Seasonal variability and predictability of monsoon precipitation in Southern Africa

Rainfed agriculture is the mainstay of economies across Southern Africa (SA), where most precipitation is received during the austral summer monsoon. This study aims to further our understanding of monsoon precipitation predictability over SA. We use three natural climate forcings, El Niño–Southern Oscillation, Indian Ocean Dipole (IOD), and the Indian Ocean Precipitation Dipole (IOPD)—the dominant precipitation variability mode—to construct an empirical model that exhibits significant skill over SA during monsoon in explaining precipitation variability and in forecasting it with a five-month lead. While most explained precipitation variance (50%–75%) comes from contemporaneous IOD and IOPD, preconditioning all three forcings is key in predicting monsoon precipitation with a zero to five-month lead. Seasonal forecasting systems accurately represent the interplay of the three forcings but show varying skills in representing their teleconnection over SA. This makes them less effective at predicting monsoon precipitation than the empirical model.


Introduction
Southern Africa (SA) is a drought-and flood-prone region where over 95% of agriculture relies on seasonal precipitation primarily occurring during the austral summer monsoon [1][2][3][4][5][6].Most of the precipitation during the peak of the monsoon season (December to February; DJF) is enhanced by tropical lows and, in some cases, tropical cyclones that form within the tropics and move westward over Africa [2,[7][8][9].
Earlier studies suggested that El Niño-Southern Oscillation (ENSO) and Indian Ocean Dipole (IOD) are the two primary mechanisms contributing to SA monsoon seasonal variability [8,[10][11][12][13], while Madden Julian Oscillation is responsible for intraseasonal variability [14,15].Beyond providing contemporaneous forcing, the IOD has also been shown to predict austral summer precipitation with several months lead [11].However, IOD and ENSO are usually not independent of each other.Therefore, it is only possible to independently attribute precipitation variability to one while accounting for the other's influence.Recent studies also suggest that the leading mode of precipitation variability in the Indian Ocean (IO) can mediate the impacts of tropical forcings, such as ENSO, in distant regions during northern hemisphere winter [16][17][18].However, it is currently unknown whether this leading precipitation mode in IO also plays a role in mediating ENSO and IOD influences over SA during monsoon, which overlaps with northern hemisphere winter, or has a distinct role that can be leveraged to predict monsoon precipitation.We address this gap in understanding by constructing an empirical model for the SA monsoon using these three natural tropical forcings.
Accurate seasonal predictability of monsoon precipitation across SA can be a key to sustainable agricultural practices.Evidence of seasonal precipitation predictability over parts of SA relates to Pacific and Indian Oceans sea surface temperature (SST) variability [11,14,[19][20][21].However, while there have been previous summer monsoon predictability studies using both empirical and dynamic models over SA [14,19,[21][22][23][24][25][26][27][28][29][30][31][32][33][34], neither the sources of predictability nor empirical and dynamical models have been fully exploited for predicting monsoon precipitation in SA, particularly at longer lead times.To further our understanding of sources of predictability for the SA monsoon at different lead times, we analyze results from the empirical model after using the three tropical modes of variability as leading precursors and contemporaneous forcings.In addition, we analyze the skillfulness of two seasonal forecasting systems: the Geophysical Fluid Dynamics Laboratory (GFDL) Seamless System for Prediction and Earth System Research (SPEAR) [35] and the European Center for Medium-Range Weather Forecasts (ECMWF) fifthgeneration seasonal forecasting system (SEAS5) [36], in predicting monsoon over SA.The ECMWF and SPEAR are among the prominent forecasting systems in Europe and North America that provide a large ensemble of multi-decadal simulations, which motivates their choice for this study.However, further investigations with other dynamical forecasting systems may be necessary to fully understand the capabilities or limitations of the current generation.
