3.1 Spatial variability of duration and intensity-based indices
The spatial distribution of the annual total wet-day precipitation amount (PRCPTOT) during the 1985–2014 period over CA, derived from the observations CHIRPS, ARC2 and TAMSAT, as well as the individual CMIP6 simulations and their MME is displayed in Fig. 2. The PRCPTOT represents the total precipitation accumulated over a given area during a specific period, and its spatial distribution provides information about the overall amount of precipitation. The distribution of PRCPTOT computed through CHIRPS datasets shows the rain-belt concentrated around the Equator (i.e., between 5ºS and 5ºN), with peaks of up to 3000 mm located along the coastal countries and over eastern DRC (Fig. 2a). The PRCPTOT presented by ARC2 and TAMSAT has similar patterns as that of CHIRPS, but with peaks more widespread in the case of TAMSAT (Fig. 2b-c). The RMSE/SD/PCC between CHIRPS and other observations (i.e., ARC2 and TAMSAT) is always between 0 and 0.5/0.75 and 1.25/0.9 and 1 (see supplementary material, Fig. S1a; PRCPTOT), indicating a good level of agreement between observations for the representation of the total wet-day precipitation amount. Generally, the MME as well as the individual CMIP6 models succeed in capturing the main features of the PRCPTOT’s patterns, with RMSE/SD/PCC always between 0 and 1/0.75 and 1.25/0.8 and 1 (see supplementary material, Fig. S1a; PRCPTOT), except for ACCESS-CM2, BCC-CSM2-MR, CanESM5, E3SM-2-0, MIROC6 and MRI-ESM2-0 which have a slightly high variability. Despite the relatively good performance of the MME as well as its individual members in simulating the total precipitation amount on rainy days, their magnitude nevertheless remains overestimated compared to CHIRPS. The overestimation shown by CMIP6 models in simulating the patterns of the total wet-day precipitation amount over CA could be attributed to an excessive representation of both divergent and rotational components of the total wind, which in consequence would lead to a poor simulation of the strength of the low-level westerlies (LLWs) as discussed in Taguela et al. (2022).
Figures 3 and S2 of the supplementary material, show the spatial distribution of the annual wet-day intensity (SDII) and their frequency (RR1), respectively, during the 1985–2014 period over CA, derived from the observations CHIRPS, ARC2 and TAMSAT, as well as the individual CMIP6 simulations and their MME. Firstly, the SDII represents the average amount of precipitation per rainy day and provides insights into the intensity of precipitation events. On the other hand, the RR1 represents the number of days with precipitation exceeding a given threshold (e.g, 1 mm in this study) and provides information on the frequency of precipitation events. Accordingly, the spatial distribution of SDII and RR1 shown by CHIRPS enables the identification of coastal and mountainous areas (see Fig. 1) as regions with higher intensities and frequencies of precipitation events, respectively and, therefore, potentially susceptible to experience more intense and frequently localized extreme weather events. Consequently, the spatial patterns of SDII shown by CHIRPS are generally greater than 3 mm/day over the study area, with peaks of up to 13–14 mm/day which coincide throughout the year with the positions of the maximum PRCPTOT recorded in Fig. 2a. Note that there is some disagreement between CHIRPS and the other observations (i.e., ARC2 and TAMSAT), especially on the less (more) intense events over the northern (central) part of the domain in the case of ARC2 (TAMSAT). Furthermore, the observed spatial patterns of RR1 are similar between CHIRPS, ARC2 and TAMSAT (see, Fig. S2a in the supplementary material). TAMSAT (CHIRPS) appears more intense (frequent) than the other observation products over CA. This behaviour could be associated with their higher spatial resolutions, and therefore their high variability as shown through the TAMSAT data which is located either outside the quadrant (in the case of SDII) or before SD = 0.75 (in the case of RR1) in the Taylor diagram (see supplementary material, Fig. S1a; SDII and RR1). Compared to the observations, the MME and the individual CMIP6 models are able to reproduce the overall structure of the patterns of both SDII and RR1, but in certain cases simulate less (more) intense (frequent) precipitation events along the rain-belt. For instance, by producing events that are less intense and more frequent than observations (also see supplementary material, Fig. S1a; SDII and RR1), the MME signal tends to neutralise the contributions of these two indices in the result of the total wet-day precipitation amount shown in Fig. 2.
