1. Introduction
Sahel is a semiarid transition zone in Africa between the arid Sahara Desert to the north and the tropical rain forest to the south, where most rain falls from June to September (JJAS) (Folland et al. 1991). The Sahel summer rainfall has a pronounced decadal variability. This decadal persistence has been linked to aggravated impacts. In particular, severe economic losses and millions of deaths were caused during the persistent and devastating drought period in the 1970s and 1980s (Mohino et al. 2011). Previous studies have demonstrated that the decadal variability of the Sahel rainfall is mainly driven by the decadal variations of sea surface temperature (SST) in various parts of the world, such as the Atlantic, Mediterranean Sea, Indian Ocean, and Pacific (Folland et al. 1986, 1991; Bader and Latif 2003; Giannini et al. 2003; Rowell 2003; Lu and Delworth 2005; Zhang and Delworth 2006; Gaetani et al. 2010; Mohino et al. 2011; Villamayor and Mohino 2015; Park et al. 2016). Other possible causes of the Sahel rainfall variations include local land–atmosphere interactions (Giannini et al. 2003) and external forcings such as greenhouse gases and aerosols (Held et al. 2005; Dong and Sutton 2015).
The prediction of Sahel rainfall is usually conducted by coupled climate model integrations initiated from observation-based states. This kind of decadal prediction experiment is part of the European ENSEMBLES project (van der Linden and Mitchell 2009), phase 5 of the Coupled Model Intercomparison Project (CMIP5) (Taylor et al. 2012), and phase 6 of the CMIP (CMIP6) (Eyring et al. 2016). The Sahel region is one of the most skillful locations over land, in particular in terms of precipitation, even if it is not always strongly significant (Kirtman et al. 2013). The prediction systems of the ENSEMBLES project show some skills in predicting the 4-yr mean Sahel rainfall, although the skill is not statistically significant (van Oldenborgh et al. 2012; García-Serrano et al. 2013). The CESM decadal prediction large ensemble has high skills in predicting Sahel rainfall and significant skill improvement over the uninitialized simulation, which is mainly attributed to the high prediction skill of the SST difference between the subtropical North Atlantic and global tropics (Yeager et al. 2018). Some of the CMIP5 prediction systems show prediction skills at lead years 2–5 and 6–9 due to the improvement of decadal SST variations of the Atlantic and Pacific (Gaetani and Mohino 2013). The results from the CMIP5 multimodel decadal hindcasts suggest that accurate SST predictions and realistic relations between SST and Sahel rainfall in the uninitialized simulation (i.e., the historical run for the twentieth-century climate) contribute to the prediction skill of Sahel rainfall (Martin and Thorncroft 2014). Sheen et al. (2017) found that their prediction system can skillfully predict Sahel rainfall on multiyear and interannual time scales, which are regulated by the North Atlantic and Mediterranean Seas and El Niño–Southern Oscillation (ENSO) and western Indian Ocean SST, respectively. In addition, increasing model horizontal resolutions also benefits the prediction skill by better coupling large-scale atmospheric circulations and Sahel rainfall processes (Vellinga et al. 2016).
For the initialization of decadal predictions, assimilating observation-based ocean data is indispensable. Approximately half of the CMIP5 and CMIP6 (CMIP5&6) decadal predictions were initialized using only ocean data, and the rest also used observation-based data of the atmosphere, land, or sea ice (Table 1). Some studies have investigated the role of ocean initialization in the prediction of Sahel rainfall (e.g., Mohino et al. 2016; Yeager et al. 2018), while others have not separated the role of initializing each component when initializing more than one component like the ocean (e.g., Sheen et al. 2017). In this study, we study the role of ocean-only initialization in Sahel rainfall predictions with a climate model called the gridpoint version 2 of the Flexible Global Ocean–Atmosphere–Land System Model (FGOALS-g2). The impact of more realistic initial ocean states on the decadal prediction skill of the Sahel rainfall is investigated. In addition, the decadal prediction skills of the CMIP5&6 models which are also only initialized with ocean data are compared to rule out the influence of the initialization of the other model components. Finally, we explore the possible mechanisms of how SSTs in various ocean basins affect Sahel rainfall predictions.
The initialization components of CMIP5 and CMIP6 models participated in the decadal prediction experiments; “i1” and “i2” are CMIP designations that represent different initialization methods adopted by CanCM4, CCSM4, and NorCPM1. “IAU” stands for incremental analysis update. Models only initializing the ocean component are highlighted in bold.
2. Methods and data
a. Climate model
FGOALS-g2 (Li et al. 2013a) is a CMIP5 model with four components connected by the coupler CPL6 (Craig et al. 2005). It was jointly developed by the State Key Laboratory of Numerical Modeling for Atmospheric Sciences and Geophysical Fluid Dynamics (LASG) of the Institute of Atmospheric Physics (IAP) and Tsinghua University. The atmospheric component is the Gridpoint Atmospheric Model of IAP LASG version 2 (GAMIL2), with a horizontal resolution of 2.8° and 26 vertical levels (Li et al. 2013b). The ocean component is LASG IAP Climate system Ocean Model version 2 (LICOM2), with a horizontal resolution of 1° meridionally increased to 0.5° in the tropics and 30 vertical layers (Liu et al. 2012). The sea ice component is CICE4-LASG, which is an improved version of Community Ice Code version 4.0 (CICE4) (Yu et al. 2008). The land component is version 3 of the Community Land Model (CLM) (Oleson et al. 2010) and has the same horizontal resolution as the atmospheric component.
b. Initialization method
The initialization method is based on the dimension-reduced projection four-dimensional variational (DRP-4DVar) data assimilation (DA). This is an economical 4DVar approach that uses the ensemble technique instead of the adjoint technique to minimize the 4DVar cost function (Wang et al. 2010, 2018). Monthly mean oceanic temperature and salinity analyses from the ds285.3 dataset (Ishii et al. 2005, 2006) are assimilated to produce initial conditions (ICs) for decadal predictions (He et al. 2017, 2020a,b). This method uses a 1-month assimilation window to obtain the optimal IC by minimizing the difference between the forecasts and the observations of monthly mean sea temperature and salinity measured by the cost function. These forecasts are produced from 1-month integrations applying the fully coupled version of FGAOLS-g2 across the assimilation window. The 1-month integration of the fully coupled model included in the minimization procedure is called the constraint of the fully coupled model, simply the model constraint. The advantage of the DRP-4DVar method, compared to the other methods used by the prediction systems listed in Table 1, is that it allows for multicomponent interactions and produces fully coordinated ICs between different model components. The other initialization methods, such as nudging used by the MPI group (Matei et al. 2012) and EnKF adopted by the NorCPM group (Bethke et al. 2021), cannot achieve this in every single analysis for the absence of time dimension in them, thus producing less consistent initial conditions with the coupled model than the DRP-4DVar method. In addition, some models use an ocean initialization obtained from an ensemble of forced ocean model experiments (e.g., Matei et al. 2012; Yeager et al. 2012). This method employs the same ocean model for generating ICs and was shown to provide accurate and dynamically balanced ICs for skillful decadal prediction of SSTs over the North Atlantic and Mediterranean regions.
c. Experiments
The initialization experiment is conducted based on the DRP-4DVar method assimilating the full-field monthly mean oceanic observational temperature and salinity from the ds285.3 analysis dataset (Ishii et al. 2005, 2006). The uninitialized simulation includes three members from the historical simulation up to 2005 and representative concentration pathway 4.5 (RCP4.5) future projection simulation after 2006. The decadal prediction or hindcast experiment is conducted once a year from 1961 to 2005 with 10 ensemble members generated by a time-lagged method for all the model components, i.e., from 1 February to 1 November in the year before each starting year (e.g., 1 February–1 November 1960 for hindcast 1961–70). For example, the ICs for the first (tenth) hindcast member are the model states on 1 February (November) from 1960 to 2013 in the initialization experiment. The uninitialized simulation is used as a benchmark to assess the impact of ICs on decadal prediction skills. All the experiments share the same external forcings.
d. Datasets
Two observed precipitation datasets are selected: Climatic Research Unit time series version 4.04 (denoted as CRU TS) (Harris et al. 2020) and Global Precipitation Climatology Centre full data product version 2018 (denoted as GPCC) (Schneider et al. 2011), both based on worldwide stations. The observed SST dataset is the Hadley Centre Sea Ice and Sea Surface Temperature version 1.1 (denoted as HadISST) (Rayner et al. 2003). The fifth-generation European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis (named ERA5) (Hersbach et al. 2019; Bell et al. 2020) and National Oceanic and Atmospheric Administration (NOAA)–Cooperative Institute for Research in Environmental Sciences (CIRES)–U.S. Department of Energy (DOE) Twentieth Century Reanalysis V3 (named 20CRv3) (Slivinski et al. 2019) are used to verify the atmospheric variables.
To perform an apple-to-apple comparison among the hindcast of FGOALS-g2, which only has ocean initialization and is started once a year for decadal predictions, and those of the CMIP5&6 models, only those CMIP models of which the hindcasts apply ocean-only initialization and are started with the same frequency are selected (Table 2). We note that the 3D ocean states are initialized in all the selected CMIP decadal prediction systems except the IPSL-CM6A-LR prediction system which only initializes the ocean surface states. The historical simulations and the RCP4.5 simulations from CMIP5 as well as the historical simulations from CMIP6 of these models are also chosen. As the prediction skill is highly dependent on the ensemble size (Scaife et al. 2014; Yeager et al. 2018), to make fair comparisons, we selected three hindcast members according to the minimum ensemble size of these hindcasts from all the available ensemble members for each CMIP5&6 prediction system. We also selected three members of the uninitialized simulations from all the available ensemble members for each CMIP5&6 model except only one member available for BCC_CSM1.1 in the RCP4.5 future scenario. The available ensemble members of both the hindcasts and the uninitialized simulations are presented in Table 2.
