1. Introduction
The recent increase in the occurrence of extreme climate events, as prolonged droughts, strong heat waves, intense hurricane activity, and the related significant social and economic impacts, has raised scientific and political interest in the near-term climate change, encouraging the development of tools for climate predictions at a decadal time scale, namely out to 10–30 years (Meehl et al. 2009). Climate evolution at the decadal time scale is expected to be influenced by both internal variability and external forcing. The internal variability is mainly related to the state of climate in the recent past, so the weight of the initialization in decadal climate simulations is comparable to that of the boundary conditions (Cox and Stephenson 2007). Several studies have been recently conducted to assess the capability of climate models to predict the climate at a decadal time scale, showing predictive skills for global surface temperature and ocean dynamics in the North Atlantic and North Pacific (Smith et al. 2007; Keenlyside et al. 2008; Pohlmann et al. 2009; Mochizuki et al. 2010; Garcia-Serrano and Doblas-Reyes 2012). Moreover, a set of near-term (10–30 yr) integrations of state-of-the-art coupled atmosphere–ocean global circulation models (AOGCMs) specifically addressed to evaluate decadal climate predictions have been conducted in the framework of the Coupled Model Intercomparison Project phase 5 (CMIP5) (Taylor et al. 2012). The evaluation of the decadal experiments is presently in an early stage; however, the CMIP5 models show improved skill in predicting the Atlantic multidecadal variability (AMV) and the interdecadal Pacific oscillation (IPO) (Kim et al. 2012; Chikamoto et al. 2013).
In the context of decadal climate variability, the study of the monsoonal rainfall in the Sahel is a prominent topic because of its strong environmental and economic impact on the sub-Saharan countries (Benson and Clay 1998) and the difficulty of climate models in producing agreed-upon predictions (Cook and Vizy 2006; Biasutti et al. 2008; Philippon et al. 2010). The main forcing for the decadal and multidecadal variability of the Sahelian rainfall observed during the twentieth century is the global ocean SST, with dominant influence of the tropical SST warming, AMV, and IPO (Bader and Latif 2003; Lu and Delworth 2005; Zhang and Delworth 2006; Ting et al. 2009; Caminade and Terray 2010; Mohino et al. 2011). However, the role of radiative forcings as aerosols and greenhouse gases concentrations is not negligible (Haarsma et al. 2005; Biasutti and Giannini 2006), adding sources of unpredictability for future projections. A recent attempt to predict the Sahelian rainfall at the decadal time scale in a multimodel environment has not clarified whether initialization improves the predictive skill (van Oldenborgh et al. 2012). The aim of this work is to assess the capability of eight state-of-the-art CMIP5 AOGCMs in predicting the monsoonal precipitation in the Sahel on a decadal time scale, and to evaluate the impact of the initialization on the predictive skill.
2. Data and methodology
The analysis focuses on the seasonal average of Sahelian precipitation during the monsoonal season, July to September (JAS). For each model, monthly data from a set of 10-yr-long hindcast experiments initialized every 5 years from 1960 to 2005 (10 experiments) and from a historical experiment run in the twentieth century up to 2005 including all forcings are analyzed, and the predictive skills are compared to evaluate the role of the initialization. The historical integrations are started from multicentury preindustrial control runs and forced by the observed atmospheric composition evolution (reflecting both anthropogenic and natural sources), whereas the decadal prediction experiments are initialized from the observed climate state, including atmosphere composition and ocean and sea ice conditions. The initial dates of the 10 decadal experiments are in the last 4 months of 1960, 1965, … , 2000, 2005, and the outputs extend through the next 10 years, namely the periods 1961–70, 1966–75, … , 2001–10, 2006–15. Ocean and sea ice initial conditions are representative of the observed anomalies or full fields for the start date. Basic information about the models (including full expansions) is reported in Table 1; more details concerning the experimental set up can be found in Taylor et al. (2012).
Models used in the analysis, institution and country, latitude–longitude resolution, and number of members in decadal and historical experiments.
