Changes in climate extremes over West and Central Africa at 1.5 °C and 2 °C global warming

Numerous approaches have recently been developed for identifying regional climate signals associated with specific global warming targets (James et al 2017). We here apply a hybrid approach that cannot be categorized into the four main techniques described in James et al 2017. We refer to it as the empirical sampling approach, which has been applied in Seneviratne et al 2016 and described in more detail in Wartenburger et al 2017. In contrast to pattern scaling (e.g. Tebaldi and Arblaster 2014, Lopez et al 2014), this approach does not need a priori assumptions about the nature of the dependency on global temperature (or other climate variables; e.g. Frieler et al 2012, Lynch et al 2016, Kravitz et al 2017) such as the linearity of the response (e.g. Fischer et al 2014). For predefined regions, this method derives the empirical dependency of changes in regional quantities as a function of global mean temperature changes based on a range of climate model projections (Herger et al 2015, Seneviratne et al 2016). The scatter plots of the scaling relationship between changes in global mean temperature (ΔTg) and in regional climate indices are computed as follows (see further detail below, in the description of the plots). Annual global mean temperatures are derived from each ensemble member and individual years before performing the 20 year moving average that we assign to the median year and then we compute the anomalies relative to the pre-industrial reference period 1861–1880. Over the land points, we do the same computation for the temperature and precipitation extreme indices averaged regionally. We will also assess the significance of this dependency relationship, against a null hypothesis of no relationship. To test this, we fit an ordinary least squares between changes in global temperature and in climate indices averaged regionally over land points, applied for each individual model realization for ΔTg >= 1 °C. The number of models for which the dependency relationship approximated by a linear regression is significantly different from zero (p = 0.01) will be shown in the discussion section to give an idea of robustness (Benjamini and Hochberg 1995).

The advantage of this approach is to provide in a single figure information on the response of a given regional quantity for different global mean temperature, on empirical assessment of this relationship (allowing e.g. to identify possible non-linearities), and on the range of inter-model and inter-scenario uncertainties. Box plots have been computed to capture the uncertainties between the models.

The regional-scale dependencies between global mean temperature and a range of climate change indices (consisting of 11 indices from the joint CCl/CLIVAR/JCOMM Expert Team on Climate Change Detection and Indices (ETCCDI; Sillmann et al 2013a, 2013b), plus three additional water cycle indices computed by Wartenburger et al 2017, see table 1) are derived from global climate model (GCM) simulations performed within the frame of the Coupled Model Intercomparison Project Phase 5 (CMIP5, Taylor et al 2012). The subset of the CMIP5 models used in this study is identical to the model subset used in Wartenburger et al 2017 (table 1 therein). The regional changes are analyzed at the spatial scale typical of the models. To assess the impact of intra-model spread, we use only one ensemble member per model (r1i1p1). To assign the same weight to all simulations, all GCM output has been bi-linearly interpolated to a horizontal resolution of 2.5° × 2.5° prior to performing the empirical sampling analysis. Wartenburger et al (2017) have studied the natural variability in investigating the spread among all ensemble members of a particular model. They found that the model spread has essentially no impact on the ensemble mean or median, and even the spread is almost similar. Rowell (2012) showed that the role of natural variability is small in this region.

We consider for this study the yearly anomalies (relative to the pre-industrial reference period 1861–1880) of global mean temperatures ΔTg and the respective anomalies of the aforementioned indices computed for all emission scenario from each ensemble member and for all models. This reference period has also been used in previous publications that apply the scaling approach (Wartenburger et al 2017 and Seneviratne et al 2016). To be compatible with those studies (and to enhance comparability), it has been decided to stick to the same reference period here.

All indices and the global mean temperatures are smoothed by applying a 20 year moving average over land points of the West Africa region (WAF; Giorgi and Francisco 2000, hereafter referred to as the WAF domain) and over land points of smaller-scale regions located within the WAF domain (figure 1): western Sahel (WSA): 10°N–20°N/−18°W–10°W; central Sahel (CSA): 10°N–20°N/10°W–10°E; eastern Sahel (ESA): 10°N–20°N/10°E–30°E; Guinea Coast (GC): 5°N–10°N/10°W–10°E; Central Africa (including Congo Basin; CAF): 7.5°S–10°N/10°E–30°E.