We aim to answer two key questions using the analytical framework of this study: (1) what are the roles of ENSO, IOD, and the dominant IO precipitation mode in monsoon precipitation variability and predictability over SA? (2) How effective are SPEAR and SEAS5 in predicting summer monsoon over SA, and can the skillfulness of seasonal forecast models be explained by their capability or shortcomings to represent the influences of these three natural forcings?

Data and methods
This study uses precipitation and atmospheric variables from ECMWF's Fifth Generation Reanalysis (ERA5) [37].We use the same source for all variables to maintain the data consistency required for teleconnection analysis [38].We analyze monthly precipitation, SST, and three-dimensional atmospheric winds, divergence, and vertical pressure velocity.ERA5 precipitation compares reasonably with the Climate Research Unit (CRU) Timeseries 4.07 [39].However, a substantial disparity exists between CRU and Climate Prediction Center (CPC) [40] over SA (figure S1), with CPC being substantially drier.Due to the lack of consistency in gridded observations and to avoid inconsistency in teleconnection analyses, we use a consistent dataset (i.e.ERA5) for all analyses.
Moreover, two seasonal forecasting systems are analyzed for their skillfulness in predicting the SA monsoon: GFDL's SPEAR with 15 members and ECMWF's SEAS5 with 25 members.In the main text, we provide analysis of zero-, two-, and five-month lead simulations, initialized in December, October, and July for SPEAR, while SEAS5 only has data for zero-and two-month lead simulations, initialized in December and October.The supplementary material contains area-averaged analysis for data initialized for each month between July and December, where available.These lead times allow us to analyze medium and longer-range forecasts consistent with the extent of predictability in previous studies [9,31].The analysis period covers 1991 through 2022, the overlapping period in all three datasets (ERA5, SEAS5, and SPEAR).All data is linearly detrended before use except for climatological analyses to isolate the impacts of natural forcings and remove the influence of climate change from the analysis.
The analyses cover land areas south of 5 • S for the austral summer months (DJF).While the rainy season substantially varies latitudinally across SA, DJF is the region's core monsoon season [1].We investigate monsoon precipitation variability and predictability using three natural modes of variability: ENSO, IOD, and the dominant mode of precipitation variability in the IO, hereafter termed the IO precipitation dipole (IOPD) due to its zonal dipole pattern [17].We define the ENSO index as the Principal Component (PC) of the first empirical orthogonal function (EOF) of monthly SSTs in the Pacific covering 160 • W-80 • E and 10 • S-10 • N (figure S2).The PC-based ENSO index strongly correlates with SST-based Niño indexes.It is preferred over choosing one of the four Niño indexes to minimize issues related to ENSO diversity.The IOD [41] is defined using the standardized difference in SSTs between the Western (50 • E-70 • E, 10 • S-10 • N) and Eastern (90 • E-110 • E, 10 • S-0 • ) IO.Some studies have used the subtropical IOD index to investigate SA's precipitation variability [9,[42][43][44].Our analyses did not find it more relevant than IOD (not shown).IOPD is the PC of the first EOF of monthly precipitation in the IO, covering 40 • E-140 • E and 10 • S-10 • N. Spatially, it shows a zonal dipole pattern in the IO, as shown in figure S2 [17].When not controlling for other variables, IOPD's positive (negative) phase is associated with more (less) precipitation near equatorial Africa and less (more) precipitation in the more southern areas of the region (figure S3).
We use multiple linear regression (MLR), simple Pearson correlation, and partial correlation analyses to investigate the individual and combined influences of three modes of variability on SA monsoon precipitation.Our MLR is defined by an ordinary least squares technique on detrended anomalies at each grid point.A two-tailed T-test determines the significance of regression coefficients, while an F-test determines the statistical significance of the MLR model at each grid point.All results are tested for significance at 95% confidence.The MLR model is further tested for overfitting by comparing the coefficient of determination (R 2 ) and predicted R 2 .For calculating predicted R 2 , we train the model based on all but one year.We then predict the missing year as a one-point data set.The process is repeated for each data point until we have a time series that is completely predicted based on the regression model from which we can calculate an R 2 value.

