Volume 43, Issue 1 p. 99-122
RESEARCH ARTICLE
Open Access

Deficient precipitation sensitivity to Sahel land surface forcings among CMIP5 models

Fuyao Wang

Corresponding Author

Fuyao Wang

Nelson Institute Center for Climatic Research, University of Wisconsin-Madison, 1225 W Dayton St, Madison, Wisconsin, USA

Correspondence

Fuyao Wang, Nelson Institute Center for Climatic Research, University of Wisconsin-Madison, Madison, WI 53706.

Email: [email protected]

Search for more papers by this author
Michael Notaro

Michael Notaro

Nelson Institute Center for Climatic Research, University of Wisconsin-Madison, 1225 W Dayton St, Madison, Wisconsin, USA

Search for more papers by this author
Yan Yu

Yan Yu

Department of Atmospheric and Oceanic Sciences, School of Physics, Peking University, Beijing, China

Search for more papers by this author
Jiafu Mao

Jiafu Mao

Environmental Sciences Division and Climate Change Science Institute, Oak Ridge National Laboratory, Oak Ridge, Tennessee, USA

Search for more papers by this author
First published: 25 May 2022
Citations: 1

Funding information: Department of Energy, Grant/Award Numbers: DE-SC0012534, DE-AC05-00OR22725; Office of Science; Biological and Environmental Research

Abstract

The overall performance of the simulated seasonal precipitation response to local terrestrial forcings, namely vegetation abundance and soil moisture, in the Sahel among the Coupled Model Intercomparison Project Phase Five (CMIP5) Earth System Models (ESMs) is systematically investigated and compared with its observational counterpart using a multivariate statistical method. The observed seasonal precipitation response is evaluated against a large ensemble of observational, reanalysis, and satellite data sets to provide quantification of uncertainties. The behaviour of models with and without a Dynamic Global Vegetation Model (DGVM) component is also explored, along with the mechanisms responsible for terrestrial feedback on rainfall. In general, the CMIP5 models can reasonably capture the seasonal evolution of Sahel precipitation and soil moisture, albeit with wet biases during the pre-monsoon period and dry biases during the peak monsoon period. The non-DGVM ESMs simulate comparable leaf area indices (LAIs) with observations, while DGVM-enabled ESMs simulate too much year-round LAI. The variance of precipitation that is attributed to oceanic forcings in CMIP5 is comparable with observations; however, the variance of precipitation that is attributed to terrestrial forcings is smaller in CMIP5 models than observed, especially for non-DGVM ESMs. CMIP5 models, especially those without DGVMs, undervalue precipitation's observed response strength to soil moisture anomalies. In both observations and CMIP5 models, none of the atmospheric variables show significant responses to direct vegetation forcing, except for the response in transpiration. Although vegetation has minimal direct effect on the atmospheric state, it can affect the atmosphere by modifying soil moisture and transpiration rate indirectly, which helps explain the more realistic simulation of rainfall in DGVM-enabled ESMs than non-DGVM ESMs. Coupling of an ESM to a DGVM is critical in generating reasonable land–atmosphere feedback and examining future ecological and climatic changes over the Sahel.

1 INTRODUCTION

The Sahel, one of the global hotspots of notable land–atmosphere interactions (Koster et al., 2004; Zeng et al., 2010), is characterized by pronounced hydrological interannual to decadal variability (Brooks, 2004). The Sahel's shift from heavy rainfall during the 1950s through the mid-1960s to intense drought during the 1970s and 1980s represents one of the greatest observed climatic shifts in the world (Giannini et al., 2008a; Giannini et al., 2008b). Although climate models can robustly capture this prolonged historical drought (Biasutti et al., 2008; Biasutti, 2013, Zebaze et al. 2019), they vastly disagree in terms of the sign and magnitude of future changes in Sahel rainfall associated with anthropogenic climatic change, with significant disagreement among the Coupled Model Intercomparison Project Phase Three (CMIP3), Phase Five (CMIP5), and Phase Six (CMIP6) models (Biasutti, 2013; Roehrig et al., 2013; Maslin, 2013; Monerie et al., 2017; Monerie et al., 2020). By running a large ensemble of regional climate models forced by output from global climate models (GCMs), Dosio et al. (2020) found two contrasting future changes in Sahel precipitation, with the strength of simulated land-atmosphere coupling largely explaining the discrepancies. Given that hydrological extremes, including droughts and pluvials, are a serious threat to the poverty-stricken Sahel, a deeper understanding of the mechanisms responsible for these hydrological extremes can help improve the forecast skill and facilitate societal planning.

Previous studies suggest that the pronounced hydrological variability over the Sahel is largely attributed to oceanic forcings (Folland et al., 1986; Rowell et al., 1995; Giannini et al., 2003) and amplified by terrestrial forcings (Charney, 1975; Charney et al., 1977; Shukla and Mintz, 1982; Xue, 1997; Zeng et al., 1999). It is challenging to isolate and quantify the contribution of terrestrial forcings to atmospheric variability due to strong atmospheric internal variability and the presence of concurrent influences from the ocean. Therefore, the scientific community's understanding of terrestrial feedback has largely been based on sensitivity experiments using various climate models (Charney, 1975; Koster et al., 2006 and references there). These studies, however, are model dependent, because each model has its own dynamical core, numerical schemes, parameterizations, and spatial resolution, with differing experiment designs (Liu et al., 2006). Taylor et al. (2012, 2013) evaluated the simulated feedback of soil moisture on rainfall in GCMs against observations at the weather time scale and found an observed negative soil moisture–precipitation feedback in direct contrast to the positive feedback simulated by most GCMs. However, few studies have endeavoured to evaluate simulated Sahel terrestrial feedback against observations at the monthly time scale to assess the models’ credibility (Koster et al., 2006; Notaro and Liu, 2008; Wang et al., 2013, 2014; Yu et al., 2017). The benchmarks for observed terrestrial feedback for the Sahel were established in our previous studies based on the application of a carefully validated multivariate statistical technique (Yu et al., 2017; Notaro et al., 2019). In this paper, the simulated biophysical feedback in CMIP5 Earth System Models (ESMs) are assessed in comparison with these observational benchmarks.

As the complexities of climate models continue to advance, the CMIP5 ESMs now largely contain interactive vegetation phenology, and in some cases, dynamic plant functional types (PFTs) as simulated by Dynamic Global Vegetation Models (DGVMs). The main difference between ESMs that are either coupled to a DGVM or not is that the PFTs simulated by a DGVM can change over time in response to climate variability and change, while in non-DGVM models, the fractional coverages of PFTs are fixed in time, generally to match observed biome distributions. Each individual DGVM addresses its own set of physical processes, such as photosynthesis, stomatal conductance, and resource competition. In this study, we focus on the simulated versus observed influence of monthly variability in leaf area index and soil moisture on the atmosphere. The source of this terrestrial variability and the differences among DGVMs are beyond the scope of this paper.

The land surface interacts with the atmosphere via a series of complex mechanisms. Vegetation influences the climatic conditions directly through altered surface albedo, roughness, and plant transpiration, which modifies energy, momentum and moisture exchanges with the atmosphere, and also indirectly through biogeochemical processes which impact atmospheric carbon dioxide levels (Pielke Sr. et al., 1998). According to the energy feedback mechanism, a reduction in Sahel vegetation greenness increases the surface albedo, since the albedo of vegetation is lower than that of desert, which cools the low-level atmosphere and results in increased atmospheric stability, low-level subsidence, and reduced precipitation. This mechanism is the classic positive vegetation-precipitation feedback hypothesis suggested by Charney (1975). According to the moisture recycling feedback mechanism, vegetation links deep soil moisture to the overlying atmosphere via transpiration (Henderson-Sellers et al., 1995; Pollard and Thompson, 1995; Betts and Ball, 1997). If soil moisture is sufficient, an increased abundance of vegetation tends to enhance evapotranspiration (ET) and atmospheric precipitable water, leading to more precipitation. At the same time, according to the stability mechanism, the enhanced ET corresponds to greater latent heat flux to the atmosphere, which cools the lower atmosphere and leads to a more stable lower atmosphere (Yu et al., 2017, 2018). If soil moisture is limited, then the increased vegetation abundance depletes the soil moisture, which decreases the subsequent ET and results in reduced subsequent precipitation. According to the momentum feedback mechanism, a positive anomaly in vegetation cover is associated with greater land surface roughness, which favours enhanced precipitation by reducing wind speed and increasing frictional convergence and ascending motion.