The related information of the FGOALS-g2 and CMIP5&6 models participated in the decadal prediction experiments with only ocean initialization that began once a year, including their external forcings, horizontal resolutions of atmospheric component, initialization classifications, and numbers of all the available ensemble members of the hindcasts and the uninitialized simulations.
e. Indices
The anomalies of all the considered variables of the ensemble mean model data are first computed by subtracting the climatology of each lead year (Mohino et al. 2016; Bilbao et al. 2021). This method reduces the mean model drift. The climatology of each lead year is computed by averaging all the hindcasts starting from 1961 to 2005 at this lead year. For example, the anomaly fields of the first lead year of the hindcasts are calculated as the 1-yr hindcasts minus their average. To highlight the dominant signal on the decadal time scale, these variables are averaged over 10 years (i.e., lead years from 1 to 10) and linearly detrended (Mariotti and Dell’Aquila 2012). The 10-yr averages of the observations and the uninitialized simulations are computed in the same way as the hindcasts after they are rearranged according to each hindcast period, i.e., 1961–70, 1962–71, …, and 2005–14. The Sahel rainfall index is defined as the normalized detrended JJAS rainfall anomalies averaged over the Sahel region (10°–20°N, 20°W–40°E) for the whole hindcast period (Zhang and Delworth 2006; Wang et al. 2012; Li et al. 2016; Sheen et al. 2017). The decadal variability of Atlantic SST is represented by the Atlantic multidecadal variability (AMV) index (Sutton et al. 2018), which is estimated as the linearly detrended annual SST anomalies averaged over the North Atlantic (0°–60°N, 80°W–0°) (Enfield et al. 2001; Knight et al. 2006; Zhang and Delworth 2006; Mariotti and Dell’Aquila 2012; Martin et al. 2014). The Mediterranean Sea SST index is calculated as the normalized detrended JJAS SST anomalies averaged over the Mediterranean Sea (30°–45°N, 0°–40°E) (Mariotti and Dell’Aquila 2012). The dominant decadal SST variability mode in the Indian Ocean is denoted by the Indian Ocean basin mode (IOBM) index, defined as the normalized time series of the first empirical orthogonal function (EOF) mode of the detrended SST anomalies over the tropical Indian Ocean (20°S–20°N, 40°–110°E) (Yang et al. 2007; Guemas et al. 2013; Han et al. 2014; Dong and McPhaden 2017). The SST decadal variations in the Pacific are indicated by the interdecadal Pacific oscillation (IPO) index, which is computed using the normalized principal component of the leading EOF mode of the detrended boreal winter (NDJFM) SST anomalies in the Pacific region (50°S–50°N, 100°–290°E) (van Oldenborgh et al. 2012; Han et al. 2014; Dong and McPhaden 2017). Note that the SST anomalies are averaged over nine years (i.e., lead years 2–10) for the IPO index due to the lack of data from November to December at lead year 1 of some of the CMIP models selected in this study.
f. Prediction skill assessment
Before computing the prediction skill, all the atmospheric variables of the observational data and CMIP models are interpolated to the horizontal grid of FGOALS-g2, and the SST variable of all the models is interpolated to the horizontal grid of HadISST. The prediction skill is measured by the anomaly correlation coefficient (ACC), ACC difference, root-mean-square error (RMSE), and mean square skill score (MSSS). The ACC is the correlation between the predicted (or simulated) and observed anomalies (Jolliffe and Stephenson 2012). The statistical significance of ACC is tested using a nonparametric moving block bootstrap method (Mignani and Rosa 1995; Mudelsee 2010; Wilks 2019). The time series at each space location are randomly resampled 1000 times with 5-yr blocks to account for serial correlation (Smith et al. 2013; Goddard et al. 2013). Then all the bootstrap correlations generated from the resampled time series are sorted in ascending order. The ACC that lies above the 95th percentile is statistically significant at a 95% confidence level (one tailed). The ACC difference between the prediction and the uninitialized simulation measures the impact of initialization on the prediction skill. The ACC difference between the hindcasts by FGOALS-g2 and the other CMIP5&6 models indicates whether the FGOALS-g2 system initialized with the DRP-4DVar method provides better skill. Statistically significant ACC differences are assessed using the method proposed by Siegert et al. (2017). The RMSE measures the differences between the predictions or the uninitialized simulations and the observations. The MSSS is defined as one minus the ratio of the mean square error (MSE) of the prediction over the MSE of the uninitialized simulation, which also evaluates the added value of initialization (Goddard et al. 2013). The MSE only includes the error variance and excludes the bias error component. A positive MSSS indicates that the prediction is more accurate than the uninitialized simulation, and a negative MSSS suggests the opposite. The statistical significance of MSSS is estimated based on a random walk test introduced by DelSole and Tippett (2016). The above indices have been used by many decadal prediction skill assessments (e.g., Bilbao et al. 2021; Borchert et al. 2021; Volpi et al. 2021).
3. Results
a. Model drift assessment
Before assessing the decadal prediction skill of the Sahel rainfall, we analyze the model drift at the beginning of the prediction. The FGOALS-g2 model drifts toward its climatology during the first several years of the prediction (i.e., initial shock) (Fig. S1a in the online supplemental material) when the hindcast is started using the ICs directly from the initialization experiment. To reduce the initial shock, we reconstruct the ICs (including sea surface height, seawater temperature, salinity, and zonal and meridional current) by replacing their partial climatology with that from the corresponding uninitialized simulation during 1955–2006. In this way, the full internal variabilities in the ICs are retained. The hindcasts start from the ICs containing different proportions (25%, 50%, 75%, 100%) of the climatology of the uninitialized simulation. When 75% of the IC climatology is replaced by that from the uninitialized simulation, not only is the initial shock problem greatly reduced (Figs. S1b–e) but also the best prediction result is obtained (Table S1). It suggests that partly reserving the climatology of the observation in the ICs properly is more beneficial to decadal predictions than completely removing it, which verifies its positive impact on decadal predictions.
The time evolutions of global mean surface air temperature anomalies of the hindcasts by six CMIP5&6 models selected in this study are illustrated in Fig. S2. The hindcasts by BCC_CSM1.1 and CMCC-CM2-SR5 present serious initial shocks during the first several years as these two models apply full-field initialization. The initial shock is reduced in the hindcasts by the other four CMIP models as they adopt anomaly initialization (Table 2). However, the initial shock cannot be completely prevented, as disparities between the hindcasts and the observations still occur in some prediction periods.
To reduce the impact of the initial shock, the assessment of the prediction skill is based on the anomaly fields, i.e., the original field minus the climatology that is computed as a function of the lead year (see section 2e).
b. Decadal prediction skill of the Sahel rainfall
The spatial decadal prediction skill of the Sahel rainfall in the FGOALS-g2 prediction system is illustrated in Fig. 1. The predicted Sahel rainfall shows high ACC skill against both the CRU TS and GPCC observational datasets in almost the whole Sahel region (Figs. 1a,b). The uninitialized simulation has a significantly positive ACC in the western part of the Sahel (Figs. 1c,d), which implies that the western Sahel is partly driven by external forcings, especially in the west, along 10°E, in the southern part of Sahel. The differences in ACC skill between the hindcast and uninitialized simulation (Figs. 1e,f) are statistically significant in most of the Sahel region, especially in the northwestern and southeastern parts. The MSSS (Figs. 1g,h) of the hindcast relative to the uninitialized simulation is significantly positive over large regions of the Sahel and shows a similar pattern to the ACC difference. This result suggests that the decadal prediction skill of Sahel rainfall outperforms the uninitialized simulation in most regions, indicating the benefit of initialization.
(a)–(d) ACC for the detrended JJAS 10-yr mean rainfall anomaly in the hindcast with 10 ensemble members (denoted as HCST) and uninitialized simulation (denoted as UNINIT) of FGOALS-g2 against (left) CRU TS and (right) GPCC data in the Sahel region. (e),(f) ACC difference between HCST and UNINIT of FGOALS-g2. (g),(h) MSSS in the FGOALS-g2 HCST using the UNINIT as a reference. Black dots indicate significance at the 95% confidence level. Grid points with missing values are masked in gray.
Citation: Journal of Climate 36, 7; 10.1175/JCLI-D-21-0729.1
The time evolution of the linearly detrended Sahel rainfall index is displayed in Fig. 2a. The observed rainfall undergoes a marked decadal variation with anomalously high rainfall during the 1960s and since the 1990s and a persistent dry period in between. The hindcast by FGOALS-g2 is in good agreement with the observed variability, with a high correlation of 0.83 and 0.77 with CRU TS and GPCC data, respectively. Nevertheless, the correlation is significantly higher than that of the uninitialized simulation, which is also significantly correlated with observations (0.53 and 0.43). This indicates that the initialization further enhances the prediction skill through more accurate ICs than the uninitialized simulation.