The predictive skill is evaluated by computing the anomaly correlation coefficient (ACC) and the root-mean-square error (RMSE) related to the simulated and observed precipitation seasonal anomalies, in order to assess the capability of the models in reproducing the observed precipitation amount and variability. The subtraction of the climatology is particularly necessary when comparing models to observations because models produce their own climatology, which can be quite different from the observed one. Observed precipitation monthly means are extracted from the University of East Anglia Climate Research Unit (CRU)-TS3.1 database (Mitchell and Jones 2005). The CRU database covers the period 1901–2009, is based on observations from meteorological stations, and extends over the global land surface at 0.5° resolution. The CRU data are widely used to describe the Sahel precipitation multidecadal variability and validate model simulations (Haarsma et al. 2005; Cook and Vizy 2006; Zhang and Delworth 2006; Caminade and Terray 2010; Mohino et al. 2011) and compare well with station precipitation in the Sahel (Fink et al. 2010). The different-resolution datasets are regridded on a T42 Gaussian grid (~2.8° latitude–longitude), which is consistent with the coarsest resolutions of the analyzed models (Table 1). The anomalies are computed by subtracting the climatology at each forecast year (e.g., the forecast year n = 1, … , 10 corresponds to years 1960 + n, 1965 + n, … , 2000 + n, 2005 + n in the 10 studied decades); thus, for each forecast year a subset is composed using the selected years and the climatology and anomalies within the subset are computed. The anomalies are computed considering the simulations' and observations' overlapping periods, 1961–2009 for the decadal hindcasts and 1961–2005 for the historical experiments. Therefore, in the analysis of the decadal hindcasts the forecast years 1–4 include 10 values, the forecast years 5–9 include 9 values, and the forecast year 10 includes 8 values, whereas in the analysis of the historical experiments all the forecast years include 8 values. The above procedure is a standard approach to compute anomalies for decadal predictions. It is based on the World Climate Research Programme recommendations (see ICPO 2011) and it is widely used in the assessment of decadal predictions (Garcia-Serrano and Doblas-Reyes 2012; Goddard et al. 2013; Kim et al. 2012). Garcia-Serrano and Doblas-Reyes (2012) show that this approach systematically yields better ACC scores than the removal of the classic long-term mean climatology. Finally a 4-yr running mean is applied to the observed and simulated anomalies and the verification measures, ACC and RMSE, are applied to the resulting time series in order to assess the evolution of the predictive skill along the forecast time dimension. In the framework of decadal predictions, the 4-yr running mean is a common approach for filtering the data (e.g., van Oldenborgh et al. 2012; Kim et al. 2012; Goddard et al. 2013). Garcia-Serrano and Doblas-Reyes (2012) suggest that a 4-yr running mean filtering is a good compromise between removing some of the unpredictable interannual phenomena (such as El Niño) and still showing the skill evolution along the forecast time. Because of the small size of the samples for each 4-yr forecast period, the significance of the correlation is computed using a Monte Carlo test with 200 permutations.
Possible sources of predictability are investigated in the simulations and observations through a singular value decomposition analysis (SVDA) of the precipitation over West Africa (5°–20°N, 20°W–30°E) and the SST in the 50°S–70°N belt, aiming to detect the main modes of covariance. The SVDA is applied for each model to the standardized variables, concatenating the ensemble members in the time dimension (Philippon et al. 2010). Observed SST monthly means are extracted from the Extended Reconstructed Sea Surface Temperature version 3b (ERSST.v3b) (Smith et al. 2008) at 2° resolution for the period 1854 to the present.
To obtain a more robust evaluation of the initialization weight in the decadal experiments, the model's variability associated with the external forcing is detected in the historical simulations through the signal-to-noise maximizing empirical orthogonal function methodology (Venzke et al. 1999) applied to West Africa precipitation (5°–20°N, 20°W–30°E) and 50°S–70°N SST. The forcing patterns in precipitation and SST are subtracted from the decadal hindcasts and the verification measures and the SVDA are then applied to the resulting decadal-forcing residuals. Residuals associated with the multidecadal variability are also computed for the observed SST by subtracting a global warming (GW) pattern obtained through the regression of the yearly SST fields onto a GW index defined as the time series of globally averaged SST (45°S–60°N) filtered using a low-pass Butterworth filter with 40-yr cutoff frequency (Mohino et al. 2011). The residuals are then smoothed using a 4-yr running mean.