Soil moisture anomalies can likewise support land-atmosphere coupling and persistence in the climate system (Seneviratne et al., 2010). Soil moisture can regulate the surface albedo, transpiration rates, and partitioning of surface-incident solar insolation into sensible and latent heat fluxes, thereby impacting air temperature, stability and moist static energy of the atmospheric boundary layer, and rainfall (Eltahir, 1998; Seneviratne et al., 2010). Likewise, heterogeneous patterns of soil moisture can support the formation of mesoscale circulations with enhanced convergence across zones of strong sensible heat gradient, leading to greater rainfall (Pielke, 2001; Taylor et al., 2007, 2011, 2012). According to the Global Land-Atmosphere Coupling Experiment (GLACE), one of the key hotspots of soil moisture-rainfall coupling in boreal summer among GCMs stretches across the Sahel, West African Monsoon (WAM) region, and equatorial Africa, although the coupling strength differs dramatically across models (Koster et al., 2004). GLACE has demonstrated the notable soil moisture-precipitation coupling strength across such regional zones of hydroclimatic transition between dry and wet climates (Koster et al., 2004). Taylor et al. (2012) confirmed this enhanced coupling across semi-arid regions using global satellite data, although they surprisingly found evidence that drier soils favoured afternoon rainfall, indicative of an observed negative soil moisture-precipitation feedback operating at weather time scales in clear contrast to the positive feedback simulated by most coarse GCMs. Hohenegger et al. (2009), using the Consortium for Small-Scale Modelling Model in Climate Mode (CCLM), found that the sign of soil moisture–precipitation feedback varied by model resolution. The coarser version with 25-km grid spacing simulated strong positive feedback, while the finer version that applied a grid spacing of 2.2 km predominantly simulated negative feedback. Taylor et al. (2013) expressed concern about the capability of coarse resolution GCMs to capture the sensitivity of daytime convection to soil moisture and the resulting soil moisture–precipitation feedback loop, even noting that the sign of the feedback reversed in regional climate models if the convective parameterization was turned on or off. Furthermore, Müller et al. (2021) concluded that model resolution strongly impacts the spatial distribution and coupling strength of land-atmosphere interactions among GCMs, especially for the Sahel where changes in resolution impact moisture fluxes, the African easterly jet, wind shear, atmospheric stability, and resulting rainfall.

Most previous studies of land surface feedback focused either on vegetation feedback or soil moisture feedback and not on their individual or combined contributions (Liu et al., 2010). In this study, the isolated contributions of individual terrestrial forcings on the Sahel's regional climate are quantified using a validated multivariate statistical method, which is detailed in section 2. The present paper focuses on simulated land surface feedback on the Sahel's regional climate in historical (1960–2000) simulations from the CMIP5 ESMs. The overall performance of the models, either with or without coupling to a DGVM, is equally weighted, and the multi-model mean (MMM) is used to measure the general reliability of the models. The data and method are introduced in section 2. The observed and simulated seasonal cycles of precipitation, soil moisture, and LAI are compared in section 3. The predictability indicated by explained variance (EV) in precipitation is discussed in section 4, and the rainfall response to soil moisture and LAI forcings is investigated in section 5. The discussion and conclusions are presented in section 6.

2 DATA AND METHOD

2.1 Data

The current study focuses on 23 CMIP5 models, representing those models containing interactive vegetation phenology (Table 1); models that apply a fixed seasonal cycle of LAI without interannual variability are excluded. Among the 23 models, the land components lack a DGVM in the first 8 models in Table 1 (and thus apply temporally invariant PFT fractional coverages) and include a DGVM in the last 15 models (and thus apply temporally variant PFT fractional coverages). The number of ensemble members per model varies from one to seven. The MMM is the average of the ensemble mean of all models. Monthly output from historical CMIP5 experiments during the period of 1960–2000 is analysed in comparison with observations during 1982–2012. All of the model output is regridded to a 1° × 1° grid for direct comparison, and the seasonal cycle and third-order polynomial trend are eliminated to exclude seasonality and long-term trends and instead focus on interannual anomalies.

TABLE 1. The CMIP5 models and number of ensemble members analysed in the present study, along with whether or not they are coupled to a dynamic global vegetation model (DGVM)
Model index Model name Full name Number of ensemble members Coupled to DGVM?
1 CCSM4 National Center for Atmospheric Research (NCAR) community climate system model, version 4 6 N
2 CESM1-BGC NCAR Community earth system model (CESM), version 1 with biogeochemistry (BGC) 1 N
3 CESM1-CAM5 NCAR CESM, version 1 with community atmosphere model, version 5 4 N
4 CESM1-WACCM NCAR CESM, version 1 whole atmosphere community climate model 7 N
5 NorESM1-ME Norwegian climate center (NCC) earth system model, version 1, medium resolution with prognostic BGC cycling 1 N
6 NorESM1-M NCC earth system model, version 1, medium resolution 3 N
7 HadGEM2-CC Hadley Centre global environmental model, version 2 carbon cycle 3 N
8 INMCM4 Institute of Numerical Mathematics Coupled Model, version 4 1 N
9 CanESM2 Canadian Centre for Climate Modelling and Analysis Earth System Model, version 2 5 Y
10 GFDL-CM3 National Oceanic and Atmospheric Administration (NOAA) geophysical fluid dynamics laboratory (GFDL) climate model, version 3 5 Y
11 GFDL-ESM2G NOAA GFDL earth system model with Generalized Ocean layer dynamics component 1 Y
12 GFDL-ESM2M NOAA GFDL earth system model with Modular Ocean model component 1 Y
13 HadGEM2-ES Hadley Centre global environmental model, version 2 earth system 4 Y
14 IPSL-CM5A-LR L'Institut Pierre-Simon Laplace (IPSL) coupled model, version 5a, low resolution 6 Y
15 IPSL-CM5A-MR IPSL coupled model, version 5a, medium resolution 3 Y
16 IPSL-CM5B-LR IPSL coupled model, version 5b, low resolution 1 Y
17 MIROC-ESM-CHEM Model for interdisciplinary research on climate (MIROC) earth system model with atmospheric chemistry 1 Y
18 MIROC-ESM MIROC earth system model 3 Y
19 MPI-ESM-LR Max Planck institute (MPI) for meteorology earth system model, low resolution 3 Y
20 MPI-ESM-MR MPI earth system model, medium resolution 3 Y
21 MRI-ESM1 Meteorological research institute earth system model, version 1 1 Y
22 BCC-CSM1-1-M BCC climate system model, version 1.1 moderate resolution 3 Y
23 BCC-CSM1-1 Beijing climate center (BCC) climate system model, version 1.1 3 Y

Multiple observational, reanalysis, and satellite data sets (Table 2) are used to provide uncertainty estimates among different data set combinations, considering the relatively poor observational climate data network across the Sahel. Five vegetation index data sets, including LAI from Global Land Surface Satellite (GLASS), Long-Term Global Mapping Version 2 (GLOBMAP2), Global Inventory Monitoring and Modelling System (GIMMS) LAI3g, Integrated Global Monitoring for Environment and Security Project on Land Cover and Vegetation version 2 (GEOLAND2) and NDVI from Normalized Difference Vegetation Index (GIMMS NDVI3g) are applied to calculate vegetation forcings. Three ET data sets, including Mao's Merged Diagnostic ET, Penman–Monteith–Leuning (PML) estimated ET, and Global Land Evaporation Amsterdam Model (GLEAM) Global ET, are used to calculate the response field. Five precipitation data sets, such as UDEL Terrestrial Precipitation v4.01, CRU high resolution gridded data TS4.01, Global Precipitation Climatology Project (GPCP) v2.3, Global Precipitation Climatology Centre (GPCC) v7, and CPC Merged Analysis of Precipitation (CMAP) are used to determine precipitation response fields. By applying multiple data sets the uncertainty in SGEFA fields can be addressed.