Temporal evolution of the Sahel rainfall index. Black and reddish purple lines represent CRU TS and GPCC data, respectively. Blue and red lines represent the uninitialized simulation (denoted as UNINIT) and hindcast with 10 ensemble members (denoted as HCST) by FGOALS-g2, respectively. The correlations of UNINIT and HCST with CRU TS and GPCC are given in parentheses before and after the slash, respectively. The years on the x axis correspond to the 10-yr centered means. For example, 1965 stands for the 10-yr mean Sahel rainfall index of the 10-yr (1961–70) prediction initialized from 1961. The asterisk stands for the correlation that passes the 95% confidence level.
Citation: Journal of Climate 36, 7; 10.1175/JCLI-D-21-0729.1
The ACC of the Sahel rainfall index of the FGOALS-g2 hindcast with observations is one of the highest among the indices of all the considered hindcasts with only ocean initialization, which also shows significant ACC with observations (Fig. 3a and Fig. S3a). We note that ACC varies widely in the uninitialized simulation of CMCC-CM2-SR5 and IPSL-CM6A-LR for each selection of three members. The ACC for the CMCC-CM2-SR5 uninitialized simulation is even comparable with the FGOALS-g2 hindcast and higher than the CMCC-CM2-SR5 hindcast using the first three members. If we compute the ACC skills of the Sahel rainfall index based on the ensemble mean of all the available members, the CMCC-CM2-SR5 hindcast shows few improvements over the uninitialized simulation (Table S2), which is consistent with Nicolì et al. (2023). This reveals that there is a strong internal variability in the CMCC-CM2-SR5 model that cannot be averaged out with only three members. Besides, the ACC spreads of both CMCC-CM2-SR5 and IPSL-CM6A-LR hindcasts are much smaller than those of their uninitialized simulations. This indicates that the initialization plays an important role in the reduction of the ensemble spread, which is also verified in Nicolì et al. (2023). As the evaluation is based on the anomaly fields, the RMSE skill score excludes the bias component. The RMSE of the Sahel rainfall index of the FGOALS-g2 hindcast is one of the lowest (Fig. 3b and Fig. S3b). We note that there are differences in the ACC and RMSE hindcast skills between the CRU TS reference and the GPCC reference for FGOALS-g2 with 10 ensemble members, and BCC_CSM1.1 and MPI-ESM-LR with the first 3 members. However, the differences for the other prediction systems with the first three members are rather small (Table S3). The reason may be due to some disparities between CRU TS and GPCC after 1990 (Fig. 2). For the FGOALS-g2 hindcast with 10 ensemble members, the time series of the Sahel rainfall index is more consistent with CRU TS than with GPCC from 1990 to 2000, which may lead to higher ACC and lower RMSE referenced to CRU TS than to GPCC. For the BCC_CSM1.1 and MPI-ESM-LR hindcasts (Fig. S4), the time series of the Sahel rainfall index is more consistent with GPCC than with CRU TS after 1990, which may lead to higher ACC and lower RMSE when referenced to GPCC than to CRU TS.
Box-and-whisker plots for (a) ACC, (b) RMSE, (c) ACC difference between the hindcasts and the uninitialized simulation, and (d) MSSS skills of the Sahel rainfall index. The hindcasts and the uninitialized simulations in (a) and (b) are represented by red and blue, respectively. The first column of each panel stands for the results from the 10-member ensemble mean of the FGOALS-g2 hindcasts and the 3-member ensemble mean of the FGOALS-g2 uninitialized simulations. The other columns in each panel represent the results from the 3-member ensemble mean of all the CMIP members. The observational reference is CRU TS. The ranges of the skill metrics with three members are represented by boxes and whiskers. The bottom and upper bounds of the boxes represent the 25th and 75th percentiles. The horizontal line inside the box stands for the median. The lower and upper whiskers outside the boxes mark the minimum and maximum, respectively. The triangles indicate the results using the ensemble mean of the first three members. Statistically significant ACC, ACC differences, and MSSS at a 95% confidence level are denoted by filled triangles. The light black line denotes the value 0 on the y axis.
Citation: Journal of Climate 36, 7; 10.1175/JCLI-D-21-0729.1
The ACC of the Sahel rainfall index of the FGOALS-g2 hindcast is significantly higher than that of most selected CMIP5&6 hindcasts except the MIROC5 and IPSL-CM6A-LR hindcasts with the first three members (Fig. 4). However, a large spread of the ACC difference between the hindcasts by FGOALS-g2 and other selected CMIP5&6 prediction systems occurs for each selection of the three members. Compared with the uninitialized simulation, only the FGOALS-g2 and IPSL-CM6A-LR prediction systems show significantly higher ACC with the first three members (Fig. 3c and Fig. S3c). The wide range of the ACC difference between the hindcasts and the uninitialized simulations of CMCC-CM2-SR5 and IPSL-CM6A-LR possibly results from the large spread of ACC in the uninitialized simulations of these two models (Fig. 3a and Fig. S3a). For MSSS (Fig. 3d and Fig. S3d) of the Sahel rainfall index, the hindcast by FGOALS-g2 has the second largest positive significant MSSS value with the first three members, which is slightly lower than the IPSL-CM6A-LR hindcast. The FGOALS-g2 hindcasts with 10 ensemble members show slightly better MSSS performances than those with the first 3 members when referenced to CRU TS. This results from the fact that the MSEs of these two hindcasts are significantly lower than those of their respective uninitialized simulations, which reveals the added value of initialization. Similar to the ACC difference, the MSSS spans considerably for CMCC-CM2-SR5 and IPSL-CM6A-LR due to the large spread of RMSE in the uninitialized simulations of these two models (Fig. 3b and Fig. S3b).
Box-and-whisker plots for ACC difference of the Sahel rainfall index between the FGOALS-g2 hindcast and the selected CMIP5&6 hindcasts referenced to CRU TS (green) and GPCC (orange) with three members from all the available hindcast members. The ranges of the ACC difference with three members are represented by boxes and whiskers. The bottom and upper bounds of the boxes represent the 25th and 75th percentiles. The horizontal line inside the box stands for the median. The lower and upper whiskers outside the boxes mark the minimum and maximum, respectively. The triangles indicate the results using the ensemble mean of the first three hindcast members. Statistically significant positive ACC differences at a 95% confidence level are denoted by filled triangles. The light black line denotes the value 0 on the y axis.
Citation: Journal of Climate 36, 7; 10.1175/JCLI-D-21-0729.1
c. Relationships between Sahel rainfall and SST
As many studies point out that Sahel rainfall is closely associated with global SST (Folland et al. 1986, 1991; Bader and Latif 2003; Giannini et al. 2003; Rowell 2003; Lu and Delworth 2005; Zhang and Delworth 2006; Gaetani et al. 2010; Mohino et al. 2011; Villamayor and Mohino 2015; Biasutti 2016), we analyze the correlation between the Sahel rainfall index and global SST based on the ensemble mean of the first three members (Fig. 5). Significantly positive correlations in the midlatitude western Pacific, North Atlantic, and Mediterranean Sea and significantly negative correlations in the tropical eastern Pacific, South Atlantic, and Indian Ocean are observed (Figs. 5a,b). The pattern in the Atlantic is very similar to that in the AMV, while an IPO-like pattern is shown in the Pacific. The correlation patterns of the hindcast by FGOALS-g2 with both 10 and the first 3 ensemble members (Figs. 5c,d) are consistent with the observed correlation patterns in the midlatitude western North Pacific, Atlantic, Mediterranean Sea, and Indian Ocean, although the FGOALS-g2 hindcast fails to show a realistic correlation pattern in the southern Pacific. The spatial correlations between the observed and predicted SST–Sahel rainfall correlation patterns with the first three ensemble members in all the ocean regions and the ocean regions excluding the southern Pacific are shown in Table S4. The FGOALS-g2 pattern is one of the two best hindcasts among the selected CMIP5&6 hindcasts over the ocean regions excluding the southern Pacific although it fails to perform better than most CMIP hindcasts over all the ocean regions. In particular, most of the decadal prediction skill of the Sahel rainfall come from the North Atlantic and Mediterranean Sea in both FGOALS-g2 and most of the selected CMIP5&6 prediction systems (Figs. 5f,h–j). However, these CMIP5&6 prediction systems reproduce only the observed Pacific and Indian Ocean SST patterns associated with the decadal Sahel rainfall variability to some very limited degree (Figs. 5e–j). These results indicate that the Pacific and Indian Ocean SSTs are of secondary importance for decadal Sahel rainfall prediction. A similar conclusion has been drawn by Sheen et al. (2017).
Correlation map of 10-yr mean annual SST with the JJAS Sahel rainfall index of (a) CRU TS, (b) GPCC, (c),(d) FGOALS-g2 hindcast with 10 and the first 3 ensemble members, and (e)–(j) the hindcasts by 6 selected CMIP5&6 prediction systems using the averages of the first 3 members. The observed SST in (a) and (b) is HadISST. Statistically significant ACC at a 95% confidence level is denoted by dots.