Results and discussion
The rainy season in SA varies significantly with latitude, from three months south of 20 • S to over six months at 10 • S (figure 1(a)).The seasonal march of monsoon rains over SA starts in November (figures 1(a) and S1, S4), the onset month [1].DJF is the core monsoon season as zonal average precipitation exceeds 2 mm d −1 throughout the latitudinal belt between 5 • S and 30 • S. Monsoon withdraws from most of the region in March (figure S4) [1].The seasonal maximum of average precipitation and its variability occurs at the boundary of the dryline or Congo air boundary (figures 1(b) and (c)) [8].A comparable seasonal precipitation distribution with a low interannual variability is also observed between northern Mozambique and Angola.South of that, precipitation exhibits a latitudinally expanding east-west gradient with higher magnitudes east of the Kalahari Desert and little precipitation in its west.The proportion of precipitation variability relative to the mean increases over regions south of 15 • S (figure 1(b)).The north-south precipitation gradient also extends from mainland Africa to Madagascar.SA receives the most moisture from continental recycling, while the warm IO (figure 1(c)) contributes the major oceanic moisture source to continental precipitation [45].
Several key dynamical features regulate the spatially complex distribution of monsoon precipitation over SA.In the lower atmosphere (850 hPa), these features include the Angolan Low in the northwest, the Mozambique Channel Trough (MCT) between central Mozambique and Madagascar, and the diagonally oriented South Indian Convergence Zone (SICZ) off the southeast coast of SA (figure 1(c)).The upper atmosphere circulations (200 hPa; figure 1(d)) are characterized by a high extending between the South Atlantic and Indian oceans with an approximate center at the border of Zambia, Botswana, and Zimbabwe, commonly called the Botswana High.The role of these dynamical features in maintaining the summer monsoon over SA has been described extensively in several earlier studies [2,7,10,22,29,46,47].
Several studies have investigated how prevailing ENSO and IOD forcing or their preconditioning shape precipitation distribution in SA [4, 10-13, 48, 49], however, despite its proximity, IOPD's role is currently unknown.The three modes exhibit varying contemporaneous correlations throughout the year (figure 2(a)).The strongest and most consistent is the relationship between ENSO and IOPD, which peaks in boreal winter.ENSO also correlates with IOD in the latter half of the year, but this correlation is mostly insignificant after accounting for the effect of IOPD on their relationship, suggesting that IOPD acts as a mechanism for physically connecting SST variability in the Pacific and Indian oceans.The relationship between IOD and IOPD is most substantial during the austral spring.
The contemporaneous simple Pearson correlations of the three modes with monsoon precipitation over SA suggest a similar dipolar influence by ENSO and IOPD, transitioning between negative in the south and positive in the north around 15 • S (figure 2(b)).Ibebuchi [9] found a pattern similar to one associated with the IOPD in our analyses triggered by the IOD at a five month lag.Most of the IOD's influence is positive but statistically significant only in areas north of Mozambique.The similarity between ENSO and IOPD is remarkable but not surprising.In a recent study, it was determined that ENSO's influence on boreal winter precipitation variability over some regions is primarily driven by atmospheric diabatic heating anomalies caused by ENSOdriven precipitation variability in the IO because of the strong coupling between IOPD and ENSO at a seasonal scale [17,50].This is also true in SA, where the direct influence of contemporaneous ENSO forcing on monsoon precipitation becomes insignificant after controlling for the effects of IOPD and IOD (figure 2(b)).IOD, in contrast, becomes a significant positive forcing over SA after controlling for ENSO and IOPD, while IOPD's influence remains mostly unchanged over SA's eastern half, north of SICZ, after controlling for ENSO and IOD.