TABLE 2. Observational, reanalysis, and remote sensing data sets applied in this study
Variable Data set Period Spatial resolution References
Forcing fields
Sea-surface temperature (SST) Hadley Centre sea ice and sea surface temperature data set (HadISST) 1870–2016 1° × 1° Rayner et al. (2003)
Leaf area index (LAI) Global land surface satellite (GLASS) 1982–2012 0.05° × 0.05° Xiao et al. (2013)
Long-term global mapping version (v) 2 (GLOBMAP2) 1982–2015 0.0732° × 0.0732° Liu et al. (2012)
Global inventory monitoring and modelling system (GIMMS) LAI3g 1982–2015 0.0833° × 0.0833° Pinzon and Tucker (2014)
Integrated global monitoring for environment and security project on land cover and vegetation v2 (GEOLAND2) 1982–2012 0.0082° × 0.0082° Baret et al. (2013)
Normalized difference vegetation index (NDVI) GIMMS NDVI3g 1982–2015 0.0833° × 0.0833° Pinzon and Tucker (2014)
Root-depth soil moisture Global land evaporation Amsterdam model (GLEAM) v3.2a 1980–2016 0.25° × 0.25° Martens et al. (2017)
Response fields
Evapotranspiration (ET) Mao's merged diagnostic ET 1982–2010 0.5° × 0.5° Mao et al. (2015)
Penman–Monteith–Leuning (PML) estimated ET 1981–2012 0.5° × 0.5° Zhang et al. (2016)
GLEAM global ET v3.2a 1980–2016 0.25° × 0.25° Mu et al. (2007)
ET partitioning: Transpiration, evaporation from vegetation, and evaporation from soil PML estimated ET partitioning 1981–2012 0.5° × 0.5° Zhang et al. (2016)
GLEAM global ET v3.2a 1980–2016 0.25° × 0.25° Martens et al. (2017)
Precipitation University of Delaware (UDEL) terrestrial precipitation v4.01 1901–2014 0.5° × 0.5° Matsuura and Willmott (2012)
University of East Anglia Climatic Research Unit (CRU) high resolution gridded data TS4.01 1901–2016 0.5° × 0.5° Harris et al. (2014)
Global precipitation climatology project (GPCP) v2.3 1979–2017 2.5° × 2.5° Huffman (1997)
Global precipitation climatology Centre (GPCC) v7 1901–2013 0.5° × 0.5° Schneider et al. (2014)
Climate prediction center (CPC) merged analysis of precipitation (CMAP) 1979–2017 2.5° × 2.5° Xie and Arkin (1997)
2-m air temperature CRU TS4.01 1901–2016 0.5° × 0.5° Harris et al. (2014)
UDEL v4.01 1901–2014 0.5° × 0.5° Matsuura and Willmott (2012)
Precipitable water, 2-m specific humidity, 10-m wind, sea-level pressure, vertical motion Japanese 55-year reanalysis (JRA55) 1979–2013 1.25° × 1.25° Kobayashi (2015)
National Aeronautics and Space Administration (NASA) modern-era retrospective analysis for research and applications (MERRA) v2 1980–2017 0.625° × 0.5° Gelaro et al. (2017)
National Centers for environmental prediction (NCEP)-climate forecast system analysis (CFSR) 1979–2010 0.5° × 0.5° Saha et al. (2010)
European Centre for Medium-Range Weather Forecast (ECMWF) interim reanalysis (ERA-interim) 1980–2017 0.7° × 0.7° Dee et al. (2011)
Specific humidity, zonal wind, meridional wind at 925 hPa NCEP-NCAR reanalysis I 1948–Present 2.5 × 2.5 Kalnay (1996)
Outgoing longwave radiation (OLR) National Oceanic and Atmospheric Administration (NOAA) OLR-daily climate record (NOAA-CDR) 1979–2012 1° × 1° National Research Council (2004)
NOAA interpolated OLR 1974–2017 2.5° × 2.5° Liebmann and Smith (1996)
Cloud area fraction The advanced very high resolution radiometer (AVHRR) pathfinder atmospheres extended (PATMOS-x) 1979–Present 0.1° × 0.1° Andrew et al. (2014)

2.2 Stepwise generalized equilibrium feedback assessment

Built on the stochastic climate theory (Hasselmann, 1976; Frankignoul and Hasselmann, 1977), Frankignoul et al. (1998) proposed a univariate statistical assessment to investigate the impact of oceanic variability on the atmospheric state. By extending the univariate statistical assessment to a multi-variate statistical assessment, Generalized Equilibrium Feedback Assessment (GEFA) was developed by Liu et al. (2008), which can be used to study the impact of slowly-evolving forcings, for example, sea-surface temperature (SST), vegetation index, or soil moisture on the atmosphere (Wen et al. 2010). The validation of GEFA has been investigated using additional statistical methods, such as fluctuation-dissipation theorem and linear inverse modelling, by Liu et al. (2012a, 2012b). Wang et al. (2013, 2014) further demonstrated the reliability of GEFA using fully coupled climate models in comparison with dynamical ensemble experiments. SGEFA represents an advanced version of the traditional GEFA method with more stable performance and smaller sampling errors, which is especially valuable when the available time series are limited in duration (Wang et al., 2017; Yu et al., 2018). The reliability of SGEFA in separating and quantifying the terrestrial and oceanic feedback on the climate of North African has been demonstrated through a comparison of SGEFA-based statistical assessments against dynamical assessments, using experiments with modified oceanic or terrestrial forcings, in the Community Earth System Model (CESM) (Wang et al., 2017; Yu et al., 2018). Prior applications of SGEFA include its validation regarding oceanic and terrestrial feedback by Wang et al. (2017) and Yu et al. (2018); assessment of observed land surface feedback over sub-Saharan Africa by Yu et al. (2017) and Notaro et al. (2019), the La Plata River Basin by Chug and Dominguez (2019), the Australian monsoon system by Yu and Notaro (2020), and the North American monsoon system by Wang and Quiring (2021); analysis of the observed drivers of African fire activity by Yu et al. (2020); and projected changes in the oceanic and terrestrial regulators of North African climate variability by Notaro et al. (2020). The SGEFA method is briefly described here, with further details provided by Wang et al. (2017), including the difference between SGEFA and traditional GEFA and the two approaches' advantages and disadvantages.