Citation: Journal of Climate 36, 7; 10.1175/JCLI-D-21-0729.1
We also compare the SST–rainfall relationships of the first 3-member ensemble mean in the uninitialized simulation with observations (Fig. S5). The skills of the uninitialized simulation are mainly due to external forcings. High skills of the uninitialized simulation indicate that external forcing also plays a role while low skills denote the major role of internal variability. Therefore, the comparison of the SST–rainfall relationships in the uninitialized simulation with those in the observations is to identify the prediction skills that might be attributed to external forcings (Doblas-Reyes et al. 2013; Martin et al. 2014). The uninitialized simulation of FGOALS-g2 correctly shows a consistent correlation pattern with that observed in the tropical North Atlantic and subtropical North Pacific. This indicates that external forcings play an important role in these regions for FGOALS-g2. The first three member averages of the CMIP5&6 uninitialized simulations show high correlation skills between the Sahel rainfall and SST in substantial areas. For example, the SST–rainfall correlation pattern is consistent with observations in the North Atlantic by MIROC5, the Indian Ocean by MIROC5 and IPSL-CM6A-LR, and the Pacific by CMCC-CM2-SR5. However, we note that large uncertainties exist in these models in representing Sahel rainfall (Fig. 3) and SST (Fig. S6) when selecting three members from all the available members, leading to a large correlation spread in the rainfall–SST relationships (Fig. S7).
For FGOALS-g2, as Sahel rainfall is highly consistent with SSTs in the Atlantic, Mediterranean Sea, Indian Ocean, and Pacific on the decadal time scale, we then evaluate the SST prediction skill in these four regions. For the AMV and Mediterranean Sea SST indices, the FGOALS-g2 hindcast shows warm phases in the 1960s and 2000s and cold phases from the 1970s to 1990s, as those observed (Figs. 6a,b), with high correlations of 0.85 and 0.78 against the HadISST data, respectively. Both indices are consistent with observations in certain periods (e.g., the 1960s for the AMV index and before 1990 for the Mediterranean Sea SST index) in the uninitialized simulation. The correlations are 0.34 and 0.51 for the AMV and Mediterranean Sea SST index, which pass the 85% and 95% confidence level, respectively. This implies that AMV and Mediterranean Sea SST are partly controlled by external forcings (Giorgi 2006; Doblas-Reyes et al. 2013; Mann et al. 2021). For the IOBM and IPO indices, observations have similar phase transitions to the AMV and Mediterranean Sea SST indices but with opposite signs (Figs. 6c,d). The FGOALS-g2 hindcast is highly consistent with observations, with correlations of 0.76 and 0.86, respectively. These two indices in the uninitialized simulation exhibit nearly the opposite phases to the observations. This indicates that the IOBM and IPO indices are largely determined by internal variability. Thus, ocean initialization plays a more significant role in the accurate prediction of the IOBM and IPO indices than the AMV and Mediterranean Sea SST indices in FGOALS-g2.
Time series of the (a) AMV, (b) Mediterranean Sea SST, (c) IOBM, and (d) IPO indices. Black solid lines represent the HadISST observational dataset. The uninitialized simulation (denoted UNINIT) and the hindcast with 10 ensemble members by FGOALS-g2 (denoted HCST) are represented by sky blue and vermillion lines, respectively. The correlation of each experiment with observations is given in parentheses. The years on the x axis correspond to the 10-yr centered means. The asterisk stands for the correlation which passes the 95% confidence level.
Citation: Journal of Climate 36, 7; 10.1175/JCLI-D-21-0729.1
The FGOALS-g2 hindcast for the decadal SST indices is one of the best among the hindcasts by CMIP5&6 models which only apply ocean initialization (Fig. 7). The spread of the FGOALS-g2 hindcast with three members is small, especially for the AMV and IPO indices. However, a large spread occurs for the IPO index of the hindcasts by MIROC5, CMCC-CM2-SR5, and IPSL-CM6A-LR. In terms of ACC, the ensemble mean of the first three members shows that the FGOALS-g2 hindcast has significant high ACCs with observations for all the SST indices (Fig. 7a). The AMV and Mediterranean Sea SST indices can also be skillfully predicted by the MIROC5, MPI-ESM-LR, CMCC-CM2-SR5, and IPSL-CM6A-LR prediction systems. The IPO index can also be accurately predicted by the BCC_CSM1.1 and BCC_CSM2-MR prediction systems. However, none of these CMIP5&6 prediction systems have significant ACCs for all four SST indices. The ACCs of the FGOALS-g2 hindcast are significantly larger than 66% of the hindcasts by the selected CMIP5&6 prediction systems for the AMV and Mediterranean Sea SST indices while all the hindcasts for the IOBM and IPO indices (Fig. 8). The FGOALS-g2 hindcast with 10 ensemble members further improves the ACC of the Mediterranean Sea SST and IOBM indices compared with the hindcast with the first 3 ensemble members (Fig. 7a). The RMSE skill (Fig. 7b) shows similar performance to the ACC skill.
Box-and-whisker plots for (a) ACC and (b) RMSE of the SST indices predicted by FGOALS-g2 and six CMIP5&6 models with only ocean initialization. The AMV, Mediterranean Sea SST, IOBM, and IPO indices are represented by red, green, blue, and purple, respectively. The first column of each figure stands for the 10-member ensemble mean of the FGOALS-g2 hindcast. The other columns of each figure represent the 3-member ensemble mean from all the available hindcast members. The observational reference is HadISST. The ranges of the ACC and RMSE with three members are represented by boxes and whiskers. The whiskers mark the minimum and maximum. The bottom and upper bounds of the boxes represent the 25th and 75th percentiles. The horizontal line inside the box stands for the median. The red circle, green square, blue diamond, and purple triangle indicate the results using the ensemble mean of the first three members of the AMV, Mediterranean Sea SST, IOBM, and IPO indices, respectively. Statistically significant ACC at a 95% confidence level is denoted by filled symbols. The light black line denotes the value 0 on the y axis.
Citation: Journal of Climate 36, 7; 10.1175/JCLI-D-21-0729.1
Box-and-whisker plots for ACC difference of the (a) AMV, (b) Mediterranean Sea SST, (c) IOBM, and (d) IPO indices between the FGOALS-g2 hindcast and the selected CMIP5&6 hindcasts with three members from all the available hindcast members. The ranges of ACC difference with three members are represented by boxes and whiskers. The bottom and upper bounds of the boxes represent the 25th and 75th percentiles. The horizontal line inside the box stands for the median. The lower and upper whiskers outside the boxes mark the minimum and maximum, respectively. The circles indicate the results using the ensemble mean of the first three hindcast members. Statistically significant positive ACC difference at a 95% confidence level is denoted by filled circles. The light black line denotes the value 0 on the y axis.
Citation: Journal of Climate 36, 7; 10.1175/JCLI-D-21-0729.1
For the ACC difference skill between the hindcast and the uninitialized simulation (Fig. 9), the FGOALS-g2 hindcast shows significantly higher ACC than its uninitialized simulation of all four SST indices not only with 10 members but also with the first 3 members. However, only two or three selected CMIP5&6 hindcasts are better than their uninitialized simulations when using the ensemble mean of the first three members. None of the CMIP hindcasts are superior to their uninitialized simulations consistently for all the indices. A large spread of ACC difference occurs for CMCC-CM-SR5 and IPSL-CM6A-LR when using all the available three members. The MSSS values (Fig. 10) of the FGOALS-g2 prediction system are significantly positive for the AMV index with the ensemble mean of the first 3 members, IOBM and IPO indices with the ensemble mean of the first 3 or 10 members. This suggests that the SST decadal predictions are much more accurate than those made by the uninitialized simulation for these indices. Among the first 3-member ensemble mean hindcasts by the selected CMIP5&6 prediction systems, only the hindcasts by MPI-ESM-LR and IPSL-CM6A-LR for the AMV index, IPSL-CM6A-LR for the Mediterranean Sea SST index, and CMCC-CM2-SR5 for the IOBM index show significantly positive MSSS skills.
Box-and-whisker plots for ACC difference between the hindcasts and the uninitialized simulation of the four SST indices by FGOALS-g2 and six selected CMIP5&6 models with three members from all the available ensemble members. The ranges of ACC difference with three members are represented by boxes and whiskers. The bottom and upper bounds of the boxes represent the 25th and 75th percentiles. The horizontal line inside the box stands for the median. The lower and upper whiskers outside the boxes mark the minimum and maximum, respectively. The circles indicate the results using the ensemble mean of the first three members. Statistically significant positive ACC difference at a 95% confidence level is denoted by filled circles. The light black line denotes the value 0 on the y axis.
Citation: Journal of Climate 36, 7; 10.1175/JCLI-D-21-0729.1
Box-and-whisker plots for MSSS of the SST indices of the FGOALS-g2 hindcast and six selected CMIP5&6 hindcasts using the uninitialized simulation as a reference with three members from all the available ensemble members. The ranges of MSSS with three members are represented by boxes and whiskers. The bottom and upper bounds of the boxes represent the 25th and 75th percentiles. The horizontal line inside the box stands for the median. The lower and upper whiskers outside the boxes mark the minimum and maximum, respectively. The circles indicate the results using the ensemble mean of the first three members. Statistically significant MSSS at a 95% confidence level is denoted by a filled circle. The minimum MSSS of the IPO index for IPSL-CM6A-LR is too small (i.e., −36.5) to be fully expressed in the figure, as we only plot the minimum value of the y axis to −13.0 in all the figures. The light black line denotes the value 0 on the y axis.