Our analysis shows that predicting monsoon precipitation in SA with significant skill is possible using ENSO, IOD, and IOPD precursors.We demonstrate this by examining the preconditioning of these natural modes in December, October, and July as predictors of the summer (DJF) monsoon in SA.First, we examine their lead simple Pearson and partial correlations with SA's monsoon precipitation to explain their predictive power (figures 2(c) and S5).The strength of correlations among the three modes varies during these months (figure 2(a)), resulting in different contributions to SA's precipitation predictability.In July, correlations between the three natural modes are weaker, which means each mode can have a more distinct and independent role in predicting monsoon precipitation.ENSO and IOPD lead correlations retain dipolar patterns like their contemporaneous correlations (figure 2(c), dotted red).However, unlike ENSO's limited contemporaneous role, controlling for IOD and IOPD retains most of its lead correlation, except for southeast SA, north of SICZ, where lead IOPD forcing has a more significant negative impact (figure 2(c), solid red).The July IOD shows precipitation controls like those in its contemporaneous relationship with the SA monsoon.In October, the strongest IOPD-IOD coupling is observed (figure 2(a)).As a result, the October IOD exhibits a dipolar correlation with precipitation, like ENSO and IOPD (figure 2(c); dotted blue, figure S5).After controlling for the other two factors, the IOPD correlation becomes mostly negative, while the IOD correlation becomes positive.ENSO retains a dipolar influence pattern.In December, dipolar patterns persist, although IOD's negative influence is relatively insignificant (figure S5).ENSO (IOD) partial correlations with the precipitation are mostly negative (positive), while the IOPD relationship remains dipolar (figure 2(c); solid green).These analyses suggest that all three forcings play a role in monsoon variability in SA, and their interplay helps determine its predictability.
Next, we construct an MLR model using ENSO, IOD, and IOPD as independent variables or predictors and gridded precipitation over SA as the dependent variable to examine the extent to which precipitation variability can be explained through their contemporary forcings or preconditioning.Given the latitudinal contrast in their influences, the results are presented spatially (figure S6) and in zonal averages (figure 3).Several key points from this analysis can be summarized: (1) the strong coupling between ENSO and IOPD in DJF eliminates the independent role of contemporary ENSO forcing on monsoon precipitation beyond what is already propagated by IOPD (figure 3(b); black).As for the ENSO preconditioning, its predictive power is also the weakest, with statistically significant influence limited to SA's northernmost and southernmost parts (figure 3(c); black, figure S6). ( 2) The IOD influence is predominantly positive and significant across all latitudes in DJF and with December preconditioning.However, its most robust influence is limited to the northernmost parts, with July and October preconditioning (figure 3(a); blue and red).The IOPD is the most prominent force, exerting strong dipolar influence in the north (positive; 5 • S-12 • S) and south (negative; 17 • S-25 • S), except in October when strong coupling with IOD limits its distinct role in predicting SA monsoon.Spatiotemporally varying roles of these natural forcings suggest that they can counteract or amplify one another's effect.Notably, while empirical model predictability improves in the northern region as we reach shorter lead times over the five months, consistency is less in the southern region.One reason for this could be the inconsistency in IOD influence over the southern latitudes at different lead times (figure 2); however, further investigation is necessary to fully understand the causes.On the other hand, SEAS5 and SPEAR tend to be more consistent in improving predictability over both regions, though they underperform the empirical model at all lead times in both correlation and error metrics (figure S7).
Overall, the empirical model explains ∼50% to >75% (∼35% to >50%) zonally averaged precipitation variability in areas along the 5 • S-12 • S (17 • S-25 • S).Spatially, the skill is notably higher in southern Kenya and Tanzania, certain parts of Zambia, and within the region north of SICZ, encompassing parts of Mozambique, Zimbabwe, and Botswana (figure S6).However, between 12 • S and 17 • S, predictability is limited due to the fluctuation between negative and positive influences and inherent low precipitation variability (figures 1(b) and 3(a)).Regions with low predictability skills include northern Mozambique, central Zambia, southern Malawi, and southern Angola.