According to the stochastic climate theory, the coupled climate system can be represented by a quickly-varying ‘weather’ component (e.g., atmosphere) and a gradually responding ‘climate’ component (e.g., SST, vegetation greenness index, and soil moisture). At time t , the atmospheric variable A t , can be broken down into two terms: B × O t , which is the feedback response of the atmosphere to the slow-process forcings, and N t , which is the atmospheric internal variability, such that:
A t = B × O t + N t . (1)
The feedback response matrix, B , consists of the responses of atmospheric variable A to each individual forcing in O. Through multiplying by O t τ T and applying covariance to both sides of Equation (1), the equation becomes:
A t O t τ T = B × O t O t τ T + N t O t τ T . (2)
where τ is a time lag that is longer than the atmospheric persistence time (here, the lag is 1 month) and superscript ‘T’ represents a transpose, and x , y represents the covariance between variables x and y. Since the atmospheric internal variability at time t, N t , is independent of the variability of slow-evolving variables at an earlier time, O t τ , the following approximation can be made:
N t O t τ T = 0 , (3)
Finally, the feedback response to slowly evolving climatic forcings can be solved as:
B = A t O t τ T O t O t τ T . (4)
In this paper, following the methodology of Notaro et al. (2019), potential predictors are identified by dividing the global ocean into eight non-overlapping basins and northern tropical Africa into four climatically and ecologically distinct regions (Table 3), namely the Sahel, Horn of Africa (HOA), WAM region, and Congo. The oceanic forcings in SGEFA's forcing matrix is represented by the time series for the principal components (PCs) associated with the top two empirical orthogonal function modes of SST variability in each ocean basin and the area-averaged Mediterranean SSTs (MED). The time series of area-averaged LAI and soil moisture over the four sub-Saharan regions correspond to the terrestrial forcings within the forcing matrix. The anomalies of the forcing data and response fields are obtained by removing the seasonal cycle and third-order polynomial trend. To summarize, the full forcing matrix, as computed in this study, is represented as:
O = [ TP 1 , TP 2 , NP 1 , NP 2 , TA 1 , TA 2 , NA 1 , NA 2 , TI 1 , TI 2 , SP 1 , SP 2 , SA 1 , SA 2 , SI 1 , SI 2 , MED , V Sahel , V WAM , V HOA , V �Congo , S Sahel , S WAM , S HOA , S �Congo ] (5)
For example, NP1 and TI2 represent the first PC of North Pacific SST and the second PC of tropical Indian SST, respectively. V and S indicate LAI and soil moisture, respectively, and the subscripts indicate the region. For instance, VSahel and SCongo represent the area-averaged LAI over the Sahel and area-averaged soil moisture over the Congo, respectively.
TABLE 3. Geographic extent and abbreviation for different ocean basins and land areas for performing area-averages and principal component analysis for SGEFA
Ocean basin Abbreviation Region
Tropical Pacific TP 20°S–20°N, 120°E–60°W
North Pacific NP 20°N–60°N, 120°E–100°W
South Pacific SP 60°S–20°S, 150°E–70°W
Tropical Indian TI 20°S–20°N, 35°E–105°E
South Indian SI 60°S–20°S, 20°E–120°E
Tropical Atlantic TA 20°S–20°N, 70°W–20°E
North Atlantic NA 20°N–60°N, 90°W–10°W
South Atlantic SA 60°S–20°S, 70°W–20°E
Mediterranean MED 30°N–50°N, 0°E–50°E
Land region Abbreviation Region
Sahel Sahel 12°N–17°N, 20°W–40°E
West African monsoon WAM 5°N–12°N, 20°W–30°E
Horn of Africa HOA 10°S–10°N, 30°E–50°E
Congo Congo 10°S–5°N, 10°E–30°E
The reliability of traditional GEFA method is limited by the sampling errors associated with simultaneously considering too many forcing; here, 25 forcings are included. By applying a backward-selection stepwise method (Hocking, 1976), SGEFA improves on the traditional GEFA method by removing relatively irrelevant forcings from the complete forcing matrix. The backward selection regression approach consists of starting with all candidate forcings, testing the deletion of each forcing one by one according to a chosen model fit criterion, deleting the forcing whose elimination yields the most statistically insignificant decline in model fit, and repeating the process until no additional forcings can be eliminated without a statistically significant reduction in fit. Then, the statistical prediction model is formed using the most relevant forcings as predictors of a select atmospheric variable, according to an automated procedure. Using the Akaike Information Criteria (AIC, Akaike, 1974), the relative quality of the statistical prediction model is estimated as:
AIC = 2 × N f 2 × ln L ̂ (6)
N f is the number of individual forcings within the forcing matrix, and L ̂ is the maximized likelihood function of the model, which is determined by:
L ̂ = L 2 ln t = 1 L A ̂ t A t 2 / L + C 1 (7)
A ̂ t = B × O t (8)
where L is the length of the data record, C 1 is a constant term that is independent of B , and A ̂ t is the estimated atmospheric state at time t . In this study, on average, about 50% of the forcings are dropped from the full forcing matrix by the stepwise selection process, which significantly reduces the sampling errors and enhances the reliability of GEFA.

The reliability of SGEFA is limited by the length of the time series due to sampling errors (Wang et al., 2014; Yu et al., 2018). To achieve sufficient data record, which is about 100 years based on our previous studies (Yu et al., 2018), the monthly anomalies are combined into bins of three consecutive months to essentially triple the time series' record length (Czaja and Frankignoul, 2002; Yu et al., 2017, 2018). Then, the seasonal SGEFA is estimated using this aggregated time series. For example, the response in July–August–September (JAS) rainfall (R) anomalies to soil moisture anomaly forcings (S) is estimated using the forcing time series of (SJune yr 1, SJuly yr 1, SAugust yr 1, SJune yr 2, SJuly yr 2, SAugust yr 2,…) and the atmospheric time series of (RJuly yr 1, RAugust yr 1, RSeptember yr 1, RJuly yr 2, RAugust yr 2, RSeptember yr 2,…). This approach ensures that the forcing always precedes the response by 1 month in seasonal SGEFA, as the land surface's response to the atmosphere likely dominates over the atmosphere's response to the land surface, and thus forcing and response variables cannot be considered at overlapping times.

The statistical significance, based on 95% confidence level, of the SGEFA feedback matrix is assessed using the Monte Carlo bootstrap method by randomly iterating the atmospheric variable's time series 10,000 times (Czaja and Frankignoul, 2002; Wang et al., 2013, 2014, 2017; Notaro et al., 2019).

3 FIDELITY OF SIMULATED SAHEL CLIMATE AND LAND SURFACE CONDITIONS

The Sahel is bordered by the Saharan Desert to the north, Sudanian savanna to the south, Atlantic Ocean to the west, and Red Sea to the east. In this semi-arid area, the primary land cover types are bare ground (27%), grassland (24%), savannas (23%), woody savannas (13%), and shrublands (10%) (Yu et al., 2017). As part of the WAM system, Sahel rainfall has a clear observed seasonal cycle, with a peak in JAS (Figure 1a). The rainfall during this peak season accounts for close to 70% of the observed total annual mean rainfall. The seasonal peaks in root-depth soil moisture (Figure 1b) and LAI (Figure 1c) lag precipitation by 1 month due to the observed dominant control of rainfall on land surface conditions in this semi-arid region.

Details are in the caption following the image
Seasonal cycle of median (a) precipitation (unit: Mm day−1), (b) normalized total column soil moisture (unitless), and (c) leaf area index (unit: M2 m−2) from observations (red), multi-model mean (MMM) of CMIP5 models with DGVM (green), and MMM of CMIP5 models without DGVM (blue) across the Sahel. The red, green, and blue shading represents the extent of the 10th and 90th percentiles across multiple ‘observed’ data sets, CMIP5 models with DGVM, and CMIP5 models without DGVM. Because only one soil moisture datum, GLEAM, is used, no red shading is provided for soil moisture. The month here represents three consecutive months starting from the current month. When comparing the pool of DGVM-enabled models and non-DGVM models, the red shading indicates significant difference between the two multi-model ensembles at the 95% confidence level and the unfilled red box indicates significant difference at the 90% confidence level

The simulated MMM seasonal cycle of precipitation is generally consistent with that of observations, with both displaying a unimodal seasonal cycle, which peaks in JAS. However, the MMM overestimates the pre-monsoon rainfall in spring and underestimates the seasonal peak in rainfall in summer-autumn, leading to a broader-than-observed peak in the seasonal cycle of simulated precipitation (Figure 1a). The dry bias during the peak monsoon season is believed to be mainly attributed to the models' insufficient spatial extent and intensity of the Atlantic Cold Tongue (Steinig et al., 2018; Monerie et al., 2020). Due to the MMM's positive SST bias in the eastern tropical Atlantic Ocean, the simulated thermal contrast between the Saharan Desert and equatorial Atlantic Ocean is underestimated, resulting in a weaker monsoon system with limited northward penetration and thus insufficient simulated monsoonal rainfall across the Sahel (Davey et al., 2002; Nicholson, 2013; Richter et al., 2014). During the pre- to early monsoon period of March–July, 83% of non-DGVM models and 51% of DGVM-enabled models exhibit a wet bias (Figure 1a). The difference in the MMM precipitation simulated by DGVM-enabled ESMs and non-DGVM ESMs is statistically significant (p-value <0.05) according to the Student's t-test, suggesting that ESM coupling to DGVMs can help decrease the wet bias found in non-DGVM ESMs.