The spatial distributions of the annual maximum consecutive dry day (CDD) and maximum consecutive wet day CWD, respectively, during the 1985–2014 period over CA, derived from the observations CHIRPS, ARC2 and TAMSAT, as well as the individual CMIP6 simulations and their MME are shown in Figs. 4 and S3 of the supplementary material. CDD and CWD represent the duration of consecutive dry and wet periods, respectively and provide information on the presence of prolonged dry and wet spells. Concerning the variability of CDD, all observation datasets exhibit dry spells of up to 180 days over the northern and southern parts of the domain. For instance, the central part of the study area has fewer numbers of dry spells (< 20 days). As expected, the patterns of the lowest CDD are located in areas with higher PRCPTOT (see, Figs. 2a-c). Regarding the MME and the individual CMIP6 models, they are able to realistically reproduce the CDD fields, with almost all models recording a RMSE/SD/PCC between 0 and 1/0.75 and 1.25/0.8 and 1 (see supplementary material, Fig. S1a; CDD). On the other hand, the MME and the individual CMIP6 models generally overestimate the CWD magnitude along the Atlantic coast and over countries located over the central part of the domain, showing wide areas experiencing more than 50 consecutive precipitation days. This result thus highlights the poor performance of CMIP6 models in simulating wet spell patterns over CA, as also shown by the data from the models which are almost all located outside the quadrant in the Taylor diagram (see supplementary material, Fig. S1a; CWD). Previous studies based on either GCMs (e.g., Sonkoué et al., 2019; Akinsanola et al., 2021; Klutse et al., 2021; Faye and Akinsanola, 2022; Agyekum et al., 2022) or regional climate models (RCMs; e.g., Klutse et al., 2016; Fotso-Nguemo et al., 2019; Fotso-Kamga et al., 2020) have also revealed an overestimation of wet spell over specific regions of Africa. This shortcoming could be attributed, among others to: i) the complexity of interactions between the ocean, atmosphere and vegetation, which are still poorly represented in climate models (Patricola and Cook, 2010); ii) the spatial resolution, which is still too low compared with the scale of the mesoscale processes that control precipitation variability (Roehrig et al., 2013); and iii) the lack of reliable historical precipitation data over sub-Saharan Africa, which limits the quality of climate models evaluation (Oluwagbemi et al., 2022).
3.2 Spatial variability of intense precipitation indices
Figures 5 and S4 of the supplementary material show the spatial distribution of the annual maximum consecutive 5-day precipitation amount (RX5DAY) and maximum consecutive 1-day precipitation amount (RX1DAY), respectively, during the 1985–2014 period over CA, derived from the observations CHIRPS, ARC2 and TAMSAT, as well as the individual CMIP6 simulations and their MME. RX5DAY and RX1DAY provide information about the intensity and duration of extended precipitation events. The spatial distribution of RX5DAY generally shows similar patterns to that of RX1DAY, with elevated values of up to 180 mm and 140 mm recorded along the Atlantic coast for RX5DAY and RX1DAY, respectively. Despite the deficiencies exhibited by the individual CMIP6 models in representing RX5DAY and RX1DAY patterns compared to observations (most pronounced for BCC-CSM2-MR, CanESM5, GFDL-ESM4 and MIROC6), the MME succeeds to satisfactory representing the overall structure of the patterns of these two indices, with RMSE/SD/PCC always between 0.5 and 1/0.75 and 1.25/0.6 and 0.7 (see supplementary material, Fig. S1b; RX5DAY and RX1DAY). This suggests that the MME succeeds to capture reasonably well the broad-scale patterns of prolonged extreme precipitation events over CA. Previous studies conducted over Africa have also highlighted the relatively good performance of GCMs and RCMs in simulating prolonged extreme precipitation during all seasons of the year (e.g., Sonkoué et al. 2019; Fotso-Nguemo et al., 2019; Fotso-Kamga et al., 2020; Faye and Akinsanola, 2022). This result emphasized the importance of climate models to accurately simulate RX5DAY and RX1DAY in order to assess the potential impacts of prolonged heavy precipitation on water resources management and ecosystem dynamics.
