. 2011; Christensen et al. 2013), with Indian Ocean SSTs identified as the potentially dominant driver (Behera et al. 2005). Therefore it is plausible that long-term greenhouse gas–forced SST pattern changes in these regions (Fig. 1) could contribute to precipitation change over East Africa.
It has been shown here that the ΔPshift component is the primary driver of model uncertainty in precipitation projections across East Africa. The ΔPshift component is strongly linked to SST changes in many tropical regions (see section 4; see also Chadwick et al. 2014), but not all (Rowell 2012), and it is likely that the main way in which SST pattern change will drive precipitation change is through influencing the position of convective regions (e.g., Xie et al. 2010; Watanabe et al. 2014). Therefore, to remove the influence of other contributing mechanisms (such as thermodynamic moisture increases) on the ΔP signal, the correlation between ΔPshift and SST anomalies in the equatorial Pacific and Indian Oceans (Fig. 1) can be assessed to identify the relationship between climate change SST anomalies and shifts in convection over East Africa.
The correlation scores across models between ΔPshift averaged over East Africa and ΔT for a number of subregions in both the long (MAM) and short (OND) rainy seasons are detailed in Table 2. For the short rains (OND), which exhibit a clear ΔPshift, and therefore ΔP, signal for increased precipitation, there is little correlation with the SST anomalies in the Pacific and Indian Oceans. Instead we find small correlations with the local GHA (negative) and GHC (positive) surface warming, and medium negative correlation with the local land–sea temperature gradient anomaly in the East Africa region. This correlation with the local land–sea surface warming gradient is statistically significant at the 5% level. Those models that show the largest changes in ΔPshift also exhibit the smallest changes in the local land–sea surface warming gradient, and vice versa. This may be due to a feedback effect in which the increased precipitation cools the land surface in the East African region, and needs further investigation.
In contrast, when assessing the long rains (MAM), which exhibits strong model disagreement on the sign of ΔP, we find medium to strong correlations. Excluding the local land–sea gradient anomaly and the Niño-3.4 region, all correlations are statistically significant at the 5% level (Table 2). The correlations increase from the Pacific Ocean through to the Indian Ocean (e.g., Niño-3.4 to the west Indian Ocean), highlighting the relationship with the Indian Ocean, and to a lesser extent Pacific Ocean, SST anomalies and therefore potentially the corresponding Walker cells. Furthermore, the intermodel variance in SST changes over the Indian Ocean is relatively small in the east but increases in the west and northwest (Figs. 7 and 8), particularly in the GHC region, which shows the strongest correlation (0.67) with the East African ΔPshift component. Although these results are different from the previously mentioned studies of interannual variability, which show SST correlations in OND but not MAM, they are consistent with analyses assessing the connection between East African MAM precipitation variability and SST variability over longer multidecadal time scales (Funk et al. 2014; Lyon and DeWitt 2012).
This would imply that the model uncertainty in ΔP during the long rains (MAM) is partly linked to model uncertainty in SST pattern anomalies (and uncertainty in related atmospheric teleconnections) across the equatorial Pacific and Indian Oceans and in particular off the Greater Horn of Africa coast. Therefore the large uncertainty in East Africa long rains precipitation change may potentially be able to be reduced if a constraint can be found on future Indian and Pacific Ocean SST pattern change.
6. Summary and conclusions
a. Tropical precipitation uncertainty
Human activities have been shown to drive changes in the climate system and this will continue and amplify during the twenty-first century. Projected changes in the hydrological cycle are uncertain over the tropics with disagreement on the sign of change for precipitation in many regions. Using a 20-member ensemble from CMIP5 driven by the RCP8.5 “business as usual” greenhouse gas scenario, the spatial patterns of the projected change in seasonal and monthly precipitation, and the intermodel variance, are found to not be strongly influenced by uncertainty in global mean temperature and climate sensitivity. Comparing estimates of intermodel uncertainty with and without accounting for the intermodel spread in global mean temperature change, differences are usually less than 10%, particularly in tropical regions. This means that uncertainty in regional precipitation projections is not primarily a consequence of uncertainty in climate sensitivity, although reduction in underlying GCM model biases or improvement in process simulation may lead to the reduction of uncertainty in both cases.