The simulated MMM seasonal cycle of normalized total column soil moisture is consistent with normalized root-depth soil moisture as estimated from the GLEAM in terms of the seasonal timing (Figure 1b). Since the details of the coupled land component, including the depth of the total soil column and the number and depths of each individual soil layer, vary among CMIP5 ESMs (Mahowald et al., 2016), it is difficult to extract soil moisture from the surface to root-depth level for each model to directly compare with GLEAM soil moisture. Therefore, the simulated normalized total column soil moisture, rather than the raw soil moisture data, is used. Among the standard CMIP5 outputs, only top layer and total column soil moisture are available for analysis. Given that GEFA requires that the memory of the forcing field under consideration exceeds that of the atmospheric variable, the total column soil moisture is more appropriate than top layer soil moisture to analyse given its greater persistence time. The persistence time, measured by the monthly autocorrelation coefficient of total column soil moisture anomalies, is at least 1 month, which exceeds the atmospheric internal variability.

As a result of the dominant role of precipitation in regulating LAI across the Sahel, the simulated MMM seasonal cycle of LAI in both DGVM-enabled ESMs and non-DGVM ESMs has a broader-than-observed peak due to the overly broad precipitation peak in the MMM (Figure 1c). The non-DGVM ESM ensemble is characterized by a wet bias with excessive LAI during the pre-monsoon season and dry bias with insufficient LAI during the peak monsoon season. Although DGVMs are capable of simulating temporal variability in vegetation cover and PFT distributions, the accuracy of their simulated vegetation patterns depends on the accuracy of the simulated atmospheric state and the DGVMs' applied parameterizations, broad definitions of PFTs, and treatment of vegetation competition, among other land surface processes (Scheiter et al., 2013). Thus, the inter-model range in the seasonal cycle of LAI is much larger in DGVMs than in non-DGVMs, while the inter-model range in non-DGVM is closer to observations. Most DGVM-enabled ESMs overestimate the year-round area-average LAI, while most non-DGVM models simulate comparable LAI to observations across the Sahel. Mahowald et al. (2016) also found globally overestimated LAI in ESMs. In general, the ESMs without DGVMs simulate a more realistic LAI climatology over the Sahel, as such models receive realistic PFT distributions as fixed boundary condition inputs, while the DGVM-enabled ESMs dynamically generate their own PFT distributions and end up overestimating year-round LAI (Figure 1c).

In summary, the CMIP5 models reasonably capture the Sahel's seasonal cycle in precipitation and soil moisture, but frequently exhibit a wet pre-monsoon bias and dry peak monsoon bias. The non-DGVM ESM ensemble outperforms with DGVM-enabled ESM ensemble in simulating the Sahel's LAI by prescribing reasonable PFT distributions to reflect observations.

4 PRECIPITATION PREDICTABILITY

The simulated precipitation predictability in terms of EV by oceanic (SST PC time series) and terrestrial (soil moisture and LAI) forcings is investigated and compared with its observed counterpart. According to observational data, the annual mean total variability in Sahel rainfall that can be explained by oceanic and terrestrial drivers together is 36%, with an across-data set range of 33% for the 10th percentile to 38% for the 90th percentile, depending on the combination of observational data sets considered (Figure 2a). In contrast, almost all of the CMIP5 models underestimate the total EV of Sahel precipitation by oceanic and terrestrial forcings, with an annual MMM value of 29%, with an across-model range of 25% for the 10th percentile of models to 33% for the 90th percentile (Figure 2a). The annual mean total EV in Sahel rainfall in the CMIP5 MMM total is significantly different from the observed EV, based on the Student's t-test at the 95% confidence level. The primary observed driver of Sahel rainfall variability across its full annual cycle is soil moisture, with an annual mean EV of 17% (13–20%), followed by SSTs, with an annual mean EV of 14% (12–16%), and finally LAI, with an annual mean EV of 5% (2–7%). In comparison, the primary CMIP5 driver is SSTs, with an annual mean EV of 17% (13–23%), followed by soil moisture, with an annual mean EV of 9% (5–16%), and finally LAI, with an annual mean EV of 3% (1–5%). In observations, the relative importance of soil moisture and SSTs in regulating year-round Sahel rainfall is comparable, in contrast to the CMIP5 models, in which SSTs dominate. The limited direct contribution of LAI forcings on rainfall variability is consistent between the observations and models. The annual mean EV of Sahel rainfall due to variability in SST, soil moisture, and LAI in the CMIP5 MMM is significantly different from observations according to the 95% confidence level.

Details are in the caption following the image
Percent explained variance of (a) annual and (b) JJA precipitation anomalies across the Sahel attributed to oceanic (blue), vegetation (green), soil moisture (red/orange), and total (black/grey) forcings in observations, CMIP5 models, CMIP5 models with DGVM, and CMIP5 models without DGVM during 1960–2000. The shading indicates the 10th and 90th percentiles across either multi-observational data sets or multi-CMIP5 models. The solid horizonal lines indicate the 50th percentile. The red boxes in DGVM and non-DGVM columns indicate that the difference between DGVM-enabled ESMs and non-DGVM ESMs are statistically significant at the 90% (dash boxes) and 95% (solid boxes) confidence levels

In both observations and CMIP5 models, Sahel rainfall is most predictable, based on slowly evolving oceanic and terrestrial forcings, during the growing season, especially the peak monsoon season (Figure 3a, b). The dominant role of soil moisture in regulating observed Sahel precipitation is more obvious during the peak monsoon season compared to the annual mean (Figure 2b). The total observed EV of Sahel rainfall during JJA is 50% (44–58%), which can be decomposed into 33% (21–42%) due to soil moisture as the observed dominant driver, 12% (6–20%) due to SST, and 5% (0–16%) due to LAI. As also true for the annual mean EV, almost all CMIP5 models underestimate the EV of peak monsoon season rainfall, with a MMM value in JJA of 35% (26–44%), which can be decomposed into 21% (11–32%) due to SST as the simulated dominant driver, 11% (2–29%) due to soil moisture, and 3% (0–6%) due to LAI. For JJA, the differences between simulated and observed mean EV in rainfall due to total combined forcings, SST forcings alone, soil moisture forcings alone, or LAI forcings alone are all statistically significant at the 95% confidence level.

Details are in the caption following the image
Seasonal cycle of percent explained variance of seasonal precipitation anomalies across the Sahel attributed to oceanic (blue), vegetation (green), soil moisture (red), and total (black/grey) forcings in (a) observations, (b) CMIP5 models, (c) CMIP5 models without DGVM, and (d) CMIP5 models with DGVM during 1960–2000. The shading indicates the 10th and 90th percentiles across either multi-observational data sets or multi-CMIP5 models. The solid horizonal lines indicate the 50th percentile. The month here represents three consecutive months starting from the current month

The overestimated SST contribution and underestimated soil moisture contribution to Sahel precipitation variability during the growing season are typically more severe in CMIP5 models without an active DGVM, consisting of 25% (19–34%) from SSTs and 5% (2–11%) from soil moisture in non-DGVM ESMs compared with 19% (10–29%) and 14% (4–30%) in DGVM-enabled ESMs, respectively. The difference between DGVM-enabled ESMs and non-DGVM ESMs in JJA mean EV of Sahel rainfall attributed to soil moisture forcings is statistically significant at the 95% confidence level based on the Student's t-test and further explored mechanistically in section 5.

In summary, almost all CMIP5 models underestimate the annual mean total EV of Sahel precipitation attributed to oceanic and terrestrial forcings (Figure 2a, b). The main difference between observed and simulated total EV of Sahel precipitation is the proportion of contributions from SSTs versus soil moisture. Inaccuracies in the simulated proportion of precipitation EV due to oceanic versus terrestrial forcings are most apparent in non-DGVM models, compared with DGVM-enabled ESMs, especially during the growing season.

5 SAHEL RAINFALL RESPONSE TO LOCAL TERRESTRIAL FORCINGS: CMIP5 MODELS VERSUS OBSERVATIONS

The simulated precipitation response to local fluctuations in soil moisture and LAI is investigated and compared against its observed counterpart in this section. Here, ‘local’ refers to within the Sahel, so local terrestrial forcings are area-average LAI and soil moisture over the Sahel. Correspondingly, ‘non-local’ indicates outside of the Sahel. For example, the oceanic forcings, along with the area-average LAI and soil moisture over the other terrestrial regions, are all considered non-local forcings. To measure the response consistency across either different observational data sets or different CMIP5 models, the response consistency index is computed as the percentage of response fields among all data set pairs that agrees in sign with the mean of the significant responses across data sets.