The spatial distribution of the annual 95th percentile of precipitation patterns (R95), during the 1985–2014 period over CA, derived from the observations CHIRPS, ARC2 and TAMSAT, as well as the individual CMIP6 simulations and their MME is presented in Fig. 6. The R95 represents the contribution of heavy precipitation events to the total precipitation amount and can help to localize regions with a high proportion of heavy precipitation events. Generally, the spatial patterns of R95 follow those of SDII (see, Fig. 3) with maximum values of ARC2 and TAMSAT (up to 50 mm/day) consistently greater than those of CHIRPS (up to 30 mm/day). The high values of R95 detected by ARC2 and TAMSAT could be explained by the reconstruction algorithms used by these 2 microwave satellite-based sources, which tend to generate higher estimates when there is no in-situ data, compared with CHIRPS. However, the MME, along with the individual CMIP6 models fail to satisfactory reproduce the overall structure of the R95 patterns, with RMSE > 1, SD > 1,25 and PCC < 0.6 (see supplementary material, Fig. S1b; R95), except for MPI-ESM1-2-LR which shows reasonable performances. Although the switch from CMIP5 to CMIP6 has allowed to better simulate the interactions of dynamical and thermodynamic processes in these models over CA (Taguela et al., 2022; Kuete et al., 2023), the fact nevertheless remains that in CMIP6 the contribution of the northern component of the African Esterly Jet (AEJ) remains overestimated as compared to its southern component (Kuete et al., 2023). This overestimation of the AEJ also could explain the simulation of excessive precipitation obtained for the majority of individual models considered in this study, especially in BCC-CSM2-MR, CanESM5, GFDL-ESM4 and MIROC6, which also simulate very heavy precipitation.
Figures 7 and S5 of the supplementary material show the spatial distribution of the annual total wet-day precipitation amount above the 95th percentile (R95PTOT) and total wet-day precipitation frequency above the 95th percentile (R95P), respectively, during the 1985–2014 period over CA, derived from the observations CHIRPS, ARC2 and TAMSAT, as well as the individual CMIP6 simulations and their MME. R95PTOT and R95P provide additional information on both the intensity and frequency of extreme precipitation events. Thus, R95PTOT represents the total precipitation amount associated with heavy precipitation events, while R95P represents the number of days with precipitation originating from these events. Concerning R95PTOT, its spatial pattern is very close to that of PRCPTOT but with maximum precipitation of about 400–600 mm recorded along the coastal area through all the observed datasets (i.e., CHIRPS, ARC2 and TAMSAT; see, Fig. 7a-c). Moreover, the patterns of R95P shown by observations also suggest that these heavy precipitation peaks are experienced around 14–16 times during the year (see, Fig. S5a-c). This implies that over CA, the recorded total wet-day precipitation amount is not induced by heavy precipitation events, but probably to the fact that rainy days are more frequent. Furthermore, the MME and the individual CMIP6 models are able to reproduce the observed spatial pattern field of R95PTOT and R95P, but with higher magnitude. For example, the results of the heavy precipitation simulated by BCC-CSM2-MR reach up to 1500 mm over most parts of the study domain (representing half of the precipitation received; see, Fig. 7f), and occurs up to 22 times a year (see, Fig. S5f)
3.3 Statistical evaluation of extreme precipitation indices over climatic subregions
The ability of CMIP6 models to accurately represent extreme precipitation indices is also assessed at the scale of the different CA climatic subregions, based on several metrics, including percentage BIAS, RMSE, PCC and TSS.