Zoom In Zoom Out Reset image size

The limitation with the WAF domain is that it mixes several climatic and ecological zones. Also, this region is facing a multitude of issues in terms of adaptation and mitigation (Redelsperger et al 2006). These regions have different landscapes and surface conditions (arid and semi-arid areas with savanna and trees over Sahel (WSA, CSA and ESA), forest over GC and equatorial forest in CAF with complex interactions with the West African monsoon. Due to substantial differences in their location relative to the tropical Atlantic and to the Hadley cell, those regions are characterized by very diverse hydrometeorological regimes. The mean annual precipitation in these regions is 575 mm in WSA, 450 mm in CSA, 347 mm in ESA, 1250 mm in GC and 2000 mm in CAF. Besides, the seasonality of precipitation in these regions is different with two rainy seasons over GC, while the Sahel (WSA, CSA and ESA) is characterized by only one wet season (Sultan et al 2013).

The plots below are computed from the CMIP5 simulations of one ensemble member r1i1p1 of all models available for 1961–2099 for each scenario (26 models for the scenarios 'Historical' and 'RCP8.5', 16 for RCP6.0, 23 for RCP4.5 and 19 for RCP2.6).

The scaling plots show the changes in climate indices in the WAF domain and WAF subregions with global mean temperature. Each color represents the multimodel mean of the corresponding scenario. The gray shading represents the minimum and maximum values simulated by the 110 (26+26+16+23+19) projections.

The boxplots represent the distribution of the 84 individual ensemble members for RCP8.5 (26 models), RCP6.0 (16 models), RCP4.5 (23 models) and RCP2.6 (19 models). For these plots, the scaling plots are the 20 year averages of the indices for which the respective 20 year average global mean temperature exceeds (or reaches) 1.5 °C (2 °C) for the first time.

In each boxplot, the central mark indicates the median value. The bottom and top edges of the box indicate the first and third quartile, respectively. The whiskers extend to the most extreme data points not considered outliers (that is any value comprised between the first (third) quartiles and 1.5^⁎ the interquartile range). The outliers (that is any value lower (higher) than 1.5^⁎ the interquartile range) are plotted individually using the ∣'+'∣ symbol.

The verification of the scaling of the temperature values is done in comparing the modeled (26 CMIP5 historical simulations, one member per model) scaling with observation-based (GSWP3) scaling for the period 1920–2010 (using a reference period of 1961–1990). To test our results against independent data, we have applied the same methodology to compute ETCCDI indices derived from the 3rd phase of the Global Soil Wetness Project (GSWP3, http://hydro.iis.u-tokyo.ac.jp/GSWP3/) forcing data set, which is the successor to GSWP2 (Dirmeyer et al 2006). We have opted for this recently developed data set, as it has gained high popularity in the modeling community and provides high-resolution (daily) estimates of climate variables back to 1900. Due to its limited temporal coverage, we use the period 1961–1990 as a reference for this comparison.