The majority of CMIP5 models underestimates the observed Sahel rainfall response to local soil moisture forcings across the seasonal cycle, according to SGEFA (Figure 4a–c). The Sahel rainfall response to soil moisture forcings is typically significant and moderate in intensity in DGVM-enabled ESMs (Figure 4b) and trivial in non-DGVM ESMs (Figure 4c). In observations, a positive soil moisture-precipitation feedback is detected during AMJ to SON, with high consistency (76–100%) among different data sets combinations (Figure 4a). The observed precipitation feedback peaks in JJA and ASO, with strengths of +0.41 and +0.35 mm day 1 σ SM 1 (representing the increase in rainfall induced by an increase of one standard deviation in soil moisture), respectively. For the DGVM-enabled ESMs, a weak but significant positive soil moisture-precipitation feedback is detected from MJJ to SON, with moderate consistency (60–80%) among DGVM-enabled ESMs. The precipitation response is rather uniform in time across the months of MJJ to SON, with the strength ranging from +0.12 to +0.14 mm day 1 σ SM 1 . For the non-DGVM ESMs, a trivial positive soil moisture-precipitation response is detected, with models showing moderate consistency (62–75%) limited only to JAS and SON, and the response strength is one order smaller than that of the observations. The underestimated rainfall response to soil moisture in CMIP5 models, especially the non-DGVM models, is consistent with the underestimated EV in both DGVM-enabled and non-DGVM ESMs as discussed in section 4.

Details are in the caption following the image
Seasonal cycle of precipitation response to local soil moisture (a–c) and leaf area index (d–f) anomalies over the Sahel in observations (a, d), Earth System Models (ESMs) with DGVM (b, e), and non-DGVM ESMs (c, f). For observations, the multi-data-set combination is analysed, and for the models, the multi-model mean is analysed. The mean response is represented by squares, and the 10th, 50th, and 90th percentiles across the multi-data set/multi-model combination are shown with horizontal black tick marks. The shading indicates the sign consistency among either the different data set combination for observations or among different CMIP5 models. The unit is mm day−1 σ (forcing)−1. The month here represents three consecutive months starting from the current month

Precipitation feedback responses to either soil moisture or LAI can result from several competing mechanisms, such as the moisture recycling mechanism, albedo feedback, stability mechanism, and roughness mechanism. To explore the mechanisms, the responses of key variables, such as ET and its components (transpiration, soil evaporation, and canopy evaporation), 2-m specific humidity, outgoing longwave radiation (OLR), precipitable water, sea-level pressure, low-level (925 hPa) vertical motion (ω) and moisture convergence, and 2-m air temperature, to soil moisture and LAI forcings are investigated.

On the monthly time scale, the mechanisms of moisture recycling and moisture convergence are primarily responsible for the SGEFA-identified observed soil moisture–rainfall feedback across the Sahel. In observations, positive soil moisture anomalies lead to significant increases in total ET from AMJ to NDJ, with high consistency (87–100%) among multi-data set combinations (Figure 5a). Among the components of ET, the observed ET response to soil moisture, which stretches across the entire pre- to post-monsoon period, mainly consists of a dominant soil evaporation response during AMJ - NDJ (Figure 5g), significant transpiration response during SON - NDJ (Figure 5d), and trivial response from canopy evaporation (Figure 5j). Positive soil moisture anomalies generally support an observed enhancement of ET (Figure 5a), which leads to increases in 2-m specific humidity (Figure 6a), precipitable water (Figure 6d), convective activity as indicated by reductions in OLR (Figure 6g), and precipitation. The observed ET responses are about one-third of the magnitude of the precipitation responses (Figures 4a, 5a), indicating that beyond this moisture recycling mechanism, moisture convergence also plays a role. During JJA - ASO, the response of moisture convergence is highly consistent among the combination of observational data sets, supporting moisture influx into the Sahel and increased rainfall. Other feedback mechanisms compete against each other, resulting in the observed rainfall response to soil moisture drivers. For example, the increased ET leads to near-surface atmospheric cooling (Figure 7j) and greater stability of the lower atmosphere, as indicated by anomalous descent in 925 hPa vertical motion (Figure 7g), with high consistency (87–100%) among different combinations of data sets, which support less rainfall. Sea-level pressure generally increases in response to the atmospheric cooling and stabilization but with low across-data set consistency (Figure 7d).

Details are in the caption following the image
Seasonal cycle of the response of evapotranspiration and its components, namely transpiration, evaporation from soil, and evaporation from canopy, to local soil moisture anomalies over the Sahel in observations (a, d, g, j), DGVM-enabled Earth System Model (b, e, h, k), and non-DGVM ESMs (c, f, i, l). For observations, the multi-data-set combination is analysed, and for models, the multi-model mean is analysed. The mean response is represented by squares, and the 10th, 50th, and 90th percentiles across the multi-data set/multi-model combination are shown with horizontal black tick marks. The shading indicates the sign consistency among either the different data set combination for observations or among different CMIP5 models. The unit is mm day−1 σ (forcing)−1. The month here represents three consecutive months starting from the current month
Details are in the caption following the image
Seasonal cycle of the response of 2-m specific humidity [g kg−1 σ (forcing)−1], precipitable water [kg m−2 σ (forcing)−1], cloud area fraction [% σ (forcing)−1], and OLR [W m−2 σ (forcing)−1] to local soil moisture anomalies over the Sahel in observations (a, d, g, j), DGVM-enabled Earth System Models (ESMs) (b, e, h, k), and non-DGVM ESMs (c, f, i, l). For observations, the multi-data set combination is analysed, and for models, the multi-model mean is analysed. The mean response is represented by squares, and the 10th, 50th, and 90th percentiles across the multi-data set/multi-model combination are shown with horizontal black tick marks. The shading indicates the sign consistency among either the different data set combination for observations or among different CMIP5 models. The month here represents three consecutive months starting from the current month
Details are in the caption following the image
Seasonal cycle of the response of moisture convergence at 925 hPa [g/kg s−1 σ (forcing)−1], sea-level pressure [hPa σ (forcing)−1], 925 hPa vertical motion [Pa s−1 σ (forcing)−1], and 2-m air temperature [°C σ (forcing)−1] to local soil moisture anomalies over the Sahel in observations (a, d, g, j), DGVM-enabled Earth System Models (b, e, h, k), and non-DGVM ESMs (c, f, i, l). For observations, the multi-data-set combination is analysed, and for models, the multi-model mean is analysed. The mean response is represented by squares, and the 10th, 50th, and 90th percentiles across the multi-data set/multi-model combination are shown with horizontal black tick marks. The shading indicates the sign consistency among either the different data set combination for observations or among different CMIP5 models. The month here represents three consecutive months starting from the current month

The observed response of near surface air temperature to soil moisture forcing is the competing result of concurrent changes in ET and surface albedo. Increased soil moisture leads to both an enhancement in ET, which cools the near-surface atmosphere, and a lower surface albedo (Yu et al., 2017), which warms the near-surface atmosphere. In observations, a net reduction in the near-surface air temperature is detected in response to anomalously wet soils (Figure 7j), indicating that the ET-associated cooling effect dominates over the albedo-associated warming effect across the Sahel.