Citation: Journal of Climate 36, 7; 10.1175/JCLI-D-21-0729.1
The relationships of the four SST indices with the Sahel rainfall index in the hindcast and uninitialized simulation by FGOALS-g2 are illustrated in Table 3. Observations show that enhanced Sahel rainfall is often related to the positive phases of AMV and Mediterranean Sea SST indices and the negative phases of IOBM and IPO indices. The uninitialized simulation fails to reproduce the observed relationships in a consistent way for all four SST indices. The hindcast by FGOALS-g2 correctly captures all the relationships, which presents statistically significant positive correlations for the AMV and Mediterranean Sea SST indices and negative correlations for the IOBM and IPO indices. The FGOALS-g2 hindcast also shows better performances than its individual member of the uninitialized simulation. It indicates that the initialization plays an essential role in accurately representing these relationships. Thus the high decadal prediction skill of the Sahel rainfall probably depends on not only the high decadal prediction skill of the four SST indices but also their accurate relationships with the Sahel rainfall during predictions to translate the SST decadal variability to Sahel rainfall decadal variability, particularly the IOBM and IPO indices.
Correlation between the Sahel rainfall index and four SST indices for CRU TS, GPCC, the range of three individual members of the uninitialized simulation (denoted as UNINIT range), and hindcast with 10 ensemble members (denoted as HCST) by FGOALS-g2. The positive correlation coefficients for the AMV and Mediterranean Sea SST indices and negative correlation coefficients for the IOBM and IPO indices statistically significant at a 95% confidence level are highlighted in bold.
The predicted rainfall–SST index relationships by FGOALS-g2 with both 10 and 3 ensemble members and the selected CMIP5&6 models are presented in Fig. 11. The FGOALS-g2 hindcast with 10 ensemble members performs better than that with the first 3 ensemble members in representing the relationships between the Sahel rainfall index and the AMV, Mediterranean Sea SST or IOBM index. The first 3-member ensemble mean of MIROC5, BCC_CSM2-MR, CMCC-CM2-SR5, and IPSL-CM6A-LR prediction systems can successfully capture the relationships between the Sahel rainfall index and the AMV or Mediterranean Sea SST index. This possibly results from the high prediction skills of these prediction systems on the Sahel rainfall, AMV, and Mediterranean Sea SST indices. The ACC skills of these three indices in the uninitialized simulations are also significant in MIROC5 (Fig. 3a and Fig. S6), which indicates an important role of external forcings in accurately predicting the rainfall–SST relationships. The first 3-member ensemble mean hindcast by the MPI-ESM-LR model well represents the relationship between the Sahel rainfall index and the IOBM or IPO index, and that by BCC_CSM2-MR and CMCC-CM2-SR5 accurately describe the IPO-rainfall relation. A majority of the selected CMIP5&6 hindcasts initialization fail to capture the relationship between the Sahel rainfall and the IOBM or IPO index with the first three members, which is largely due to the failure of predicting the IOBM or IPO index by these prediction systems. However, none of these prediction systems are capable of representing all the rainfall–SST index relationships reasonably with the first three members. Note that the rainfall–AMV relation in the MPI-ESM-LR hindcast contradicts the result in Mohino et al. (2016), which is possibly due to different validation periods.
Box-and-whisker plots for correlations between the Sahel rainfall index and four SST indices, including the AMV index (red), Mediterranean Sea SST index (green), IOBM index (blue), and IPO index (purple), for observations and hindcasts. The first two columns show the results of observations. The third column stands for the 10-member ensemble mean of the FGOALS-g2 hindcast. The other columns represent the 3-member ensemble mean from all the available hindcast members of FGOALS-g2 and six selected CMIP5&6 hindcasts. The ranges of correlations with three members are represented by boxes and whiskers. The whiskers mark the minimum and maximum. The bottom and upper bounds of the boxes represent the 25th and 75th percentiles. The horizontal line inside the box stands for the median. The red circle, green square, blue diamond, and purple triangle indicate the results using the ensemble mean of the first three members of the AMV, Mediterranean Sea SST, IOBM, and IPO indices, respectively. Statistically significant ACC at a 95% confidence level is denoted by filled symbols. The light black line denotes the value 0 on the y axis.
Citation: Journal of Climate 36, 7; 10.1175/JCLI-D-21-0729.1
The rainfall–SST relationships in the uninitialized simulations (Fig. S7) are generally worse than those in the predictions. Only the relationships between the Sahel rainfall index and the AMV index in FGOALS-g2, the Mediterranean Sea SST index in FGOALS-g2, MIROC5 and CMCC-CM2-SR5, and the IPO index in CMCC-CM2-SR5 and IPSL-CM6A-LR are correctly represented with the first three members. We note that the CMCC-CM2-SR5 model simulates the most realistic rainfall–SST linkages among all the models and prediction systems using the first 3-member ensemble mean (Fig. 5 and Fig. S5). This likely leads to the high ACC skills of Sahel rainfall in the uninitialized simulation of CMCC-CM2-SR5 with the first three members (Fig. 3a). However, a large spread occurs for all the rainfall–SST relationships for the CMCC-CM2-SR5 model (Fig. S7).
To further investigate whether there is an added value of initialization in predicting the rainfall–SST index relationships, individual members of the rainfall–SST index relationships in the uninitialized simulations are presented in Fig. S8. Only one member for the AMV index and one member for the Mediterranean Sea SST index of FGOALS-g2 show significant correlation skills. The selected CMIP5&6 models show similar results or even better performance in their uninitialized simulations (e.g., MPI-ESM-LR, CMCC-CM2-SR5, IPSL-CM6A-LR) (Table S5). These results indicate that these models fail to naturally and consistently reproduce all the rainfall–SST relationships. Combined with the fact that the hindcasts in FGOALS-g2 do show a good representation of the relationships for all indices, which does not occur so clearly in the hindcasts of the other models (Fig. 11), the DRP-4DVar initialization helps improve the representation of the teleconnections.
d. Possible mechanisms
In this section, we investigate how the predicted SST indices possibly affect the Sahel rainfall predictions. Separate composite analyses were performed for each SST index. For the influence of the Atlantic SST, we divided the AMV index into positive (1961–67 and 1993–2005) and negative (1968–92) phases according to the observations in Fig. 6a (black line). The composite difference of the moisture flux at 850 hPa between the positive and negative phases of the AMV index is illustrated in Fig. 12. Both ERA5 and 20CRv3 reanalyses (Figs. 12a,b) show that the anomalous westerly and southwesterly flows lead to increase of moisture transport from the tropical Atlantic to the Sahel, which may result from the northward shift of the intertropical convergence zone (ITCZ) during the positive AMV phase (Zhang and Delworth 2006). Note that the moisture fluxes over the Sahel region in the ERA5 and 20CRv3 reanalyses are inconsistent, and this inconsistency is common between different global reanalysis products, especially on land (e.g., Ramon et al. 2019). The hindcast by FGOALS-g2 (Fig. 12d) is more consistent with the reanalyses in presenting the moisture flux in the tropical Atlantic than its uninitialized simulation (Fig. 12c), although the intensity is slightly weaker than in the reanalyses. We further computed the spatial correlation of 850-hPa zonal moisture flux over the tropical Atlantic (0°–10°N, 50°W–15°W) of each experiment. The spatial correlations of the 10-member ensemble mean hindcast referenced to ERA5 and 20CRv3 are 0.50 and 0.69, respectively, significantly better than those of the uninitialized simulation (−0.36 and −0.54). The FGOALS-g2 hindcast with three ensemble members shows performance similar to that of the hindcast using 10 ensemble members (Fig. S9a). Only the first 3-member ensemble mean hindcasts by MPI-ESM-LR, CMCC-CM2-SR5, and IPSL-CM6A-LR of the selected CMIP5&6 prediction systems show significant spatial correlations over the tropical Atlantic (Fig. S9a). In addition, the spatial correlation of the FGOALS-g2 hindcast with the first three ensemble members is significantly higher than those of almost all the selected CMIP5&6 prediction systems (Fig. S10a). Thus, the FGOALS-g2 hindcast is more capable of representing the moisture transport from the tropical Atlantic during the positive phase of the AMV index than its uninitialized simulation and the selected CMIP5&6 prediction systems, which contributes to enhanced Sahel rainfall.
Composite difference of the 850-hPa moisture flux (m s−1 g kg−1) between the positive and negative phases of the AMV index of (a) ERA5 and (b) 20CRv3 reanalyses, (c) uninitialized simulation (denoted as UNINIT), and (d) hindcast by FGOALS-g2 with 10 ensemble members (denoted as HCST). Only the differences that are statistically significant with a 95% confidence level are plotted.
Citation: Journal of Climate 36, 7; 10.1175/JCLI-D-21-0729.1
For the influence of the Mediterranean Sea SST, there are large discrepancies in the Mediterranean Sea region between the ERA5 and 20CRv3 reanalyses (Fig. S11). We computed the spatial correlation of 850-hPa meridional moisture flux in the Mediterranean Sea region (30°–40°N, 0°–40°E) of each experiment. The spatial correlations of the FGOALS-g2 hindcast referenced to ERA5 and 20CRv3 are 0.27 and 0.42, respectively. The spatial correlation is significantly higher than that of its uninitialized simulation (0.15) referenced to 20CRv3. However, the spatial correlation of the FGOALS-g2 hindcast with the first 3 ensemble members is not as good as that with 10 ensemble members (Fig. S9b). A large spread occurs for each selection of 3 members from all the available 10 ensemble members. This implies that three members are not enough to accurately estimate the spatial correlation. The spatial correlation of the first three hindcast members by FOGALS-g2 fails to have advantages over that by the selected CMIP5&6 prediction systems (Fig. S10b). Combined with the fact that there are large uncertainties for each selection of three members, such as in MIROC5, BCC_CSM2-MR, and IPSL-CM6A-LR, the differences across models might be largely influenced by noise using only the first three members. Thus, the FGOALS-g2 hindcast with 10 ensemble members can reproduce the moisture transport from the Mediterranean Sea during the positive phase of the Mediterranean Sea SST index, leading to increased Sahel rainfall.