3. Predictive skills
A Sahelian precipitation index is defined averaging the precipitation in the domain (10°–20°N, 15°W–15°E) and its time evolution in the study period is displayed in Fig. 1 for CRU data, historical simulations, and decadal hindcasts. The observed precipitation shows a marked interannual variability modulated by a multidecadal variability, with a drying period from the 1960s to the 1980s and a partial recovery in the recent period. The observed multidecadal variability is partially reproduced by some models in the historical experiment (CanCM4, CNRM-CM5, HadCM3, and MIROC5; all model names are expanded in Table 1), while a slight positive trend characterizes the long-term evolution in the other models. The historical simulations show a general negative bias (except for MIROC5; very large for IPSL-CM5A-LR and MRI-CGCM3) and reduced variability. Similar biases are observed in the decadal hindcasts, along with a drift toward the model climate for some models (BCC-CSM1.1, CanCM4, and CNRM-CM5). The computation of the anomalies at each forecast year removes the effect of the stationary climate drift.
(top to bottom) Sahelian precipitation index for different models: (left) CRU time series (black line), historical simulations (red line), (middle) decadal hindcasts (red and blue lines), and (right) for each decade, the 6–9-yr average (1966–69, 1971–74, … , 2006–09) for CRU data (black circles), decadal hindcasts (blue stars), and historical simulations (red crosses). Each model is represented by the ensemble mean.
Citation: Journal of Climate 26, 19; 10.1175/JCLI-D-12-00635.1
ACC and RMSE are computed between the observed and simulated Sahelian precipitation indices and are displayed in Fig. 2. The RMSE is in the 0.2–0.4 mm day−1 range, with lower values for long lead times (5–10 yr; Figs. 2d–f). The RMSE is generally higher than the standard deviation (STD) of the simulated Sahelian index, except for MIROC5 (Figs. 2d–f), indicating that the bias in the representation of the precipitation anomaly amount (see also Fig. 1) can be sizeable and should be taken into account when assessing decadal predictability (Goddard et al. 2013). Although it highly varies among models, a general improvement in the ACC is observed for long lead times in the decadal hindcasts (Fig. 2a). This feature is also found when using longer time windows for the filtering (five years instead of four), although a general reduction of the significance is observed (not shown). On the other hand, by increasing the sampling for those models with simulations launched every year instead of every five years (CanCM4, HadCM3, MIROC5, and MPI-ESM-LR), the general improvement of the skill with the lead time is conserved showing significant values at long lead times (not shown). This tendency to show improved ACC scores at long lead times could be related to nonstationary drifts that are not correctly removed by the subtraction of the climatology at each lead time (Goddard et al. 2013). The CanCM4 model shows significant correlations for both short (1–6 yr) and long lead times, while CNRM-CM5 and MPI-ESM-LR models are skillful for long lead times. The multimodel ensemble (MME) is skillful with a 6–9-yr lead time. In the historical experiments the ACC shows a wide spread among the models with few significant values (Fig. 2b): the CanCM4 and MIROC5 models are skillful for long lead times, and the MME is skillful with a 6–9-yr lead time. The different performances of the decadal and historical experiments at long lead times are also evident when comparing the time evolution of the observed and simulated Sahelian index averaged over the 6–9-yr interval of each decade. The CanCM4, CNRM-CM5, and MPI-ESM-LR decadal hindcasts follow the observations more closely than the historical simulations, and the MIROC5 model produces very similar results in both experiments, while the other models show discrepancies in reconstructing the observed time evolution in both the decadal and historical simulations (Fig. 1, right panels). When the ACC is computed for the decadal-forcing residuals, the results are similar to those from the decadal hindcasts, with the same skills for the CanCM4, CNRM-CM5, and MPI-ESM-LR models and the MME (Fig. 2c), indicating a contribution of the initialization to the predictive skills of these models. The MIROC5 skill is lost in the residuals and is even lower than in the decadal experiments. In this case the removal of the model's forced component eliminates any predictive skill, suggesting that the initialization does not improve the decadal predictability. To represent a confidence limit for the significance of the predictive skills, the verification measures are tested against persistence, assuming that the observed precipitation at the initialization year persists throughout the decade, namely precipitation in 1960 (1965, 1970, … , 2000, 2005) is assumed to be persistent from 1961 to 1970 (1966–75, 1971–80, … , 2001–09, 2005–09). Figure 2 displays that most of the models show better skill than the persistence prediction skill at long lead times (higher ACC and smaller RMSE) and all the significant ACC are above the persistence threshold.
