We consider the three following subregions: West Sahel (10–20∘ N 18–10∘ W), Central Sahel (10–20∘ N 10–10∘ E), and Guinea Coast (5–10∘ N 10–10∘ E), shown as black boxes in Fig. 1a. We focus on annual values over the period 2006–2099. Following previous studies (Froidurot and Diedhiou, 2017; Bichet and Diedhiou, 2018a, b), we define a wet (dry) day using the threshold of 1 mm d-1. We define a dry spell as a sequence of 2 or more consecutive dry days that are preceded and followed by a wet day. Hence, the duration of a dry spell, as defined in our study, spans from 2 to 365 d. We compute the annual precipitation intensity (INT), number of wet days (RR1), maximum length of consecutive dry days (CDDs), and maximum length of consecutive wet days (CWDs) following the definition of the Expert Team of Climate Change Detection and Indices (ETCCDI; Zhang et al., 2011). Note that because INT corresponds to the precipitation averaged over wet days, a change in the INT value directly translates into a change in the intensity of wet events, regardless the number of wet events. In addition, we compute the annual contribution of very heavy rain (C98) following Eq. (1): 1C98=PRCPTOT98PRCPTOT, where PRCPTOT98 is the sum of daily precipitation greater than or equal to the 98th percentile annual value on wet days (Pctl98), and PRCPTOT is the sum of daily precipitation on wet days during that year. Following previous studies (Giorgi et al., 2011; Bichet and Diedhiou, 2018a, b), we compute the annual average duration of dry spells (DSLs) following Eq. (2): 2DSL=NDDNDS, where NDD is the annual number of dry days excluding isolated dry days (single dry day preceded and followed by a wet day), and NDS is the total number of dry spells during that year. Note that the annual number of dry days is directly derived from the annual number of wet days (RR1). Based on previous studies (Giorgi et al., 2011; Mohan and Rajeevan, 2017), we compute the annual hydroclimatic intensity index (HY-INT) following Eq. (3): 3HY-INT=DSLnINTn, where DSLn and INTn are the normalized DSL and INT, respectively. The normalization consists, for each grid point, in dividing the annual time series of INT(DSL) by its mean value over the period 2006–2099.

We use an ensemble of 18 high-resolution regional climate projections taken from the most up-to-date ensemble produced in the recent years for Africa: CORDEX-Africa (Giorgi et al., 2009; Jones et al., 2011; Hewitson et al., 2012; Kim et al., 2014). All simulations available online at the time of the analysis have been used. In this ensemble, 5 regional climate models (RCMs) are used to downscale 10 global climate models (GCMs) under the climate scenario RCP8.5 (Table 1). Out of the 50 combinations possible, only 18 were available, from which 8 use the same RCM. Whereas this imbalance presents the disadvantage to slightly bias the results towards this RCM, it also presents the advantage of representing a large number of GCMs, not accessible otherwise. Because the impact of the heterogeneity of the CORDEX-Africa GCM–RCM matrix on future precipitation changes is found mostly over central and West Africa (Dosio et al., 2019), we choose to represent a maximum diversity of RCMs and GCMs. Furthermore, although averaging model output may lead to a loss of signal (such that the true expected change is very likely to be larger than suggested by a model average), there is too little agreement on metrics to separate “good” and “bad” models to objectively weight the models (Knutti et al., 2010). In the following, we thus use the equal-weighted model average to illustrate the mean response of our ensemble (multimodel mean maps in Figs. 1–2) and show the individual responses of each simulation using scatterplots (Figs. 3–5). The simulations span the period 2006–2099 at a daily mean time step and cover Africa (24.64∘ W–60.28∘ E; 45.76∘ S–42.24∘ N) at about 50 km (∼0.5∘) spatial resolution in latitude and longitude. For each simulation and each grid cell, daily time series of surface air temperature and precipitation are retrieved.