Figure 2 shows the verification of the scaling of temperature values from the CMIP5 models (HIST, in black) and observations (GSWP3, in red) for the ETCCDI indices (top to bottom, respectively) ΔTAS (hereafter ΔT as mean air temperature at 2 m), ΔDTR, ΔTX90p, ΔWSDI against global temperature ΔT_g averaged across the subregions: WSA, CSA, ESA, GC and CAF. Both models and observations show an increase in temperature, but the warming is around 1 °C in HIST and 2 °C in GSWP3 confirming that the models likely underestimate the temperature rise (Sherwood et al 2014). Except over CAF, ΔDTR decreases with temperature rising both in the models and in the observations. In all the subregions, ΔTX90P increases with temperature both in the models and in the observations. ΔWSDI increases with temperature rising mainly over the Sahel band (WSA, CSA and ESA) both in the models and observations, but over GC and CAF the models underestimate the magnitude and rate of change with temperature.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Figure 3 shows the scaling plot of ΔTg and changes in temperature (ΔT) for the WAF domain and the five subregions. The box plots show the responses at a global warming of 1.5 °C (hereafter ΔTg1.5C) and of 2 °C (hereafter ΔTg2C). The linear increase of regional temperatures with global mean temperature is confirmed whatever the size of the domain, with the regional projections for a specific global temperature increase marked by a relatively narrow ensemble spread (Engelbrecht et al 2015). Note, however, that if one takes the perspective of limiting carbon emissions, rather than specifying a temperature target, then modeling uncertainty is substantially enhanced (see the analysis of Rowell et al 2016 for Africa). Nevertheless, from either perspective, larger regional warming than global warming is highly likely. For the WAF region, a global warming of 1.5 °C (2 °C) induces a multi-model mean increase of 1.7 °C (2.3 °C). This warming is more intense in the driest regions of Sahel (CSA and ESA) where a ΔTg1.5C (ΔTg2C) is associated with a warming of 2 °C (2.7 °C) over CSA and of 1.9 °C (2.6 °C) over ESA. Over WSA, GC and CAF, the rates are weaker. We also note that the annual maximum of daily maximum temperature (ΔTXX, not shown) is projected to increase more rapidly than the global mean temperature with a particularly large magnitude over the Sahel. The annual minimum of daily minimum temperature (ΔTNN, not shown) in all subregions is projected to increase with ΔTg at even higher magnitudes than ΔTXX.

Figure 4 shows the box plots of the responses of the diurnal temperature range (ΔDTR, top), percentage of very hot days (ΔTX90P, middle), and of the warm spell duration index (ΔWSDI, bottom) to a ΔTg1.5C and ΔTg2C in the WAF subregions. The changes of the diurnal temperature range in all the WAF subregions present a disparity in the sign with a large spread either side of zero. However, at ΔTg1.5C, most models (62.5% of models in WSA, 95% in CSA, 100% in ESA, 93.8 in GC and 82.3% in CAF) projected a decrease ΔDTR with increase of ΔTg. We verified that this trend is mainly associated with the projected increase of the daily minimum temperature in most models (ΔTNN, not shown). The least pronounced changes are found over WSA, while the magnitudes in the other regions are similarly moderate.

There is a large spread across models for changes in above-average days, but in general the warmest areas of Sahel are projected to have the weakest increase of very hot days, while the wetter regions of GC and CAF are projected to experience the largest changes in ΔTX90p. The trend of the warm spell duration index (ΔWSDI) is similar to the trend of ΔTX90p with a large spread across the ensemble models. GC and CAF are projected in most of the models to have the longest warm spells. We note also that the cold spell duration index (ΔCSDI, not shown) is projected to decrease with increase of global temperature until a threshold of around ΔTg = 2 °C and remains constant thereafter. However, there is a large ensemble spread that is weaker in the drier subregions of Sahel and very large in the wetter areas of the domain.

As illustrated in figures 3 and 4, a 2 °C target for ΔTg implies increases in warm temperature extremes exceeding 2 °C in the WAF subregions. This may be due to the changes in the West African monsoon dynamic associated with land–sea contrast (Sutton et al 2007) in response to radiative forcing, or to feedbacks from decreases in soil moisture (Serreze and Barry 2011, Seneviratne et al 2013), which may further amplify changes in extreme temperatures in some of the regions (Seneviratne et al 2014). Higher temperatures expected in ΔTg1.5C and ΔTg2C will cause crops to grow faster, with thus less time for photosynthesis and then less biomass built per year. When occurring at specific development stages of most grain crops, very high temperature may severely reduce or even completely destroy the harvest of grain crops (Rosenzweig et al 2014, Varhammar et al 2015, Ahmed et al 2015). The increase of the frequency and intensity of warm spells will have serious implications for agriculture and human health (Weiss et al 2014, D'amato et al 2015, Partanen et al 2018).