The CMIP5 models, both with and without DGVMs, can capture the observed positive response of total ET to positive soil moisture anomalies, with moderate to high consistency (75–88%) (Figure 5b, c). DGVM-enabled ESMs generate an ET response of comparable magnitude to observed estimates in response to Sahel soil moisture forcings, with a feedback of +0.084 mm day 1 σ SM 1 in observations and +0.079 mm day 1 σ SM 1 in DGVM-enabled ESMs during April–November (Figure 5a, b). In contrast, ESMs without DGVMs can only reproduce approximately half of the observed response (+0.044 mm day 1 σ SM 1 , Figure 5c). In ESMs without DGVMs, both soil evaporation and transpiration responses to soil moisture forcings are underestimated compared to observations (Figure 5f, i). The response magnitude of canopy evaporation to local soil moisture forcing is negligible compared with the other two ET components in both observations and CMIP5 models (Figure 5j–l). Compared with the high consistency among different observed data sets, both DGVM-enabled ESMs and non-DGVM ESMs exhibit large across-model discrepancies in terms of the individual responses of each ET component to soil moisture forcings (Figure 5f,h,i,k,l), with low across-model consistency (<50%). One exception is the simulated transpiration response to soil moisture forcings in DGVM-enabled ESMs (Figure 5e), which achieves moderate across-model consistency (60–80%). The difference in ET responses to Sahel soil moisture anomalies between DGVM-enabled ESMs and non-DGVM ESMs is mainly attributed to transpiration. The transpiration response magnitude of DGVM-enabled ESMs is about triple that of non-DGVM ESMs (Figure 5e, f). The consistency of the transpiration response among DGVM-enabled ESMs is moderate, while that of non-DGVM ESMs is low. Although the non-DGVM ESMs have reasonably prescribed vegetation types, the vegetation cannot transfer soil moisture to the atmosphere efficiently, which warrants further investigation.

In DGVM-enabled ESMs, the soil moisture-induced enhancement of ET leads to increases in 2-m specific humidity (Figure 6b), precipitable water (Figure 6e), and total cloud cover fraction (Figure 6h) and reductions in OLR (Figure 6k), but with weaker response magnitudes and lower consistency compared to their observed counterparts (Figure 6a, d, g, j). The simulated response of 2-m specific humidity in DVGM-enabled ESMs can only reproduce approximately one-third of the observed response (+0.16 g kg 1 σ SM 1 in DGVM-enabled ESMs versus +0.46 g kg 1 σ SM 1 in observations). As a consequence, the soil moisture-induced responses in precipitable water and convection are also underestimated and rather inconsistent among DGVM-enabled ESMs (Figure 6e, h, k). The simulated ET response to soil moisture forcing and the simulated precipitation response to soil moisture are of comparable magnitude (Figure 5b, Figure 4b), implying the dominant role of the moisture recycling mechanism in regulating Sahel precipitation. Unlike the observations, where both moisture recycling and moisture convergence contribute to the positive soil moisture - precipitation feedback, a weak moisture divergence response to soil moisture forcings instead exists in DGVM-enabled ESMs (Figure 7b) with low to moderate across-model consistency. The missing signal of enhanced moisture convergence, as seen in observational analyses, helps explain the underestimated response of precipitation to soil moisture anomalies in most DGVM-enabled ESMs.

The precipitation response to soil moisture forcing is trivial in the non-DGVM ESMs, unlike in the DGVM-enabled ESMs and observations (Figure 4a–c). Responses among all atmospheric variables and surface-atmosphere flux variables, except for total ET (Figure 5c), are negligible, with high across-model inconsistency, for the non-DGVM ESMs (Figures 5-7). This suggests that coupling to a DGVM (or equivalent dynamical land model) is needed to study land-atmosphere feedback and their role in future climate change projections for the Sahel.

The spatial pattern of the SGEFA response in JJA, when the observed precipitation response to soil moisture is strongest, is further examined. The observed wetter and colder response to Sahel soil moisture anomalies is mainly local, across the whole Sahel region (Figure 8a, g), with high consistency among different data set combinations (Figure 8d, j). The observed positive ET response to soil moisture anomalies is also mainly local and more limited to the eastern Sahel (Figure 9a) with moderate to high consistency among different data sets (Figure 9d). The spatial pattern of the 2-m air temperature response (Figure 8g) matches that of the ET response (Figure 9a), indicating that the ET-associated cooling effect dominates over the albedo-associated warming effect across the Sahel. The low-level moisture convergence responses are different between the western and eastern Sahel, with a moisture convergence response across the western Sahel and moisture divergence response across the eastern Sahel (Figure 9m). This can be explained by the greater cooling effect in the eastern Sahel, which prevents the WAM dynamic circulation from penetrating far inland, leading to anomalous moisture divergence there. The anomalous moisture convergence response to wet soils in the western Sahel leads to more precipitable water (Figure 9g). Consistent with the regional mean SGEFA results discussed above, the simulated positive precipitation response is also mainly local but much weaker than observed (Figure 8b) and with moderate consistency among different models (Figure 8e); specifically, for non-DGVM ESMs, the precipitation response is trivial with low multi-model consistency.

Details are in the caption following the image
Spatial pattern of the mean response of precipitation [mm day−1 σ (forcing)−1] and 2-m air temperature [°C σ (forcing)−1] to local soil moisture anomalies over the Sahel during JJA in observations, DGVM-enabled Earth System Models (ESMs), and non-DGVM ESMs. For observations, the multi-data sets combination is analysed, and for models, the multi-model mean is analysed. The corresponding sign consistency among either the different data set combination for observations or among different CMIP5 models is shown. Only sign consistency larger than 50% is shown. Green box indicates Sahel region
Details are in the caption following the image
Spatial pattern of the mean response and sign consistency of ET [mm day−1 σ (forcing)−1], precipitable water [kg m−2 σ (forcing)−1], and moisture convergence [g/kg s−1 σ (forcing)−1] to local soil moisture anomalies over the Sahel during JJA in observations, DGVM-enabled Earth System Models (ESMs), and non-DGVM ESMs. For observations, the multi-data set combination is analysed, and for models, the multi-model mean is analysed. The corresponding sign consistency among either the different data sets combination for observations or among different CMIP5 models is shown. Only sign consistency larger than 50% is shown. Green box indicates Sahel region

Sahel rainfall is much less sensitive to LAI anomalies than soil moisture anomalies according to both observations and CMIP5 models (Figure 4). In observations, a negative vegetation-rainfall feedback is detected, but only significant during ASO (Figure 4d), with 76% consistency among observed data sets. Among CMIP5 models, this negative vegetation-rainfall feedback typically occurs during JAS in most of the DGVM-enabled ESMs (Figure 4e), with negligible rainfall responses in non-DGVM ESMs (Figure 4f). Except for significant transpiration responses to positive LAI anomalies (Figure S1d–f), no consistent MMM response is detected in other examined atmospheric variables and surface-atmosphere flux variables, including soil evaporation, canopy vegetation (Figure S1), near-surface specific humidity, precipitable water, OLR (Figure S2), near-surface wind speed, sea-level pressure, low-level vertical motion, and near-surface air temperature (Figure S3). The negligible precipitation response to LAI, but significant response to soil moisture, indicates that vegetation has minimal direct effect on the atmospheric state over the Sahel, but its indirect effect of modifying soil moisture and transpiration rate is important for DGVM-enabled ESMs to better simulate rainfall responses to soil moisture forcings than non-DGVM ESMs.

In summary, Sahel rainfall exhibits a greater sensitivity to soil moisture anomalies than LAI anomalies in both the observations and CMIP5 ESMs, although both DGVM-enabled ESMs and non-DGVM ESMs underestimate the precipitation response to local soil moisture forcings over the Sahel. The atmospheric responses to terrestrial forcings in the Sahel are typically trivial among non-DGVM ESMs. In observations, both moisture recycling and moisture convergence contribute to the Sahel's rainfall response to soil moisture anomalies, while the moisture recycling mechanism dominates the response to soil moisture anomalies in DGVM-enabled ESMs and non-DGVM ESMs.