Applying a decomposition framework to separate different mechanisms of precipitation change, it is found that the future precipitation pattern anomaly does not follow the thermodynamically driven wet-get-wetter pattern of change at seasonal and monthly climatological time scales, consistent with previous results at annual mean climatological time scales (Chadwick et al. 2013a). Combining the thermodynamic component with the tropics-wide weakening circulation and surface relative humidity change components results in a strong cancellation between the three mechanisms, leaving only a weak residual surface warming–driven pattern of change. Because of this cancellation, the spatial pattern of the precipitation anomaly is instead found to be almost entirely driven by spatial shifts in convective mass flux. The spatial correlation between the total change and this component is found to be extremely high (>0.97) across all months for the ensemble mean and for each individual model (>0.95).
Assessing zonally averaged changes in precipitation shows the seasonal cycle to be predominantly driven by these spatial shifts in convergence zones, with a smaller influence from the residual thermodynamic component. We hypothesize that the seasonal migration of the spatial shifts in convergence is driven by 1) the nonlinear spatial response of convection to SST changes causing it to track the summer hemisphere (climatologically warmer) side of the equatorial SST anomaly, and 2) a general movement of convection from land to ocean in response to uniform SST increases, which also favors precipitation increases in the summer hemisphere.
As well as the spatial pattern of ensemble mean precipitation change, the intermodel uncertainty in precipitation change is also mainly driven by spatial shifts in convection and convergence, with extremely high spatial correlations between intermodel variances of the spatial shift component and total precipitation change throughout the annual cycle (>0.99). Over the oceans much of this may be associated with uncertainty in SST pattern change (Ma and Xie 2013) and uncertainty in the atmospheric response to a given SST pattern change. Over land the drivers are less clear (He et al. 2014), with land–sea warming contrasts likely to be important (Giannini 2010; Bayr and Dommenget 2013).
To understand and reduce uncertainty in regional tropical precipitation projections, the focus should be on understanding the causes of these spatial shifts in convection. Analysis carried out on the ensemble data regridded to a 10° × 10° resolution made little difference to the results, indicating that the uncertainty stems from coherent large-scale reorganization of convection rather than smaller-scale local shifts.
b. Application to regional East African rainfall uncertainty
A strength of the decomposition framework and uncertainty analysis methodology is that it can be used to assess precipitation anomalies over any tropical region. As a demonstration of this, applying the framework to the East Africa region found that in both rainy seasons (MAM and OND) the precipitation anomaly intermodel variability is primarily driven by intermodel uncertainty in the spatial shifts in convective mass flux.
During the long rains (MAM) it is the thermodynamically driven residual wet-get-wetter component that drives the magnitude of the ensemble mean precipitation anomaly, whereas in the short rains (OND) it is a combination of both the spatial shifts in convergence and thermodynamic components. Throughout the annual cycle it is found that model uncertainty in the spatial shifts of convergence, and its covariances with the other components, accounts for 80%–90% of the total precipitation anomaly intermodel variability in the East Africa region.
Previous studies have highlighted strong links between East African interannual precipitation variability and remote SST anomalies in OND, so it is plausible that long-term greenhouse gas–forced SST pattern changes may contribute to East African precipitation changes. As the East African precipitation anomaly uncertainty is driven by the uncertainty in spatial shifts in convergence, and to remove the influence of other mechanisms, the relationship between SST anomalies and shifts in convection over East Africa are assessed.
For the short rains (OND) only small correlations with SSTs were found. However, for the long rains (MAM) moderate correlations with all SST anomaly regions analyzed—particularly in the region local to the Greater Horn of Africa, but excluding the local land–sea gradient anomaly and Niño-3.4 SST anomalies—were found. This highlights the relationship of the East African rainfall with the Indian SST anomalies, and therefore potentially the corresponding Walker cell. We hypothesize that the large uncertainty in projected East Africa long rains precipitation may be reduced with a constraint on future SST pattern changes. Further analysis of the capability of models to represent present-day SST–rainfall links, and any relationship with model projections, may contribute to constraining the uncertainty in projected East Africa long rains. Further questions that remain to be answered include the following: Why is the component of spatial shifts in convergence systematically different between the long and short rains? Why do we see such contrasting teleconnections between interannual and long-term precipitation changes with SST anomalies? How do the main mechanisms that drive future change in models relate to the mechanisms that lead to present-day model climatology errors? And why is the reduction in relative humidity over East Africa so small compared to other tropical land regions?