Sahelian precipitation index: (top) ACC and (bottom) RMSE between the CRU dataset and (left) decadal hindcasts, (middle) historical experiments, and (right) decadal-forcing residuals. (bottom) Dash-dotted lines display the STD of the simulated Sahelian index. Dashed lines display ACC and RMSE assuming persistence for the CRU data. Each model is represented by the ensemble mean. Circles indicate the 95% significant positive correlations.
Citation: Journal of Climate 26, 19; 10.1175/JCLI-D-12-00635.1
The capability of the models in reproducing the observed spatial pattern of the Sahelian precipitation is explored correlating the CRU Sahelian index with the simulated precipitation fields. In Fig. 3 the correlation maps for the decadal hindcasts for each model and the MME are shown. The 2–5, 4–7, and 6–9 forecast years are displayed in order to represent the predictive skill for a short, medium, and long lead time. The CanCM4 model shows coherent significant correlation patterns in the Sahel at 2–5 and 6–9 forecast years; the CNRM-CM5, MIROC5, and MPI-ESM-LR models and the MME show significant correlations in the Sahel for long lead time. The HadCM3 model shows significant values over the ocean and the IPSL-CM5A-LR model produces a coherent but not significant pattern over the eastern Sahel at long lead time. The BCC-CSM1.1 and MRI-CGCM3 models do not show coherent correlation patterns. When the CRU Sahelian index is correlated with the simulated precipitation fields from the historical simulations no clear Sahelian patterns are observed, except for the MIROC5 model and the MME for 6–9-yr lead time (Fig. 4).
ACC at (left to right) forecast years 2–5, 4–7, and 6–9 between the CRU Sahelian index and the precipitation fields from the decadal hindcasts for (top to bottom) the different models. Each model is represented by the ensemble mean. Shadings indicate the 95% significant positive correlations.
Citation: Journal of Climate 26, 19; 10.1175/JCLI-D-12-00635.1
As in Fig. 3, but for historical simulations.
Citation: Journal of Climate 26, 19; 10.1175/JCLI-D-12-00635.1
The correlation maps for the decadal-forcing residuals are displayed in Fig. 5. Significant predictive skill for the Sahelian precipitation is observed for the models CanCM4 (2–5 and 6–9 forecast years), CNRM-CM5 (6–9 forecast years), MPI-ESM-LR (6–9 forecast years), and MRI-CGCM3 (4–7 forecast years). The HadCM3 model shows a no significant southern pattern, while the IPSL-CM5A-LR model shows a coherent but not significant pattern over the eastern Sahel (6–9 forecast years). The MME shows significant but weakly coherent correlation patterns. Interestingly, the MIROC5 model shows a negative correlation over the Sahel for 6–9-yr lead time. These results suggest that in the models producing skillful decadal predictions the initialization plays an important role, except for the MIROC5 model in which the skill appears to be related to the external forcing.
As in Fig. 3, but for decadal-forcing residuals.