How do these natural modes of variability influence the monsoon circulation that eventually impacts precipitation?We explain it by regressing threedimensional divergence, vertical pressure velocity, and 850 hPa winds onto ENSO, IOD, and IOPD indexes.The divergence and vertical pressure velocity represent the background monsoon environment, while lower atmospheric winds indicate anomalies that may lead to enhanced or reduced moist flow over land areas.In DJF, the positive phase of IOD enhances the moist flow in the lower atmosphere from continental Africa and the anomalously warm IO through the Mozambique Channel and strengthens the deep convective environment throughout the region except over Angola, where lower-level subsidence induced by the positive phase of IOD suppresses convection (figures 3(b) and (c)).A slightly weaker but similar dynamic anomaly pattern dictates a similar precipitation response with December IOD preconditioning (figures S6 and S8).The enhanced moist flow from IO is absent with October and July IOD preconditioning, and so is its robust positive influence over latitudes south of 20 • S (figure S6).These analyses reveal anomalous moist flow from tropical IO as a critical factor in determining monsoon precipitation variability over southern latitudes in SA.
On the other hand, IOPD's contemporaneous forcing weakens MCT and SICZ, limiting moist flow over southeast Africa and reducing precipitation.In areas south (north) of 15 • S, it suppresses (intensifies) the deep convective monsoon environment, weakening the southward seasonal march of moist continental air.IOPD's December preconditioning yields similar dynamical and precipitation anomalies.However, IOPD's dipolar influence on precipitation is missing with October preconditioning (figures S6 and S8).This inconsistency in IOPD's influence is understandable.In October, IOPD strongly couples with IOD.Therefore, accounting for this coupling in MLR, the partial regression coefficient associated with IOPD forcing shows a substantial reduction in moist flow from IO, which cuts off regular moisture supply throughout the region and leads to seasonal precipitation deficit.
During DJF, ENSO and IOPD are strongly linked (figure 2(a)).When MLR accounts for this coupling, ENSO's influence is mainly confined to the southernmost latitudes over South Africa.Here, dry air entrainment from the southern Atlantic Ocean leads to reduced precipitation.In other areas, ENSO has limited independent influence on the background monsoonal environment (figure 3).However, at other lead times, the coupling between ENSO and IOPD has varying strengths (figure 2(a)).Therefore, ENSO exhibits a slightly more substantial influence.Nonetheless, dry entrainment from the southern Atlantic Ocean remains the unique and key ENSOinduced dynamical anomaly that causes consistent precipitation variability in southernmost regions at all lead times (figures S6 and S8).
We will now examine two seasonal forecasting systems, SPEAR and SEAS5, for their skillfulness in predicting monsoon precipitation over SA within the context of the three identified forcings (figure S9).We begin by analyzing correlations between the modes and note that models' ensemble mean can represent their varying relationships.For instance, the IOD-IOPD correlation is strongest in October (SEAS5 = 0.89; SPEAR = 0.89), as seen in reanalyses (ERA5 = 0.86).Similarly, the ENSO-IOPD coupling increases in December (SEAS5 = 0.81; SPEAR = 0.74), consistent with reanalyses (ERA5 = 0.75).July's initialized SPEAR ensemble mean also shows a relatively weak correspondence between the three modes.Note that SEAS5 does not provide forecast data for DJF with the July initialization.Despite reasonably representing the interplay of ENSO, IOD, and IOPD, models exhibit biases in representing their influences on SA monsoon variability.The biases are particularly severe in SPEAR.Compared to SEAS5, it lacks skill in representing the latitudinal variability in Pearson correlations of three indexes with DJF precipitation (figures 4 and S10, S11).Additionally, SPEAR and SEAS5 exhibit an overly strong influence of ENSO on precipitation variability across SA.Moreover, they show no significant influence of IOPD over SA's southeast in partial correlations after accounting for the effects of IOD and ENSO (figures 4 and S10, S11), which contrasts with reanalyses (figures 2 and S5).Similarly, IOD's positive association with precipitation over latitudes south of 17 • S is also missing in its partial correlations (figures 2 and 4).Both models show a negative IOD relationship instead.On the other hand, ENSO's partial correlations are overly strong in both cases.SPEAR also fails to accurately represent the IOD and IOPD lead influences in July initialized simulations.