Regarding the influence of the Indian Ocean, the negative IOBM index increases the surface temperature gradient between the Tibetan Plateau and the Indian Ocean, which strengthens the upper-level tropical easterly jet (TEJ) (Nicholson 2013; Sheen et al. 2017). The TEJ is stronger over the Sahel in both ERA5 and 20CRv3 reanalyses when the IOBM index is in the negative phase (Figs. 13a,b). The hindcast by FGOALS-g2 (Fig. 13d) presents a stronger TEJ over most of the Sahel region, while its uninitialized simulation (Fig. 13c) fails to show upper-tropospheric easterlies. We also computed the spatial correlation of 200-hPa zonal wind over the Sahel region (10°–30°N, 20°W–50°E) of each experiment. The spatial correlation of the FGOALS-g2 hindcast referenced to 20CRv3 is 0.66, which is significantly higher than that of its uninitialized simulation (0.22). However, both experiments show low spatial correlations with ERA5, i.e., −0.04 for the uninitialized simulation and 0.12 for the hindcast. The low spatial correlation of the hindcast with ERA5 may be due to the inconsistency of the wind difference in East Sahel as well as the uncertainty in different reanalysis products. The spatial correlation over the Sahel region is lower using the first 3 ensemble members than using 10 ensemble members in the FGOALS-g2 hindcast (Fig. S9c). Only the hindcasts by MIROC5, BCC_CSM2-MR, and CMCC-CM2-SR5 of the selected CMIP5&6 prediction systems show significant spatial correlations over the Sahel region (Fig. S9c). The spatial correlation of the FGOALS-g2 hindcast with the first three ensemble members is significantly higher than those of 66% of the selected CMIP5&6 prediction systems when referenced to 20CRv3 (Fig. S10c). The intensified TEJ increases convection over the Sahel region, which is associated with more rainfall over the Sahel (Hulme and Tosdevin 1989; Grist and Nicholson 2001; Sheen et al. 2017).
Composite difference of the 200-hPa wind (m s−1) between the negative and positive phases of the IOBM index of (a) ERA5 and (b) 20CRv3 reanalyses, (c) uninitialized simulation (denoted as UNINIT), and (d) hindcast by FGOALS-g2 with 10 ensemble members (denoted as HCST). Only the differences that are statistically significant with a 95% confidence level are plotted.
Citation: Journal of Climate 36, 7; 10.1175/JCLI-D-21-0729.1
Regarding the influence from the Pacific, the warming of the tropical eastern Pacific can excite an atmospheric Kelvin wave propagating across the Atlantic and Africa, which produces anomalous tropical westerlies (easterlies) in the upper (lower) troposphere over the Pacific, Atlantic, and Sahel, indicating a baroclinic vertical structure of wind anomalies (Palmer 1986). Therefore, the negative IPO index is associated with an anomalous divergence (convergence) over the tropical eastern Pacific and convergence (divergence) over the tropical Atlantic and Africa at low (high) levels (Villamayor and Mohino 2015). This pattern is clearly shown by the ERA5 and 20CRv3 reanalyses (Figs. 14a–d). The hindcast by FGOALS-g2 (Figs. 14g,h) is very close to the reanalyses, although the intensity is slightly weaker, while the divergence and convergence patterns shown in the uninitialized simulation (Figs. 14e,f) are nearly opposite to the reanalyses. This indicates that an anomalous Walker-type circulation subsides over the tropical eastern Pacific and ascends from the Sahel region, which results in increased low-level monsoon westerlies in the tropical Atlantic, leading to wet conditions over the Sahel.
Composite differences of divergent wind (vectors; m s−1) and velocity potential (shading; 106 m2 s−1) between the IPO negative and positive phases of (a),(b) ERA5 and (c),(d) 20CRv3 reanalyses, (e),(f) uninitialized simulation (denoted as UNINIT), and (g),(h) hindcast by FGOALS-g2 with 10 ensemble members (denoted as HCST). The composite differences at (left) 850 and (right) 200 hPa. Only the differences that are statistically significant with a 95% confidence level are plotted. The divergent wind blows from the maximum velocity potential to the minimum velocity potential. Positive (negative) velocity potential stands for divergence (convergence) of the wind (Tanaka et al. 2004).
Citation: Journal of Climate 36, 7; 10.1175/JCLI-D-21-0729.1
4. Conclusions and discussion
The decadal predictions of the Sahel rainfall by the FGOALS-g2 initialized only with ocean analysis data using a DRP-4DVar-based coupled DA method are assessed. Accurate ICs are obtained, which play important roles in skillfully predicting the time evolution of the Sahel rainfall. The ACC skill of the Sahel rainfall index of the FGOALS-g2 hindcast significantly outperforms that of its uninitialized simulation and a majority of the CMIP hindcasts which only apply ocean initialization. The MSSS skill is also one of the best among the CMIP5&6 hindcasts with only ocean initialization. The high decadal prediction skill of the Sahel rainfall mainly comes from two different factors. The first is a good prediction skill in the four SST areas (i.e., Atlantic, Mediterranean Sea, Indian Ocean, and Pacific) known to remotely influence Sahelian rainfall, which in the Indian and Pacific Oceans is largely due to a successful initialization. The SST predictions in these oceans of the FGOALS-g2 hindcast are more advantageous than most ocean-only initialized CMIP hindcasts in terms of ACC, RMSE, ACC difference, and MSSS. The second is a very realistic representation of the teleconnection mechanisms linking those regions with the Sahel by the FGOALS-g2 prediction system. The SST–rainfall relationships of the FGOALS-g2 hindcast are also more accurate than those of most CMIP hindcasts with only ocean initialization. Proper ocean initialization plays an indispensable role in these two factors. The accurate predictions of SSTs and SST–rainfall relationships result in anomalous moisture convergence over the Sahel during the wet periods through the response of atmospheric circulations to the SST.
Further dedicated research is needed to initialize more observation-based data from other climate components (e.g., atmosphere and land) and to explore their impacts on the decadal prediction skill of the Sahel rainfall.
Acknowledgments.
This work is funded by the National Natural Science Foundation of China (41875127 and 42005030).
Data availability statement.
The ds285.3 ocean temperature and salinity analyses can be downloaded from https://rda.ucar.edu/datasets/ds285.3/. The CRU TS and GPCC precipitation datasets are available from https://dap.ceda.ac.uk/badc/cru/data/cru_ts/cru_ts_4.04/ and https://psl.noaa.gov/data/gridded/data.gpcc.html, respectively. The HadISST is available at https://www.metoffice.gov.uk/hadobs/hadisst/. The ERA5 and 20CRv3 datasets can be accessed from https://cds.climate.copernicus.eu/#!/home and https://psl.noaa.gov/data/gridded/data.20thC_ReanV3.html, respectively. The precipitation and SST outputs from the CMIP5 and CMIP6 decadal prediction experiments can be obtained from http://www.ipcc-data.org/sim/gcm_monthly/AR5/Reference-Archive.html and https://pcmdi.llnl.gov/CMIP6/, respectively. The FGOALS-g2 uninitialized simulation output is available from CMIP5 http://www.ipcc-data.org/sim/gcm_monthly/AR5/Reference-Archive.html. The output of the FGOALS-g2 decadal prediction experiments is provided in the open repository Zenodo at https://zenodo.org/record/4759971#.YJ4VaqgzbD4.
REFERENCES
Bader, J., and M. Latif, 2003: The impact of decadal-scale Indian Ocean Sea surface temperature anomalies on Sahelian rainfall and the North Atlantic Oscillation. Geophys. Res. Lett., 30, 2169, https://doi.org/10.1029/2003GL018426.
Bell, B., and Coauthors, 2020: ERA5 monthly averaged data on pressure levels from 1950 to 1978 (preliminary version). C3S CDS, accessed 20 August 2021, https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-pressure-levels-monthly-means-preliminary-back-extension?tab=overview.
Bethke, I., and Coauthors, 2021: NorCPM1 and its contribution to CMIP6 DCPP. Geosci. Model Dev., 14, 7073–7116, https://doi.org/10.5194/gmd-14-7073-2021.
Biasutti, M., 2016: What brings rain to the Sahel? Nat. Climate Change, 6, 897–898, https://doi.org/10.1038/nclimate3080.
Bilbao, R., and Coauthors, 2021: Assessment of a full-field initialized decadal climate prediction system with the CMIP6 version of EC-Earth. Earth Syst. Dyn., 12, 173–196, https://doi.org/10.5194/esd-12-173-2021.
Borchert, L. F., V. Koul, M. B. Menary, D. J. Befort, D. Swingedouw, G. Sgubin, and J. Mignot, 2021: Skillful decadal prediction of unforced southern European summer temperature variations. Environ. Res. Lett., 16, 104017, https://doi.org/10.1088/1748-9326/ac20f5.
Craig, A. P., R. Jacob, B. Kauffman, T. Bettge, J. Larson, E. Ong, C. Ding, and Y. He, 2005: CPL6: The new extensible, high performance parallel coupler for the Community Climate System Model. Int. J. High Perform. Comput. Appl., 19, 309–327, https://doi.org/10.1177/1094342005056117.