We evaluate the spatial distribution of daily mean rainfall simulated in the CORDEX models by comparing it to the satellite Climate Hazards Group Infrared Precipitation with Station data (CHIRPS) rainfall dataset (Fig. S1 in the Supplement). The CHIRPS dataset is explicitly designed for monitoring agricultural drought and global environmental change over land. It corresponds to a gridded, quasi-global (50∘ S–50∘ N), high-resolution (0.05∘), daily rainfall dataset that covers the time period 1981–2014 (Funk et al., 2015). In addition, we evaluate the statistical distribution of daily mean and yearly mean rainfall simulated in the CORDEX models from three cities (Ouagadougou Airport (12.35∘ N, 1.52∘ W), Dakar-Yoff (14.72∘ N, 17.51∘ W), and Accra Kotoka International Airport (5.6∘ N, 0.17∘ W), as seen in Fig. 1a), by comparing it to the CHIRPS dataset (grid point comparison) and to three near-surface daily rain gauge data extracted from the BADOPLUS dataset, as described in Panthou et al. (2012). Results are shown in Figs. S1 and S2 and summarized here. Each of the 18 CORDEX simulations satisfactorily reproduces the spatial distribution of the CHIRPS daily mean precipitation and to some extent RR1 (slightly overestimated in CORDEX, especially along the Guinea Coast; Fig. S1). More disagreements are found across models for INT, even though the multimodel mean is in good agreement with the CHIRPS dataset (Fig. S1). In Dakar, the CHIRPS and the BADOPLUS datasets show similar statistical distribution of daily mean and yearly mean precipitation, as well as RR1, and to some extent INT (Fig. S2). In Ouagadougou and Accra however, CHIRPS and the BADOPLUS datasets show similar statistical distributions for daily mean and yearly mean precipitation but strongly disagree for RR1 (overestimated in the CHIRPS dataset) and INT (underestimated in the CHIRPS dataset; Fig. S2). In other words, the CHIRPS dataset shows more frequent but less intense precipitation events than the BADOPLUS dataset in Ouagadougou and Accra. Each of the 18 CORDEX simulations generally shows a similar statistical distribution with the observations for daily mean and yearly mean precipitation, but they tend to overestimate (underestimate) the observed RR1 in Accra and Ouagadougou (Dakar) and underestimate the observed (particularly as compared to the BADOPLUS dataset) INT in all three locations (Fig. S2). In other words, the CORDEX simulations show less intense precipitation events compared to observations (especially the BADOPLUS dataset) in all three locations and less (more) frequent precipitation events in Dakar (Ouagadougou and Accra) than the observations. Nevertheless, we find that the observations are always included within the range of the 18 CORDEX simulations. Hence, we conclude that the CORDEX simulations compare satisfactorily well with the observations and can be used for the purpose of our study.

Trends (2006–2100) of air surface temperature (∘C per decade), mean precipitation (mm d-1 per decade), INT (mm d-1 per decade), RR1 (days per decade), CDD (days), and CWD (days) are shown in Fig. 1, as multimodel mean maps. From Fig. 1, temperature is expected to increase on average by +0.5 ∘C per decade over the entire region, with a northward increase that reaches +0.7 ∘C per decade over the northern Sahel. More specifically, temperature is expected to increase on average by +0.5, +0.6, and +0.45 ∘C per decade over West Sahel, Central Sahel, and the Guinea Coast, respectively (Fig. 1a). Mean precipitation is expected to increase on average by +0.03 mm d-1 per decade over the Guinea Coast and decrease on average by -0.015 and -0.001 mm d-1 per decade over West Sahel and Central Sahel, respectively (Fig. 1b). INT is expected to increase on average by +0.2 mm d-1 per decade over all of West Africa, reaching up to +0.3 mm d-1 over the Guinea Coast (Fig. 1c), and RR1 is expected to decrease on average by -1.5 d per decade over all of West Africa, reaching up to -3 d per decade over West Sahel (Fig. 1d). Furthermore, CDD is expected to increase on average by +1 d per decade over the entire region, with a northward increase that reaches +5 d per decade over the northern Sahel (Fig. 1e), and CWD is expected to decrease on average by -0.5 d per decade over the Guinea Coast, and more specifically over Guinea Bissau, Guinea, Sierra Leone, Liberia, Ghana, Benin, Togo, and Nigeria (Fig. 1f). Hence, our results show that by the end of the century, precipitation is expected to intensify but rarefy (including longer dry periods and shorter wet periods) over all of West Africa, with different impacts on mean precipitation depending on the subregion: decrease in mean precipitation over the Sahel (especially over West Sahel) and increase in mean precipitation over the Guinea Coast.