Given the modeling errors in representing IOD, IOPD, and ENSO teleconnection across SA during monsoon (figure S9), its predictability is lower in dynamical models than in the empirical model.Because of the contrasting influences of the three modes along the latitude, we assess the predictability of two seasonal forecasting systems by analyzing time series of area average precipitation over two regions, one between 5 • S and 12 • S (northern region) where influence is predominantly positive and one between 17 • S and 25 • S (southern region) where influence is predominantly negative (figures 4(c) and S12).The empirical model using ENSO, IOD, and IOPD as predictors separately for July, October, and December can account for 52%, 64%, and 79% of precipitation variance in the northern part.It can also explain 46%, 44%, and 58% of precipitation variance in the southern part.Please note that the empirical model does not suffer from overfitting, as the predicted coefficient of determination (R 2 ) closely follows the actual R 2 in all instances (figure S13).While, by definition, the predicted R 2 is always lower than the actual R 2 , a significant difference between the two could indicate an overfitting issue in the MLR model.
Due to strong coupling with IOD, IOPD loses most of its independent influence over the southern part with October preconditioning.This results in a lower predictability with the October lead.SPEARexplained precipitation variance is 26%, 38%, and 25% in the northern part and 0%, 25%, and 23% in the southern part with July, October, and December initializations.Similarly, SEAS5-explained precipitation variance is 56% and 69% in the northern part and 46% and 34% in the southern part with October and December initializations.The substantially lower skill in SPEAR compared to SEAS5 is due to its much lower skill in representing the influences of IOD and IOPD over both areas and ENSO influence over the southern parts (figure 4).

Summary
We have constructed an empirical model using three natural climate variability modes related to SST and precipitation variability in the Pacific and Indian oceans (ENSO, IOD, IOPD).This model effectively explains 50%-80% of precipitation variance across SA during the core monsoon season.Furthermore, these natural modes can also be utilized to predict monsoon precipitation in SA with a five-month lead time.The three modes exhibit varying coupling strengths throughout the year.For instance, IOPD is strongly coupled with IOD in the fall and ENSO in the winter.These interdependencies between the three modes influence their distinct roles in shaping precipitation variability over SA.A significant component of ENSO's contemporaneous teleconnection with SA is indirectly through its coupling with IOPD in the IO.Its partial correlation with IOPD after controlling for IOD is 0.58.Direct ENSO forcing does not contribute anything substantial beyond IOPD's teleconnection pattern.IOPD's forcing significantly contributes to precipitation predictability at different lead times.Its weakest predictive power is in October because of the strong coupling between IOD and IOPD, which limits IOPD's distinct role.
The monthly correlations between ENSO, IOD, and IOPD in SEAS5 and SPEAR are consistent with those shown in the reanalysis.They also capture ENSO and IOPD interannual variability with correlations >0.82 for all time periods.However, their ability to describe IOD's interannual variability is less effective, with 0.63, 0.73, and 0.81 correlations between ERA5 and lead-0 SPEAR IOD in July, October, and December, respectively.SEAS5 performs better in analyzing IOD with lead-0 correlations of 0.97 and 0.96 in October and December.Moreover, SPEAR fails to capture the IOD and IOPD teleconnections in SA, particularly below 17 • S. As a result, both models underperform compared to the empirical model, with SPEAR being the least effective in SA's southeast.These findings highlight the importance of accurately representing ENSO, IOD, and IOPD teleconnections to enhance predictability in seasonal forecasting systems during the summer monsoon in SA.Some aspects need further investigation regarding the conclusions drawn in this study.Firstly, the characteristics of precipitation variability and the factors that drive it need to be investigated more thoroughly using gridded observations and other reanalyses.The empirical model formulation used in this study is limited to the forcings related to the Pacific and Indian Oceans.However, the role of other tropical modes of variability may need to be explored.Lastly, other seasonal forecasting systems need to be evaluated to better understand the current state of the dynamic forecasting skills and limitations in SA.