The authors were supported by the Joint Department of Energy and Climate Change (DECC) and the Department for Environment, Food and Rural Affairs (Defra), Met Office Hadley Centre Climate Programme, DECC/Defra (GA01101). We acknowledge the World Climate Research Programme’s Working Group on Coupled Modelling, which is responsible for CMIP, and we thank the climate modelling groups for producing and making available their model output. For CMIP, the U.S. Department of Energy’s Program for Climate Model Diagnosis and Intercomparison provides coordinating support and led development of software infrastructure in partnership with the Global Organization for Earth System Science Portals. We thank Jamie Kettleborough, Ian Edmond, and Emma Hibling for developing the software used to download CMIP5 data, Gill Martin and Kirsty Lewis for useful scientific discussions, and two anonymous reviewers for helping to significantly improve the manuscript.
Validity of Decomposition Assumptions on Monthly Time Scales
The validity of using the proxy mass flux M*, and its relationship with Mint, at climatological annual mean time scales is detailed in Chadwick et al. (2013a); in this section we discuss this relationship at climatological monthly time scales.
The strong relationship between M* and Mint identified in Chadwick et al. (2013a) is found to be consistent throughout the annual cycle (Fig. A1), for almost all models and both time periods. Although they have different units, they behave in a similar manner, evidenced by the strong positive spatial correlation values. However the correlation scores for the IPSL-CM5B-LR model (see model expansions in Table 1) were found to be close to zero, indicating that M* is not a good proxy for Mint in this model. The IPSL-CM5B convection and related physical parameterizations are quite unusual within the CMIP5 ensemble (Hourdin et al. 2013), and this is likely to account for its different behavior here. On closer examination it was found that over land the pattern of Mint in this model was not aligned with the pattern of rainfall, explaining the low correlation with M*. Possibly the convective mass-flux variable in this case also includes terms from other parameterization schemes, and so this model was not used in this analysis.
For the ensemble mean of the remaining models, the correlation does not fall below 0.92. Across the model ensemble, for 1971–2000, the median correlation score is 0.91, with minimum and maximum model values of 0.83 and 0.98 respectively. Positive correlations are also found between Δ Mint and ΔM* across the models albeit with generally smaller magnitudes; the median model value is 0.89 with a minimum value of 0.53 and a maximum of 0.98.
Values of Mint are found to be approximately twice the size of the corresponding M* values; the ratio of Mint to M* across the tropics is generally unimodal and with median values of ~0.5 (Fig. A2). The use of M* as a proxy is not valid for all regions within the tropics (Chadwick et al. 2013a); the relationship is approximately linear with some scatter of values closer to unity or higher (Figs. A1 and A2), indicating nonconvective precipitation. These values, which are several standard deviations away from the median value (Fig. A2), are predominantly found over high elevation regions and, depending on the season, at the latitudinal edge of the tropics in which the amount of nonconvective precipitation is larger (e.g., North America and northern South Asia). The decomposition framework should therefore not be used to assess precipitation change and uncertainties in these regions.
The coefficient of the tropics-wide circulation weakening, α, computed using both Mint and M* for each model and the ensemble mean, is consistent with the values derived in Chadwick et al. (2013a). The ensemble mean αM* is identical to that derived from a smaller ensemble (Chadwick et al. 2013a), highlighting the robust weakening of the tropical circulation across the CMIP5 models. Both αM* and αMint values vary throughout the annual cycle (Table A1). In relative terms, αMint has larger subannual variability than the αM*. This is to be expected as even though both represent convection, they are not identical (see also Chadwick et al. 2013a) and can therefore vary with different magnitudes throughout the annual cycle, and in their response to warming. The strong positive spatial correlation between Mint and M* remains for both the 1971–2000 and 2071–2100 time periods.