6 DISCUSSION AND CONCLUSIONS

This study systematically investigates the simulated representation of the influence of terrestrial forcings on Sahel rainfall among 23 CMIP5 ESMs during the historical period of 1960–2000 in comparison to observations. The models' performance in simulating the climatology of precipitation and the seasonal evolution of soil moisture and LAI over the Sahel, the isolated precipitation response to terrestrial forcings, and the associated competing mechanisms are evaluated. The performance of DGVM-enabled and non-DGVM ESMs in simulating the Sahel's climatology and atmospheric responses to terrestrial feedback is further explored. Although non-DGVM ESMs usually outperform DGVM-enabled ESMs in simulating the mean climatology over the Sahel, a DGVM-enabled ESM is critical in generating reasonable land–atmosphere feedback. The main findings that deepen our understanding of land–atmosphere feedback over the Sahel among CMIP5 models are as follows.
  • The CMIP5 models capture the seasonal cycle of Sahel precipitation but overestimate pre-monsoon rainfall and underestimate rainfall during the peak monsoon season. Compared to the non-DGVM ESMs, the DGVM-enabled ESMs struggle less with the pre-monsoon wet bias but fail to improve upon the dry bias during the peak monsoon season. This dry bias in both DGVM-enabled and non-DGVM ESMs is believed to be mainly attributed to the SST bias over the Atlantic cold tongue that is prevalent among CMIP5 models (Steinig et al., 2018). Whittleston et al. (2017) also examined this dry bias during the peak monsoon season, while identifying insufficient coupling between atmospheric jets in West Africa (e.g., tropical easterly jet, African easterly jet, and African westerly jet) and rainfall in most CMIP5 models. However, it remains uncertain if this deficiently simulated jet stream-rainfall coupling is the cause of the dry bias.
  • The CMIP5 models typically generate a reasonable seasonal evolution of normalized total soil moisture.
  • The non-DGVM ESMs outperform the DGVM-enabled ESMs in simulating the mean seasonal cycle of Sahel LAI due to the prescribed PFT distribution, assigned to resemble the observed global vegetation patterns, in the former set of models. DGVM-enabled ESMs generally overestimate LAI over the Sahel, which is a common bias across all latitudes in ESMs (Mahowald et al., 2016).
  • The predictability indicated by the total EV in precipitation associated with slowly-evolving oceanic and terrestrial forcings is smaller in CMIP5 models than in observations. The underestimated precipitation predictability in CMIP5 models mainly originates from the terrestrial portion of the coupling. This is particularly a concern for non-DGVM ESMs, as they substantially underestimate the EV of Sahel precipitation attributed to terrestrial forcings.
  • Sahel rainfall is more sensitive to local soil moisture forcings than LAI forcings in both observations and CMIP5 ESMs, although ESMs, especially those not coupled to a DGVM, underestimate the precipitation response to soil moisture forcings. Due to the moderate to high temporal correlations between soil moisture and vegetation indices across the Sahel (Schnur et al., 2010; Anyamba et al., 2014), the contribution of either soil moisture or vegetation index has rarely been estimated. In this study, we isolated the individual contributions from soil moisture and LAI and found a stronger influence from soil moisture than LAI on Sahel precipitation.
  • The observed Sahel's rainfall response to soil moisture anomalies is explained by both moisture recycling and moisture convergence mechanisms, while the simulated rainfall response is dominated by the moisture recycling mechanism in DGVM-enabled ESMs. The SGEFA method can address both local and non-local feedback of terrestrial forcings. However, in this study, only local terrestrial precipitation feedback are detected, and the remote impacts of soil moisture as identified by Lavender et al. (2010) and Berg et al. (2017) are not detected over the Sahel, with non-local feedback responses failing to achieve statistical significance in both observations and CMIP5 models.
Some limitations of this study are identified here. The accuracy of the observed rainfall response to terrestrial forcings is limited by the large uncertainties in ‘observed’ root-depth soil moisture, ET and its components, and other atmospheric variables, such as precipitable water, low-level wind, pressure, and humidity, some of which are not directly observed. For example, the GLEAM root-depth soil moisture data are a data assimilation product (Martens et al., 2017) generated using a multi-layer running water balance and assimilating satellite observations of soil moisture and ocean salinity. Besides GLEAM soil moisture, the Climate Prediction Center (CPC) soil moisture data set is also briefly explored and compared with soil moisture measurements from the African Monsoon Multidisciplinary Analysis (AMMA). Given that the CPC soil moisture data was developed using a simple bucket model and tuned specifically for the central United States, it yields less favourable agreement with AMMA measurements, so only GLEAM soil moisture data is utilized in this study. ET and its components should be considered as outputs from remote-sensing-based models, land surface models, and hydrological models rather than direct observations (Weerasinghe et al., 2020). To obtain more reliable estimates, multiple data sets are analysed to check the across-data-set consistency and estimate the uncertainty. However, the three ET data sets used in this study are not independent of each other, which affects the estimation of response uncertainty and range. For example, the merged ET data by Mao et al. (2015) is built upon multiple ET data sets, including GLEAM and PML ET, and both GLEAM and PML ET products are generated using the Penman-Montieth equation.

The current study does not address the potential connection between the sign and intensity of an ESM's simulated land-atmosphere interactions and its representation of dust emission, transport, and radiative effects in the Sahel. Our earlier observational analysis (Yu, 2017) suggested that diminished growth in vegetation and accompanying anomalously dry soils in the Sahel favour greater dust emissions across the southern boundary of North Africa's primary dust source regions in the Sahel, such as the Bodélé Depression (e.g., Yu et al., 2018b). The simulated land-dust feedback likely alter the simulated magnitude of land-precipitation feedback in the Sahel through dust's local and non-local radiative forcings (e.g., Miller et al., 2014; Jordan et al., 2018). Future assessments of CMIP ESMs over sub-Saharan Africa should expand the current SGEFA analysis to explore the contribution of desert dust feedback in regulating the sign and intensity of soil moisture-vegetation-atmosphere feedback.

The statistical method, SGEFA, is based on a linear assumption; therefore, the results presented here should be viewed as first-order approximations. In this study, the application of SGEFA is limited to the Sahel with a focus on precipitation and its associated drivers. However, SGEFA is a generalizable method that can be extended to other regions, data sets, and environmental variables but should be done so with caution. In particular, the length of the data record necessary to obtain stable estimates of the feedback responses should be assessed for the focus region and environmental fields. Furthermore, it is necessary to verify that the memory, or temporal correlation, of the forcings under consideration exceeds that of the atmospheric response variables in order for SGEFA to be applicable.

In this study, positive soil moisture – rainfall feedback at the monthly time scale are identified in both observations and DGVM-enabled CMIP5 ESMs, while Taylor et al. (2012, 2013) found negative feedback in observations and positive feedback in GCMs at the shorter weather time scale over the Sahel. Other studies have also determined that the sign and intensity of soil moisture-precipitation feedback can vary by time scale for other regions. For example, Duerinck et al. (2016) found strong positive correlations between late spring/early summer root-depth soil moisture and subsequent summer rainfall over Illinois (United States), but no correlation was detected on the daily to weekly time scales. Ford et al. (2015) also explored soil moisture-rainfall feedback on the sub-daily time scale and concluded that the soil moisture-precipitation feedback is closely related to the existing synoptic patterns. Beyond time scales, spatial resolution and convective parameterizations are also important for climate models to simulate reasonable soil moisture-atmosphere coupling (Taylor et al., 2013).

The present study has demonstrated the value of SGEFA in assessing the credibility of ESMs regarding their representation of the oceanic and terrestrial drivers of regional climate, which is critical in quantifying the models' credibility in their future climate change projections. In a subsequent study, the authors will investigate the potential to narrow the uncertainty in projected future Sahel rainfall (Biasutti, 2013; Roehrig et al., 2013) by weighting models based on their performance in terms of the precipitation response to terrestrial and oceanic forcings.

ACKNOWLEDGEMENTS

The authors acknowledge funding from the Department of Energy (DOE) through grant DE-SC0012534. This research was also partially supported by the Reducing Uncertainties in Biogeochemical Interactions through Synthesis and Computation Science Focus Area (RUBISCO SFA DE-AC05-00OR2272). RUBISCO is sponsored by the Regional and Global Model Analysis activity of the Earth and Environmental Systems Modelling Program in the Earth and Environmental Systems Sciences Division of the Office of Biological and Environmental Research in the U.S. DOE Office of Science. Computational resources were provided by the National Energy Research Scientific Computing Center. We acknowledge the World Climate Research Programme's Working Group on Coupled Modelling, which is responsible for CMIP. We thank the climate modelling groups listed in Table 1 for producing and making available their model output. For CMIP, the U.S. DOE 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.