Trends (2006–2100) of INTn, DSLn, and HY-INT are shown in Fig. 2, as multimodel mean maps. In agreement with Fig. 1c, Fig. 2 shows an intensification of precipitation over all of West Africa, on average by +0.02 (2 %) per decade (Fig. 2a). In addition, whereas DSL is expected to increase over the Sahel on average by +0.05 (5 %) per decade, with a latitudinal increase northward that reaches +0.1 (10 %) per decade over the western part of the Sahel, negligible changes are expected over the Guinea Coast (Fig. 2b). As a result, HY-INT is expected to increase on average by +0.05 (5 %) over West Africa, with a latitudinal increase northward that reaches +1.5 (15 %) per decade over the Sahel (Fig. 2c). Therefore, our results show that the hydrological cycle is expected to intensify by the end of the century over all of West Africa. Over the Guinea Coast, this intensification results exclusively from an increase in precipitation intensity (INT). Over the Sahel however, it results from both an increase in precipitation intensity (INT) and an increase in the average length of dry spells (DSL). Hence over the Sahel, rainfall events are expected to become more intense and separated by much longer periods of dryness.

Annual values (2006–2009) of mean precipitation (mm d-1), RR1 (days), and INT (mm d-1) are shown in Fig. 3, plotted against the corresponding annual mean values of air surface temperature (∘C), as averaged over (a) West Sahel, (b) Central Sahel, and (c) Guinea Coast. Individual CORDEX simulations are represented by the thin colored dots (see Table 1 for color references), and the multimodel mean is represented by the thick black dots. According to Fig. 3, mean precipitation decreases with temperature over the Sahel (multimodel mean decreases by -0.032 and -0.012 mm d-1 ∘C-1 over West Sahel and Central Sahel, respectively) and increases with temperature over the Guinea Coast (multimodel mean increases by +0.062 mm d-1 ∘C-1). In all three subregions, RR1 decreases with temperature (multimodel mean decreases by -2.3, -1.7 and -2.6 d ∘C-1 over West Sahel, Central Sahel and Guinea Coast, respectively) while INT increases with temperature (multimodel mean increases by +0.33, +0.27 and +0.41 mm d-1 ∘C-1, respectively). Based on the multimodel mean values, we show that changes in temperature explain less than 24 % of the changes in mean precipitation in all three subregions but more than 67 % (51 %) of the changes in RR1 (INT). As seen in Fig. 3, even though the annual mean values vary greatly from a simulation to another (e.g., NCC-NorESM1-HIRHAM5 is particularly warm over Central Sahel, ICHEC-RACMO is particularly cold over the three subregions, NCC-NorESM1-HIRHAM5 is particularly wet over the Guinea Coast, and CSIRO-SMHI is particularly wet over the Sahel), the relation between each variable and the temperature is consistent across all models, albeit with a different strength.

Annual values (2006–2009) of INTn, DSLn, and HY-INT are shown in Fig. 4, plotted against the corresponding annual mean values of air surface temperature (∘C), as averaged over (a) West Sahel, (b) Central Sahel, and (c) Guinea Coast. As for Fig. 3, individual CORDEX simulations are represented by the thin colored dots and the multimodel mean by the thick black dots. From Fig. 4, INTn increases with temperature in all three regions (multimodel mean increases by +0.031 (3.1 %) ∘C-1 over the Sahel and +0.041 (4.1 %) ∘C-1 over the Guinea Coast), whereas DSLn increases with temperature over the Sahel (multimodel mean increases by +0.093 (9.3 %) and +0.056 (5.6 %) ∘C-1 over West Sahel, and Central Sahel, respectively; Fig. 4a and b), but it decreases with temperature over the Guinea Coast (multimodel mean decreases by -0.01 (1 %) ∘C-1; Fig. 4c). As a result, HY-INT increases with temperature in all three subregions (multimodel mean increases by +0.13 (13 %), +0.089 (8.9 %), and +0.031 (3.1 %) ∘C-1 over West Sahel, Central Sahel, and the Guinea Coast, respectively). Based on the multimodel mean values, we show that changes in temperature explain about 63 % (74 %) of the changes in DSLn (HY-INT) over the Sahel and 14 % (27 %) over the Guinea Coast. As shown in Fig. 4, the annual mean values of DSLn vary greatly from a simulation to another over the Sahel but are similar across simulations over the Guinea Coast. Similar to Fig. 3, we find that the relation between each variable and the temperature is consistent across all models, albeit with different strengths. We conclude that most of the trends observed in Figs. 1 and 2 show a positive relationship with regional warming.