Given the strong spatial correlation between Mint and M*, the consistency of the relationship at subannual time frames and across time periods, and identification of regions where the approximation may not be valid (e.g., over high elevations), we are confident that the use of M* as a proxy for convective mass flux is applicable to subannual time frames. Furthermore, this framework was developed based on tropics-wide mechanisms, as opposed to region-specific mechanisms, and is not therefore dependent on seasonality by construction.
MAM and SON Seasonal Precipitation Components
The ensemble mean precipitation anomaly, ΔP, from 1979–2000 to 2071–2100 for RCP8.5, is shown for MAM in Fig. B1 and SON in Fig. B2, alongside the separated ΔP components derived using the framework and the associated change in surface temperature. As for DJF and JJA, the transition seasons of MAM and SON show a strong cancelation of the wet-get-wetter pattern of change (Figs. B1 and B2), leaving the spatial pattern of ΔP to be driven by the ΔPshift component.
Understanding the Seasonal Cycle of Precipitation Anomalies Using Idealized Experiments
Section 4 described the seasonal cycle of precipitation anomalies in RCP8.5, and the processes that drive this seasonal progression are now examined in more detail through the use of idealized experiments. Four additional atmosphere-only experiments were used from CMIP5: 1) amip—forced with time-varying SSTs and greenhouse gases over the period 1979–2008; 2) amip4K—as in amip, but with a +4-K uniform SST anomaly; 3) amip4xCO2—as in amip, but with CO2 concentrations quadrupled (note that only the radiation scheme in each model, not plant respiration/transpiration, is allowed to directly adjust to this increase in CO2); and 4) amipFuture—as in amip, but with a patterned SST anomaly taken from the CMIP3 ensemble mean 1% yr−1 CO2 increase experiment, normalized to have a global mean anomaly of +4 K. Eleven models were analyzed here, and are indicated by one or two asterisks next to the model name in Table 1.
Anomalies are taken as the 30-yr mean of each experiment minus the 30-yr mean of the amip control run. The experiments are further partitioned by subtracting amip4K from amipFuture to form amipPattern, which represents the response to only the pattern of future SST anomalies (assuming linearity of response to the combination of uniform and pattern SSTs). Finally, amipTot is the sum of amipFuture and amip4xCO2 and represents the response to the sum of uniform SST, pattern SST, and CO2 forcing (again assuming a quasi-linear framework).
The experiment amipTot (Fig. C1a) has a seasonal progression very similar to that of RCP8.5, though with a different magnitude, because amipTot is not normalized by the global mean temperature change (because this normalization is not appropriate for amip4xCO2). This suggests that analysis of the individual amip experiments can provide information relevant to the previous results shown for RCP8.5 in section 4.
The ΔPshift and ΔPtwrhc (Figs. C1b,c) components contribute approximately equally to the magnitude of amipTot throughout the seasonal cycle, in slight contrast to RCP8.5 where ΔPshift makes the larger contribution. Note that ΔPtwrhc is mainly associated with a large positive anomaly in amip4K (Fig. C1i) and a smaller negative anomaly in amip4xCO2 (Fig. C1l). This component may be larger in amipTot than RCP8.5 because of the lack of nonlinear interaction in amipTot between SST warming and the component of the weakening circulation associated with the direct CO2 effect. Note that amipPattern also makes a small contribution to this term (Fig. C1f) due to nonlinear interactions between SST pattern change and uniform SST warming being represented in ΔPcross [see Chadwick and Good (2013) for a more detailed analysis of the nonlinear interactions between precipitation change components].
The ΔPshift comes largely from a combination of amipPattern (Fig. C1e) and amip4K (Fig. C1h). The amipPattern shifts are associated with SST pattern change, and can be understood under the “modified warmer-get-wetter” framework (Huang 2014; Xie et al. 2010). The shifts associated with amip4K are less well understood, but we propose that they are largely an effect of the redistribution of convection from land to ocean that occurs in response to a uniform SST warming, although we also expect different sensitivities of convection to uniform warming in different ocean regions. As more moisture is generally available over ocean than land, this leads to a mean increase in the amount of precipitation. This anomaly follows the seasonal cycle, as the redistribution of convection follows the climatology of convective regions as they progress between the hemispheres. This effect is seen in reverse in amip4xCO2 but with a much smaller magnitude (Fig. C1k).