Zoom In Zoom Out Reset image size

Figure 5 shows the scaling plot of ΔTg and changes in total precipitation accumulated during wet days (ΔPRCPTOT) for the WAF domain and the five subregions. The box plots show the responses at ΔTg1.5C and ΔTg2C. Although anomalies, and hence inter-model spread, tend to be slightly larger at the larger warming target, it has previously been shown that at a specific time point each model's global warming has little impact on its regional precipitation change (Rowell 2012) in contrast to the strong impact seen above on regional temperature change. Hence, the model spread amongst wet day projections is large. Furthermore, more heterogeneous trends emerge when investigating changes over smaller scales, and so the spread is larger over these regions than for the WAF domain. Projections in the Sahel have a large spread. However, as shown in table 2, in most of the models, precipitation is projected to increase in the CSA and ESA (respectively for 63.75% of the models and for 66.25% at ΔTg1.5C), while a tendency of decrease of rainfall is noted over WSA (81.25% at ΔTg1.5C). This is in agreement with Diallo et al (2012), who suggested that drought conditions over WSA could likely be related to a delay in the onset of the wet season accompanied by an overall shortening of this period. Small changes are noted over the humid tropical and subtropical regions of GC and over CAF which have the weakest ensemble spread.

Zoom In Zoom Out Reset image size

Figure 6 shows the box plots of the responses of the Standard Precipitation Evapotranspiration Index (ΔSPEI12) of annual total precipitation explained by very heavy precipitation greater that the 99 percentile (ΔR99PTOT), of maximum length of consecutive wet days (ΔCWD), and of maximum length of consecutive dry days (ΔCDD) at ΔTg1.5C and ΔTg2C. Mean changes in ΔSPEI12 at ΔTg1.5C and ΔTg2C are small and projections have a large spread in all the subregions. However, table 2 shows that in most of the models ΔSPEI12 is projected to increase with global temperature over ESA and CAF, while over WSA it is projected to decrease (respectively, at ΔTg1.5C for 62.19% of the models in ESA, for 58.5% in CAF and for 86.6% in WSA). This is in line with precipitation changes. Changes in ΔR99PTOT at ΔTg1.5C and ΔTg2C over Sahel (WSA, CSA and ESA) have a large spread among the models. However, it is computed in table 2 that most of the models projected an increase of the contribution of heavy rainfall in CSA and ESA (respectively, 71.2% and 73.7% of the models at ΔTg1.5C). The highest positive changes tend to be projected to be in GC and CAF (respectively, for 77.5% of the models and 87.5% at ΔTg1.5C), but these subregions are where the ensemble spread is the largest. Mean changes in ΔCWD at ΔTg1.5C and ΔTg2C are not pronounced over Sahel (WSA, CSA and ESA), but as shown in table 2, in most of the models ΔCWD is projected to decrease (88.8% of the models over WSA, 82.5% over CSA and 63.8% over ESA). The highest ensemble spread is found over GC and CAF where most of the models (85% in GC and 74.1% in CAF) projected a decrease of ΔCWD. There are uncertainties in the projections of the ΔCDD. However, ΔCDD is often projected to have the highest changes in the WSA with most of the models (80% of the models at ΔTg1.5C) projecting an increase of the maximum length of dry spells with ΔTg. Projections of ΔCDD in the CSA and ESA are not pronounced and there is a large ensemble spread, in contrast to projections in GC and CAF where the responses are not pronounced, but where the ensemble spread is the weakest.

In summary, while there are several uncertainties represented by the large ensemble spread of the models, most of models projected a decrease of total precipitation (81.25% of the models) over WSA, an increase of the length of dry spells (80% of the models) and a decrease of SPEI12 (86.6% of the models). Over central Sahel, eastern Sahel and CAF, most of the models (between 70% and 85% of the models) projected a regime of more heavy rainfall and less frequent rainy days for both temperature targets, in agreement with Sylla et al (2015) and Diallo et al (2016). The wetter subregions of GC and CAF are the areas where the ensemble spreads of projected precipitation related indices are the highest. However, most of the models projected a small change in precipitation (PRCPTOT) with a decrease of the maximum length of consecutive wet days (CWDs, between 75% and 85% of the models), of SPEI12 (60% in GC and 40% in CAF) and an increase in the contribution of heavy rainfall (between 77% and 87% of the models).