Citation: Journal of Climate 26, 19; 10.1175/JCLI-D-12-00635.1
4. Sources of predictability
In Fig. 6 the first mode of the SVDA applied to the models and observations residuals is displayed, aiming to investigate the SST–precipitation covariance and to identify the possible sources of prediction skills once the model's response to the external forcing is subtracted. The additional prediction skills coming from the initialization are in general better at long lead times (Figs. 2c and 5), so the 6–9 forecast years are presented. In the observations wet anomalies in the Sahel are related to positive SST anomalies in the Northern Hemisphere and negative anomalies in the tropics and the Southern Ocean, typical of the AMV and IPO patterns (Mohino et al. 2011). The SST expansion coefficients (ECs) time series is significantly (95%) correlated with the AMV and IPO indices computed by Mohino et al. (2011) (0.69 and 0.41 respectively). Similar spatial patterns and EC time series that fit well the observed multidecadal variability result from the CanCM4, CNRM-CM5, HadCM3, and MPI-ESM-LR models (see Table 2), with some differences in reproducing the regional SST patterns. CanCM4 and HadCM3 well describe the North Pacific variability, CNRM-CM5 shows good results in the North Atlantic, and MPI-ESM-LR produces a covariance pattern very close to the observed one. HadCM3 shows a precipitation pattern too far south and the ECs time series is modulated by a higher-frequency component. BCC-CSM1.1, IPSL-CM5A-LR, and MRI-CGCM3 show EC time series dominated by long-term trends and SST patterns not fitting the observed ones. IPSL-CM5A-LR well reproduces the Atlantic and North Pacific variability but fails in describing the tropical and Southern Ocean observed behavior. MIROC5 shows a SST pattern very similar to the observed one but failing in the representation of the North Atlantic variability, and an anticorrelation between the simulated and observed ECs (Table 2). This result suggests that the initialization introduces the correct signature of the multidecadal variability in the Sahel precipitation but with reversed anomalies [i.e., wet (dry) anomalies are simulated when dry (wet) anomalies are observed]. The present analysis does not provide a clear explanation for this “exotic” behavior and this issue should be more deeply investigated.
First SST–precipitation SVD mode in decadal-forcing residuals for 6–9 forecast years for (top to bottom) the different models and observations: (left) SST homogeneous patterns, (middle) precipitation heterogeneous patterns, and (right) associated ECs time series, with thick lines for the models and thin lines for the observations (SST in blue, precipitation in red). In the last row, the SVD is applied to observations from 1910 to 2010. Shadings indicate the 95% significant correlations.
Citation: Journal of Climate 26, 19; 10.1175/JCLI-D-12-00635.1
Correlation coefficients between models and observations ECs time series for SST/precipitation. Bold (italic) values are 95% (90%) significant.
The covariance analysis indicates that some of the models are able to reproduce the observed SST–precipitation relationship in the residuals at a multidecadal time scale (CanCM4, CNRM-CM5, and MPI-ESM-LR) and the same models show additional skill in the decadal prediction of the JAS Sahelian rainfall coming from the initialization (Fig. 5). In Table 2 the correlation coefficients between EC time series from SVDA applied to models and observations residuals at forecast years 2–5, 4–7, and 6–9 are presented. Note that SST and precipitation ECs show significant correlation values for the models and forecast years with additional predictive skills (Fig. 5), suggesting that the ability of the models in reproducing the SST–precipitation multidecadal variability is crucial for skillful decadal predictions.
5. Discussion and conclusions
The performances of the CMIP5 models in predicting the JAS Sahelian precipitation are summarized on the basis of the covariance analysis of the decadal-forcing residuals in Fig. 6. Three models (CanCM4, CNRM-CM5, and MPI-ESM-LR) show the SST–precipitation patterns and associated time variability very close to the observed ones, although some regional differences are evident, and prediction skills are detected in the Sahel mainly at long lead times (6–9 yr; Figs. 2, 3, and 5). It indicates that the initialization remarkably affects the prediction skills, specifically improving the description of the multidecadal variability associated with the AMV and IPO. The analysis of decadal experiments launched every year (not shown), showing robust skills at long lead times, confirms the crucial contribution of the initialization. The HadCM3 covariance pattern is characterized by a Guinean precipitation mode, so significant skill is detected along the Guinean coast (Figs. 3 and 5). The IPSL-CM5A-LR model is weakly skillful over the Sahel (Figs. 3 and 5), which can be related to the partial description of the SST covariance pattern not showing significant anomalies in the Southern Ocean. The MIROC5 model is skillful over the Sahel in both the decadal and historical experiments, and no skill is observed in the residuals (Figs. 3 and 5), suggesting a negative impact on the predictability coming from the initialization. The MRI-CGCM3 model shows a covariance pattern far from the observed one with no skills in both decadal and historical simulations (Figs. 3 and 4). On the other hand, predictive skill is detected in the residuals (Fig. 5), indicating that the removal of the model's response to the external forcing improves the forecast, although it is not possible to identify the sources of predictability. It could be related to noisy forcing fields (not shown). In this case the improved skill is related to the initialization but, in contrast to the rest of the models, it does not come from the inclusion of a multidecadal SST variability. The BCC-CSM1.1 model does not show significant predictive performances.