On the one hand, Fig. 8 shows the portrait diagrams of the ten extreme precipitation indices, calculated over the five subregions using the CHIRPS observations as reference data. Accordingly, the positive biases (see, Fig. 8; first column) are predominant in the subregions located in the southern part of the domain (i.e., EQW, EQE and SE), while the negative biases are more persistent in the subregions located in the northern part of the study area (i.e., SS and NE). It should be noted that for most CMIP6 models and their MMEs, the PRCPTOT, RR1 and RX1DAY indices are strongly overestimated over all subregions, except for NE, EQW and EQE where biases < 40% are found for PRCPTOT and RR1. Furthermore, the majority of CMIP6 models and their MMEs showed negative (positive) biases for SDII (R95PTOT and R95P) over SS and NE (EQW, EQE and SE), except INM-CM5-0, which showed a high positive (negative) bias. It is worth noting that most of the models show positive biases for PRCPTOT, RR1, RX1DAY, CDD, CWD, R95, R95PTOT and R95P, while SDII, RX5DAY and R95 have negative biases except for INM-CM5-0, IPSL-CM6A-LR. Nevertheless, for most of the considered precipitation indices, the MME bias is considerably lower for the majority of indices over all subregions, except for PRCPTOT, RR1 and RX1DAY where strong positive biases persist in subregions NE, EQW and SE. Regarding the RMSE (see, Fig. 8; second column), its values are generally low (i.e., < 1 mm/day) for all indices and in all subregions, except for RX1DAY which shows a large and constant error for the majority of CMIP6 models over NE, EQW and EQE. In this case, the highest RMSE values are recorded for the CanESM5, EC-Earth3 and IPSL-CM6A-LR models. It is important to note that other models also show large and consistent errors in almost all subregions, such as BCC-CSM2-MR for PRCPTOT and CWD; and INM-CM5-0 for RR1 and R95. Nevertheless, in most of the considered extreme precipitation indices, the RMSE of the MME is considerably lower (i.e., < 0.4 mm/day) for most of the indices with the exception of RX1DAY where high RMSE persist in subregions NE, EQW and EQE. For the PCC (see, Fig. 8; third column), most of the indices coupled to the models show values > 0.6 over all the subregions, which indicates a very good level of agreement with CHIRPS observations in the representation of extreme precipitation events over the CA domain. A low level of agreement (i.e., PCC < 0.3) is found in most models for RR1, SDII, CDD and CWD over EQE and for CWD and RX1DAY over SE. Overall, the representation of the majority of extreme precipitation indices by the MME is in agreement with CHIRPS observations over all CA climatic subregions, except for SDII, CDD, CWD and R95 (PRCPTOT, RR1, SDII, CWD and RX1DAY) over EQE (SE) where low PCC values are recorded. Previous studies undertaken over various parts of Africa (e.g., Ayugi et al., 2021; Akinsanola et al., 2021; Klutse et al., 2021; Faye and Akinsanola, 2022; Agyekum et al., 2022) have shown similar performance of CMIP6 models in the simulation of extreme precipitation indices. Moreover, these studies revealed that the performance of the CMIP6 model in accurately represent the daily precipitation characteristics strongly depends on the chosen reference data.
On the other hand, Figs. 9 and 10 present the results of the metric TSS for the ten analyzed extreme precipitation indices over the five climatic subregions, using CHIRPS observations as reference data. For the PRCPTOT and RR1 indices, INM-CM5-0 is the worst-performing model over almost all subregions, except in SS in the case of RR1, where a TSS of around 0.7 is recorded. In the case of these two indices, the TSS values are generally > 0.6, except in the case of RR1 where TSS values systematically < 0.4 are recorded in EQW. Although the performance of the CMIP6 models in simulating SDII is relatively poor in EQW, EQE and SE, with the exception of INM-CM5-0 the other models are sufficiently competent to capture the variability of SDII in SS and NE with TSS values generally > 0.6. For CDD, high TSS values are recorded for almost all CMIP6 models over all subregions, except in EQW and SE where TSS < 0.6 are recorded for CanESM5, GFDL-ESM4, INM-CM5-0, MRI-ESM2-0 and NorESM2-LM. It is worth noting that most models perform better in simulating CDD in SS, NE and EQE, with TSS close to 0.8 and above. Compared to the other indices, the TSS obtained for CWD by the majority of models is generally < 0.5 in all subregions, with the exception of BCC-ESM2-MR, E3SM-2-0 and HadGEM3-CG31-LL which have TSS slightly > 0.5 in SS. The TSS results for RX5DAY, RX1DAY, R95, R95PTOT and R95P are relatively similar in all subregions. For these indices, TSS values are generally > 0.6 in SS and NE, with the BCC-ESM2-MR and INM-CM5-0 model performing the worst in all subregions. Overall, the MME realistically simulates the majority of extreme precipitation indices in most CA climatic subregions, with the highest performance found in SS and NE (with a TSS greater than 0.6), apart from CWD where the TSS is around 0.4 for these two subregions.