Notice: This manuscript has been authored by employees of UT-Battelle, LLC, under contract DEAC05-00OR22725 with the US Department of Energy (DOE).Accordingly, the publisher, by accepting the article for publication, acknowledges that the US government retains a nonexclusive, paid-up, irrevocable, worldwide license to publish or reproduce the published form of this manuscript, or allow others to do so, for US government purposes.DOE will provide public access to these results of federally sponsored research in accordance with the DOE Public Access Plan (www.energy.gov/downloads/doepublic-access-plan).
This research used the OLCF resources, a DOE Office of Science User Facility supported under Contract DE-AC05-00OR22725.This work is supported by the U.S. Air Force Numerical Weather Modeling Program and NCCR Center, located within the NNCCS at the ORNL and supported under a Strategic Partnership Project 2316T849-08 between DOE and NOAA.

Figure 1 .
Figure 1.(a) The monthly zonal average (5 • E-52 • E, land points) of precipitation.(b) zonally-averaged DJF mean and standard deviation of precipitation, along with the ratio of standard deviation to the mean.(c) Color contours: DJF surface temperature (ocean) and precipitation (land), line contours: DJF precipitation standard deviation (red, blue, white: 1, 2, 3 mm d −1 ), vectors: winds at 850 hPa.The white box indicates the core study area.(d) 200 hPa winds and topography over Southern Africa.Analyses cover 1991-2022 in ERA5 reanalyses.

Figure 2 .
Figure 2. (a) The monthly Pearson (dashed) and partial (solid) correlation between ENSO, IOD, and IOPD.(b) DJF Pearson (second row) and partial (third row) correlations between precipitation over Southern Africa and IOD (left), IOPD (center), and ENSO (right).(c) Zonal average of Pearson (dashed) and partial (solid) correlations of DJF precipitation (5 • E-52 • E, land points) with DJF, December, October, and July indexes of IOD, IOPD, and ENSO.Stippling in (b) and vertical dashed lines in (c) represent statistical significance at the 95% confidence level.

Figure 3 .
Figure 3. Zonally averaged partial regression coefficients describe the influence of contemporaneous and lead (December, October, and July) IOD (a), IOPD (b), and ENSO (c) forcings on DJF precipitation.Circles indicate statistical significance.(d) The coefficient of determination for the multi-linear regression models, shown in (a)-(c).The partial regression coefficients describe the influence of contemporaneous IOD (e), IOPD (f), and ENSO (g) forcings on DJF precipitation (colors) and 850 hPa winds (vectors).Green (purple) contours represent the statistical significance of the zonal (meridional) winds regression coefficient.(h) The coefficient of determination for the multi-linear regression model, shown in (e)-(g).Black boxes indicated the denoted northern and southern regions.(i)-(k) Same as in (e)-(g) but for the zonally averaged vertical cross-section of DJF divergence (multiplied by 10E6) and vertical pressure velocity (multiplied by −50), shown as vectors.The regression coefficients related to vertical pressure velocity are also shown in color.White contours represent the statistical significance of colored contours.Statistical significance indicates the 95% confidence level.

Figure 4 .
Figure 4. Zonally averaged partial (solid) and Pearson (dotted) correlation between DJF precipitation (5 • E-52 • E, land points) and contemporaneous and lead indexes of IOD (left), IOPD (center), and ENSO (right) in (a) SPEAR and (b) SEAS5.The vertical lines represent statistical significance (p < .05).(c) The mean area-averaged precipitation over northern (left) and southern (right) parts of Southern Africa (rectangles in figure 3(h)) in ERA5 (black), regression model (green), SPEAR (orange) ensemble mean, and SEAS5 (violet) ensemble mean.The regression model is based on December IOD, IOPD, and ENSO indexes, while the dynamical models are based on December initializations.Light circles indicate ensemble members in SPEAR and SEAS5.