In conclusion, it is possible to predict the JAS Sahelian precipitation at a decadal time scale if the multidecadal variability associated with the AMV and IPO is correctly represented, and the decadal hindcast approach is valuable for this purpose in three CMIP5 models (CanCM4, CNRM-CM5, and MPI-ESM-LR). However, it is important to highlight that the initialization procedure not only allows us to describe the SST internal variability, but could also account for part of the SST variability that is externally induced, through the correction of the model's near-term response to the external forcing (Goddard et al. 2013). As found by Shin and Sardeshmukh (2011), state-of-the-art AOGCMs run using prescribed observed radiative forcing do not capture well recent climate trends at a regional scale; whereas uncoupled atmospheric models forced by the observed SST changes are more skillful in this regard. This behavior is due to the poor representation of tropical SST in the coupled models, which can lead to large impacts on the simulation of both local and remote precipitation trends. Thus by means of the proper initialization, it is possible to correct misrepresentations of the externally forced variability in SSTs.
Acknowledgments
We acknowledge the World Climate Research Programme's Working Group on Coupled Modelling, which is responsible for CMIP, and we thank the climate modeling groups (listed in Table 1 of this paper) for producing and making available their model output. For CMIP the U.S. Department of Energy's Program for Climate Model Diagnosis and Intercomparison provides coordinating support and led development of software infrastructure in partnership with the Global Organization for Earth System Science Portals. This work was supported by the Spanish projects MICINN CGL2011-13564-E and CGL2009-10285.
We want to thank Juliette Mignot for the data provided and Luis Dinis for useful discussion.
REFERENCES
Bader, J., and M. Latif, 2003: The impact of decadal scale Indian Ocean SST anomalies on Sahelian rainfall and the North Atlantic Oscillation. Geophys. Res. Lett., 30, 2169, doi:10.1029/2003GL018426.
Benson, C., and E. J. Clay, 1998: The impact of drought on sub-Saharan African economies. World Bank Tech. Paper 401, 83 pp.
Biasutti, M., and A. Giannini, 2006: Robust Sahel drying in response to late 20th century forcings. Geophys. Res. Lett., 33, L11706, doi:10.1029/2006GL026067.
Biasutti, M., I. M. Held, A. H. Sobel, and A. Giannini, 2008: SST forcings and Sahel rainfall variability in simulations of the twentieth and twenty-first centuries. J. Climate, 21, 3471–3486.
Caminade, C., and L. Terray, 2010: Twentieth century Sahel rainfall variability as simulated by the ARPEGE AGCM, and future changes. Climate Dyn., 35, 75–94, doi:10.1007/s00382-009-0545-4.
Chikamoto, Y., and Coauthors, 2013: An overview of decadal climate predictability in a multi-model ensemble by climate model MIROC. Climate Dyn., 40, 1201–1222, doi:10.1007/s00382-012-1351-y.
Cook, K. H., and E. K. Vizy, 2006: Coupled model simulations of the West African monsoon system: Twentieth- and twenty-first-century simulations. J. Climate, 19, 3681–3703.
Cox, P., and D. Stephenson, 2007: A changing climate for prediction. Science, 317, 207–208.
Fink, A. H., J. M. Scharage, and S. Kotthaus, 2010: On the potential causes of the nonstationary correlations between West African precipitation and Atlantic hurricane activity. J. Climate, 23, 5437–5456.