DelSole, T., and M. K. Tippett, 2016: Forecast comparison based on random walks. Mon. Wea. Rev., 144, 615–626, https://doi.org/10.1175/MWR-D-15-0218.1.
Doblas-Reyes, F. J., and Coauthors, 2013: Initialized near-term regional climate change prediction. Nat. Commun., 4, 1715, https://doi.org/10.1038/ncomms2704.
Dong, B., and R. Sutton, 2015: Dominant role of greenhouse-gas forcing in the recovery of Sahel rainfall. Nat. Climate Change, 5, 757–760, https://doi.org/10.1038/nclimate2664.
Dong, L., and M. J. McPhaden, 2017: Why has the relationship between Indian and Pacific Ocean decadal variability changed in recent decades? J. Climate, 30, 1971–1983, https://doi.org/10.1175/JCLI-D-16-0313.1.
Enfield, D. B., A. M. Mestas-Nuñez, and P. J. Trimble, 2001: The Atlantic multidecadal oscillation and its relation to rainfall and river flows in the continental U.S. Geophys. Res. Lett., 28, 2077–2080, https://doi.org/10.1029/2000GL012745.
Eyring, V., S. Bony, G. A. Meehl, C. A. Senior, B. Stevens, R. J. Stouffer, and K. E. Taylor, 2016: Overview of the Coupled Model Intercomparison Project phase 6 (CMIP6) experimental design and organization. Geosci. Model Dev., 9, 1937–1958, https://doi.org/10.5194/gmd-9-1937-2016.
Folland, C. K., T. N. Palmer, and D. E. Parker, 1986: Sahel rainfall and worldwide sea temperatures, 1901–85. Nature, 320, 602–607, https://doi.org/10.1038/320602a0.
Folland, C. K., J. Owen, M. N. Ward, and A. Colman, 1991: Prediction of seasonal rainfall in the Sahel region using empirical and dynamical methods. J. Forecasting, 10, 21–56, https://doi.org/10.1002/for.3980100104.
Gaetani, M., and E. Mohino, 2013: Decadal prediction of the Sahelian precipitation in CMIP5 simulations. J. Climate, 26, 7708–7719, https://doi.org/10.1175/JCLI-D-12-00635.1.
Gaetani, M., B. Fontaine, P. Roucou, and M. Baldi, 2010: Influence of the Mediterranean Sea on the West African monsoon: Intraseasonal variability in numerical simulations. J. Geophys. Res., 115, D24115, https://doi.org/10.1029/2010JD014436.
García-Serrano, J., F. J. Doblas-Reyes, R. J. Haarsma, and I. Polo, 2013: Decadal prediction of the dominant West African monsoon rainfall modes. J. Geophys. Res. Atmos., 118, 5260–5279, https://doi.org/10.1002/jgrd.50465.
Giannini, A., R. Saravanan, and P. Chang, 2003: Oceanic forcing of Sahel rainfall on interannual to interdecadal time scales. Science, 302, 1027–1030, https://doi.org/10.1126/science.1089357.
Giorgi, F., 2006: Climate change hot-spots. Geophys. Res. Lett., 33, L08707, https://doi.org/10.1029/2006GL025734.
Goddard, L., and Coauthors, 2013: A verification framework for interannual-to-decadal predictions experiments. Climate Dyn., 40, 245–272, https://doi.org/10.1007/s00382-012-1481-2.
Grist, J. P., and S. E. Nicholson, 2001: A study of the dynamic factors influencing the rainfall variability in the West African Sahel. J. Climate, 14, 1337–1359, https://doi.org/10.1175/1520-0442(2001)014<1337:ASOTDF>2.0.CO;2.
Guemas, V., S. Corti, J. García-Serrano, F. J. Doblas-Reyes, M. Balmaseda, and L. Magnusson, 2013: The Indian Ocean: The region of highest skill worldwide in decadal climate prediction. J. Climate, 26, 726–739, https://doi.org/10.1175/JCLI-D-12-00049.1.
Han, W., and Coauthors, 2014: Intensification of decadal and multi-decadal sea level variability in the western tropical Pacific during recent decades. Climate Dyn., 43, 1357–1379, https://doi.org/10.1007/s00382-013-1951-1.
Harris, I. C., P. D. Jones, and T. Osborn, 2020: CRU TS4.04: Climatic Research Unit (CRU) time-series (TS) version 4.04 of high-resolution gridded data of month-by-month variation in climate (Jan. 1901–Dec. 2019). Centre for Environmental Data Analysis, accessed 20 August 2021, https://catalogue.ceda.ac.uk/uuid/89e1e34ec3554dc98594a5732622bce9.
He, Y., and Coauthors, 2017: Reduction of initial shock in decadal predictions using a new initialization strategy. Geophys. Res. Lett., 44, 8538–8547, https://doi.org/10.1002/2017GL074028.
He, Y., and Coauthors, 2020a: A DRP-4DVar-based coupled data assimilation system with a simplified off-line localization technique for decadal predictions. J. Adv. Model. Earth Syst., 12, e2019MS001768, https://doi.org/10.1029/2019MS001768.
He, Y., and Coauthors, 2020b: A new DRP-4DVar-based coupled data assimilation system for decadal predictions using a fast online localization technique. Climate Dyn., 54, 3541–3559, https://doi.org/10.1007/s00382-020-05190-w.
Held, I. M., T. L. Delworth, J. Lu, K. L. Findell, and T. R. Knutson, 2005: Simulation of Sahel drought in the 20th and 21st centuries. Proc. Natl. Acad. Sci. USA, 102, 17 891–17 896, https://doi.org/10.1073/pnas.0509057102.
Hersbach, H., and Coauthors, 2019: ERA5 monthly averaged data on pressure levels from 1979 to present. C3S CDS, accessed 2 April 2020, https://doi.org/10.24381/cds.6860a573.
Hulme, M., and N. Tosdevin, 1989: The tropical easterly jet and Sudan rainfall: A review. Theor. Appl. Climatol., 39, 179–187, https://doi.org/10.1007/BF00867945.
Ishii, M., A. Shouji, S. Sugimoto, and T. Matsumoto, 2005: Objective analyses of sea-surface temperature and marine meteorological variables for the 20th century using ICOADS and the Kobe collection. Int. J. Climatol., 25, 865–879, https://doi.org/10.1002/joc.1169.
Ishii, M., M. Kimoto, K. Sakamoto, and S.-I. Iwasaki, 2006: Steric sea level changes estimated from historical ocean subsurface temperature and salinity analyses. J. Oceanogr., 62, 155–170, https://doi.org/10.1007/s10872-006-0041-y.
Jolliffe, I. T., and D. B. Stephenson, 2012: Forecast Verification: A Practitioner’s Guide in Atmospheric Science. 2nd ed. John Wiley and Sons, 296 pp.
Kirtman, B., and Coauthors, 2013: Near-term climate change: Projections and predictability. Climate Change 2013: The Physical Science Basis, T. F. Stocker et al., Eds., Cambridge University Press, 953–1028.
Knight, J. R., C. K. Folland, and A. A. Scaife, 2006: Climate impacts of the Atlantic multidecadal oscillation. Geophys. Res. Lett., 33, L17706, https://doi.org/10.1029/2006GL026242.
Li, L., and Coauthors, 2013a: The Flexible Global Ocean-Atmosphere-Land System model, grid-point version 2: FGOALS-g2. Adv. Atmos. Sci., 30, 543–560, https://doi.org/10.1007/s00376-012-2140-6.
Li, L., and Coauthors, 2013b: Evaluation of Grid-Point Atmospheric Model of IAP LASG version 2 (GAMIL2). Adv. Atmos. Sci., 30, 855–867, https://doi.org/10.1007/s00376-013-2157-5.
Li, L., R. W. Schmitt, C. C. Ummenhofer, and K. B. Karnauskas, 2016: North Atlantic salinity as a predictor of Sahel rainfall. Sci. Adv., 2, e1501588, https://doi.org/10.1126/sciadv.1501588.
Liu, H., P. Lin, Y. Yu, and X. Zhang, 2012: The baseline evaluation of LASG/IAP Climate system Ocean Model (LICOM) version 2. Acta Meteor. Sin., 26, 318–329, https://doi.org/10.1007/s13351-012-0305-y.
Lu, J., and T. L. Delworth, 2005: Oceanic forcing of the late 20th century Sahel drought. Geophys. Res. Lett., 32, L22706, https://doi.org/10.1029/2005GL023316.
Mann, M. E., B. A. Steinman, D. J. Brouillette, and S. K. Miller, 2021: Multidecadal climate oscillations during the past millennium driven by volcanic forcing. Science, 371, 1014–1019, https://doi.org/10.1126/science.abc5810.
Mariotti, A., and A. Dell’Aquila, 2012: Decadal climate variability in the Mediterranean region: Roles of large-scale forcings and regional processes. Climate Dyn., 38, 1129–1145, https://doi.org/10.1007/s00382-011-1056-7.
Martin, E. R., and C. Thorncroft, 2014: Sahel rainfall in multimodel CMIP5 decadal hindcasts. Geophys. Res. Lett., 41, 2169–2175, https://doi.org/10.1002/2014GL059338.