Garcia-Serrano, J., and F. J. Doblas-Reyes, 2012: On the assessment of near-surface global temperature and North Atlantic multi-decadal variability in the ENSEMBLES decadal hindcast. Climate Dyn., 39, 2025–2040, doi:10.1007/s00382-012-1413-1.
Goddard, L., and Coauthors, 2013: A verification framework for interannual-to-decadal prediction experiments. Climate Dyn., 40, 245–272, doi:10.1007/s00382-012-1481-2.
Haarsma, R. J., F. M. Selten, S. L. Weber, and M. Kliphuis, 2005: Sahel rainfall variability and response to greenhouse warming. Geophys. Res. Lett., 32, L17702, doi:10.1029/2005GL023232.
ICPO, 2011: Data and bias correction for decadal climate predictions. International CLIVAR Project Office Publication Series 150, 5 pp. [Available online at http://www.wcrp-climate.org/decadal/references/DCPP_Bias_Correction.pdf.]
Keenlyside, N. S., M. Latif, J. Jungclaus, L. Kornblueh, and E. Roeckner, 2008: Advancing decadal-scale climate prediction in the North Atlantic sector. Nature, 453, 84–88.
Kim, H.-M., P. J. Webster, and J. A. Curry, 2012: Evaluation of short-term climate change prediction in multi-model CMIP5 decadal hindcasts. Geophys. Res. Lett., 39, L10701, doi:10.1029/2012GL051644.
Lu, J., and T. L. Delworth, 2005: Oceanic forcing of the late 20th century Sahel drought. Geophys. Res. Lett., 32, L22706, doi:10.1029/2005GL023316.
Meehl, G. A., and Coauthors, 2009: Decadal prediction: Can it be skillful? Bull. Amer. Meteor. Soc., 90, 1467–1485.
Mitchell, T. D., and P. D. Jones, 2005: An improved method of constructing a database of monthly climate observations and associated high-resolution grids. Int. J. Climatol., 25, 693–712.
Mochizuki, T., and Coauthors, 2010: Pacific decadal oscillation hindcasts relevant to near-term climate prediction. Proc. Natl. Acad. Sci. USA, 107, 1833–1837.
Mohino, E., S. Janicot, and J. Bader, 2011: Sahel rainfall and decadal to multi-decadal sea surface temperature variability. Climate Dyn., 37, 419–440.
Philippon, N., F. J. Doblas-Reyes, and P. M. Ruti, 2010: Skill, reproducibility and potential predictability of the West African monsoon in coupled GCMs. Climate Dyn., 35, 53–74.
Pohlmann, H., J. H. Jungclaus, A. Köhl, D. Stammer, and J. Marotzke, 2009: Initializing decadal climate predictions with the GECCO oceanic synthesis: Effect on the North Atlantic. J. Climate, 22, 3926–3938.
Shin, S. I., and P. D. Sardeshmukh, 2011: Critical influence of the pattern of tropical ocean warming on remote climate trends. Climate Dyn., 36, 1577–1591, doi:10.1007/s00382-009-0732-3.
Smith, D. M., S. Cusack, A. W. Colman, C. K. Folland, G. R. Harris, and J. M. Murphy, 2007: Improved surface temperature prediction for the coming decade from a global climate model. Science, 317, 796–799.
Smith, T. M., R. W. Reynolds, T. C. Peterson, and J. Lawrimore, 2008: Improvements to NOAA's historical merged land–ocean surface temperature analysis (1880–2006). J. Climate, 21, 2283–2296.
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.
Ting, M., Y. Kushnir, R. Seager, and C. Li, 2009: Forced and internal twentieth-century SST trends in the North Atlantic. J. Climate, 22, 1469–1481.
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, doi:10.1007/s00382-012-1313-4.
Venzke, S., M. R. Allen, R. T. Sutton, and D. P. Rowell, 1999: The atmospheric response over the North Atlantic to decadal changes in sea surface temperature. J. Climate, 12, 2562–2584.
Zhang, R., and T. L. Delworth, 2006: Impact of Atlantic multidecadal oscillations on India/Sahel rainfall and Atlantic hurricanes. Geophys. Res. Lett., 33, L17712, doi:10.1029/2006GL026267.