Martin, E. R., C. Thorncroft, and B. B. B. Booth, 2014: The multidecadal Atlantic SST—Sahel rainfall teleconnection in CMIP5 simulations. J. Climate, 27, 784–806, https://doi.org/10.1175/JCLI-D-13-00242.1.
Matei, D., H. Pohlmann, J. Jungclaus, W. Müller, H. Haak, and J. Marotzke, 2012: Two tales of initializing decadal climate prediction experiments with the ECHAM5/MPI-OM model. J. Climate, 25, 8502–8523, https://doi.org/10.1175/JCLI-D-11-00633.1.
Mignani, S., and R. Rosa, 1995: The moving block bootstrap to assess the accuracy of statistical estimates in Ising model simulations. Comput. Phys. Commun., 92, 203–213, https://doi.org/10.1016/0010-4655(95)00114-7.
Mohino, E., S. Janicot, and J. Bader, 2011: Sahel rainfall and decadal to multi-decadal sea surface temperature variability. Climate Dyn., 37, 419–440, https://doi.org/10.1007/s00382-010-0867-2.
Mohino, E., N. Keenlyside, and H. Pohlmann, 2016: Decadal prediction of Sahel rainfall: Where does the skill (or lack thereof) come from? Climate Dyn., 47, 3593–3612, https://doi.org/10.1007/s00382-016-3416-9.
Mudelsee, M., 2010: Climate Time Series Analysis: Classical Statistical and Bootstrap Methods. Springer, 474 pp.
Nicholson, S. E., 2013: The West African Sahel: A review of recent studies on the rainfall regime and its interannual variability. ISRN Meteor., 2013, 43521, https://doi.org/10.1155/2013/453521.
Nicolì, D., and Coauthors, 2023: The Euro-Mediterranean Center on Climate Change (CMCC) decadal prediction system. Geosci. Model Dev., 16, 179–197, https://doi.org/10.5194/gmd-16-179-2023.
Oleson, K. W., and Coauthors, 2010: Technical description of version 4.0 of the Community Land Model (CLM). NCAR Tech. Note NCAR/TN-478+STR, 257 pp., https://doi.org/10.5065/D6FB50WZ.
Palmer, T. N., 1986: Influence of the Atlantic, Pacific and Indian Oceans on Sahel rainfall. Nature, 322, 251–253, https://doi.org/10.1038/322251a0.
Park, J.-y., J. Bader, and D. Matei, 2016: Anthropogenic Mediterranean warming essential driver for present and future Sahel rainfall. Nat. Climate Change, 6, 941–945, https://doi.org/10.1038/nclimate3065.
Ramon, J., L. Lledó, V. Torralba, A. Soret, and F. J. Doblas-Reyes, 2019: What global reanalysis best represents near-surface winds? Quart. J. Roy. Meteor. Soc., 145, 3236–3251, https://doi.org/10.1002/qj.3616.
Rayner, N. A., D. E. Parker, E. B. Horton, C. K. Folland, L. V. Alexander, D. P. Rowell, E. C. Kent, and A. Kaplan, 2003: Global analyses of sea surface temperature, sea ice, and night marine air temperature since the late nineteenth century. J. Geophys. Res., 108, 4407, https://doi.org/10.1029/2002JD002670.
Rowell, D. P., 2003: The impact of Mediterranean SSTs on the Sahelian rainfall season. J. Climate, 16, 849–862, https://doi.org/10.1175/1520-0442(2003)016<0849:TIOMSO>2.0.CO;2.
Scaife, A. A., and Coauthors, 2014: Skillful long-range prediction of European and North American winters. Geophys. Res. Lett., 41, 2514–2519, https://doi.org/10.1002/2014GL059637.
Schneider, U., A. Becker, P. Finger, A. Meyer-Christoffer, B. Rudolf, and M. Ziese, 2011: GPCC full data reanalysis version 6.0° at 2.5°: Monthly land-surface precipitation from rain-gauges built on GTS-based and historic data, NCAR, accessed 20 August 2021, https://doi.org/10.5065/D6000072.
Sheen, K. L., D. M. Smith, N. J. Dunstone, R. Eade, D. P. Rowell, and M. Vellinga, 2017: Skilful prediction of Sahel summer rainfall on inter-annual and multi-year timescales. Nat. Commun., 8, 14966, https://doi.org/10.1038/ncomms14966.
Siegert, S., O. Bellprat, M. Ménégoz, D. B. Stephenson, and F. J. Doblas-Reyes, 2017: Detecting improvements in forecast correlation skill: Statistical testing and power analysis. Mon. Wea. Rev., 145, 437–450, https://doi.org/10.1175/MWR-D-16-0037.1.
Slivinski, L. C., and Coauthors, 2019: Towards a more reliable historical reanalysis: Improvements for version 3 of the Twentieth Century Reanalysis system. Quart. J. Roy. Meteor. Soc., 145, 2876–2908, https://doi.org/10.1002/qj.3598.
Smith, D. M., R. Eade, and H. Pohlmann, 2013: A comparison of full-field and anomaly initialization for seasonal to decadal climate prediction. Climate Dyn., 41, 3325–3338, https://doi.org/10.1007/s00382-013-1683-2.
Sutton, R. T., G. D. McCarthy, J. Robson, B. Sinha, A. T. Archibald, and L. J. Gray, 2018: Atlantic multidecadal variability and the U.K. ACSIS program. Bull. Amer. Meteor. Soc., 99, 415–425, https://doi.org/10.1175/BAMS-D-16-0266.1.
Tanaka, H. L., N. Ishizaki, and A. Kitoh, 2004: Trend and interannual variability of Walker, monsoon and Hadley circulations defined by velocity potential in the upper troposphere. Tellus, 56A, 250–269, https://doi.org/10.3402/tellusa.v56i3.14410.
Taylor, K. E., R. J. Stouffer, and G. A. Meehl, 2012: An overview of CMIP5 and the experiment design. Bull. Amer. Meteor. Soc., 93, 485–498, https://doi.org/10.1175/BAMS-D-11-00094.1.
van der Linden, P., and J. F. B. Mitchell, Eds., 2009: ENSEMBLES: Climate change and its impacts: Summary of research and results from the ENSEMBLES project. Met Office Hadley Centre Rep., 164 pp., https://ensembles-eu.metoffice.com/docs/Ensembles_final_report_Nov09.pdf.
van Oldenborgh, G. J., F. J. Doblas-Reyes, B. Wouters, and W. Hazeleger, 2012: Decadal prediction skill in a multi-model ensemble. Climate Dyn., 38, 1263–1280, https://doi.org/10.1007/s00382-012-1313-4.
Vellinga, M., and Coauthors, 2016: Sahel decadal rainfall variability and the role of model horizontal resolution. Geophys. Res. Lett., 43, 326–333, https://doi.org/10.1002/2015GL066690.
Villamayor, J., and E. Mohino, 2015: Robust Sahel drought due to the interdecadal Pacific oscillation in CMIP5 simulations. Geophys. Res. Lett., 42, 1214–1222, https://doi.org/10.1002/2014GL062473.
Volpi, D., and Coauthors, 2021: A novel initialization technique for decadal climate predictions. Front. Climate, 3, 681127, https://doi.org/10.3389/fclim.2021.681127.
Wang, B., and Coauthors, 2010: An economical approach to four-dimensional variational data assimilation. Adv. Atmos. Sci., 27, 715–727, https://doi.org/10.1007/s00376-009-9122-3.
Wang, B., J. Liu, L. Liu, S. Xu, and W. Huang, 2018: An approach to localization for ensemble-based data assimilation. PLOS ONE, 13, e0191088, https://doi.org/10.1371/journal.pone.0191088.
Wang, C., S. Dong, A. T. Evan, G. R. Foltz, and S.-K. Lee, 2012: Multidecadal covariability of North Atlantic sea surface temperature, African dust, Sahel rainfall, and Atlantic hurricanes. J. Climate, 25, 5404–5415, https://doi.org/10.1175/JCLI-D-11-00413.1.
Wilks, D. S., 2019: Frequentist statistical inference. Statistical Methods in the Atmospheric Sciences, 4th ed. D. S. Wilks, Ed., Elsevier, 143–207.
Yang, J., Q. Liu, S.-P. Xie, Z. Liu, and L. Wu, 2007: Impact of the Indian Ocean SST basin mode on the Asian summer monsoon. Geophys. Res. Lett., 34, L02708, https://doi.org/10.1029/2006GL028571.
Yeager, S. G., A. Karspeck, G. Danabasoglu, J. Tribbia, and H. Teng, 2012: A decadal prediction case study: Late twentieth-century North Atlantic Ocean heat content. J. Climate, 25, 5173–5189, https://doi.org/10.1175/JCLI-D-11-00595.1.
Yeager, S. G., and Coauthors, 2018: Predicting near-term changes in the Earth system: A large ensemble of initialized decadal prediction simulations using the Community Earth System Model. Bull. Amer. Meteor. Soc., 99, 1867–1886, https://doi.org/10.1175/BAMS-D-17-0098.1.
Yu, Y., and Coauthors, 2008: Coupled model simulations of climate changes in the 20th century and beyond. Adv. Atmos. Sci., 25, 641–654, https://doi.org/10.1007/s00376-008-0641-0.
Zhang, R., and T. L. Delworth, 2006: Impact of Atlantic multidecadal oscillations on India/Sahel rainfall and Atlantic hurricanes. Geophys. Res. Lett., 33, L17712, https://doi.org/10.1029/2006GL026267.