Seasonality of precipitation
WRF captures the seasonal cycle, as can be seen from the Hovmöller diagrams (Fig. 3) for the WRF ensemble mean (ENS) in comparison to TRMM. For both, the area of maximum precipitation from April to June is situated at the coast at about 5°N.
According to Hagos and Cook (2007), the Saharan Heat Low (SHL) reaches its maximum by the end of June when it is positioned around 20°N, and when the Atlantic cold tongue in the Gulf of Guinea is established which induces a pressure gradient strong enough to trigger the “monsoon jump”. This term refers to the sudden relocation of the precipitation maximum from the coast to approximately 10°N and represents the monsoon onset in the Sahel. We define the date of the monsoon jump as the first occurrence of two consecutive days with rainfall amounts within the 0.9 percentile for the period May–July between 9 and 11°N. For TRMM, the monsoon jump takes place on 1st of July, as can be seen in Fig. 3 from the extension and subsequent relocation of the precipitation maximum from the coast to ~12°N. ENS is also capturing the monsoon jump, although 3 days earlier. Most ensemble members are able to capture the monsoon jump close to the observed date with a mean absolute deviation of 4.3 days and a maximum shift of 16–20 days for three of the members.
Intense precipitation events are better represented in ENS than in ERA-I, because of the higher horizontal resolution of WRF. The rainband is slightly shifted to the south over the whole rainy season in comparison to TRMM, but less than for ERA-I. The shift is especially pronounced in August, when the monsoonal rainfall is at its peak and TRMM shows precipitation throughout the whole month in the northern Sahel between 16–20°N, while ERA-I and ENS are not able to capture all of these events. This also applies to all individual ensemble members. The retreat of the rainband to the Guinea Coast sets in by mid-August for TRMM and ERA-I. This movement is delayed in ENS, which results in a too dry coast in the late summer. On the other hand, the observed dry period at the coast during August, i.e., before the retreat of the rainbelt, is not very well represented in ENS, since for some members the rainband remains too far south for the whole period. All ensemble members show a seasonal relocation of the rainband, but strongly differ in the extent, intensity and width, which will be discussed in the following sections.
Spatial distribution of precipitation
In order to reveal the differences in precipitation with respect to certain physics influences, Fig. 4c shows the bias against the ensemble mean of the spatial average of JAS precipitation for each of the nine parameterization groups (following the approach of Pohl et al. 2011, Fig. 15), where one particular scheme is fixed for each group. For example, the KF ensemble consists of the average of the nine simulations that utilize the KF cumulus scheme, and so forth.
The rainband of ENS shown in Fig. 4a is mostly too narrow with excessive precipitation in its core zone and in Central Africa in comparison to TRMM. A persistent dry bias in the eastern part of the Gulf of Guinea is introduced, which stretches into the continent (Fig. 4b). It is most likely caused by too low SSTs in this area in the NCDC dataset. Further simulations using prescribed SSTs from ERA-I show a reduction of this dry bias in accordance with higher SSTs in that region (not shown).
The intensities of the rainbands in Fig. 4c cover the entire range from dry conditions (ACM2, WSM3) to wet conditions (TH, KF). None of the groups is able to capture the exceptional northward extension of the rainband in 1999, especially visible over Mali, which leads to a dry bias in the northern Sahel. Since this dry bias is also found for ERA-I, one might assume that the bias of the WRF simulations is caused by the bias of the driving data. However, in ERA-I the reason for the dry bias is a shift of the relatively broad rainband to the south, while in WRF it is the small North–South extent of the rainband. For GF and MYJ, the rainband is especially narrow with a daily precipitation of only 1–2 mm at the coast. ACM2 and WSM3 show very low precipitation intensities and induce an overall dry bias. BMJ shows closest rainfall amounts to TRMM, outperforming ENS with more precipitation at the coast and less overprediction in the Sudano-Sahel.
Parameterization influences on rainband intensity and position
In the following, we analyze the contribution of the three parameterization groups (CU, MP, PBL) to the above-mentioned differences in spatial rainfall distribution. Figure 5 shows boxplots of the parameterization groups, compared to the ensemble mean for the whole rainy season. The spread of the boxes indicates the tendency of a scheme towards a dry or a wet regime. Small boxes imply that the scheme is the dominating factor, since the resulting regime is hardly changed by different configuration partners. We differentiate between (1) the Guinea Coast, where peak precipitation occurs during the continental phase before monsoon onset, (2) the Sudano-Sahel, where the center of the rainbelt and thus the precipitation maximum are found after monsoon onset is found, and (3) the Sahel, where precipitation depends on the northernmost extent of the rainband (Fig. 1). With respect to the blue line that indicates the TRMM mean, ENS underestimates precipitation both, at the Guinea Coast and in the Sahel. However, except for the dry parameterization groups WSM3 and ACM2, we find an overestimation in the Sudano-Sahel. The largest bias reaches from −3.5 to 1.7 mm day−1 at the Guinea Coast, −2.1 to 3.2 mm day−1 in the Sudano-Sahel, and −1.4 to 0.5 mm day−1 in the Sahel, indicated by the whisker difference of ENS to TRMM.
The MP schemes show the same overall tendencies for the three regions, with a clear order (TH being the wettest and WSM3 being the driest). The mean precipitation difference between WSM3 and TH is 1.5 mm day−1 at the Guinea Coast, 2.4 mm day−1 in the Sudano-Sahel and 0.7 mm day−1 in the Sahel. Thus, the average intra-ensemble spread induced by the MP schemes is close to the magnitude of the bias to TRMM, which underlines the considerable impact of the MP schemes.
Looking at the PBL schemes, the picture is more diverse for the different regions: While MYJ is dry at the Guinea Coast and wet in the Sahel, ACM2 behaves the opposite way. This indicates a shift of the monsoon rainband, dependent on the choice of the PBL scheme. YSU shows an almost as strong northward shift as MYJ, but with much wetter conditions at the Guinea Coast due to a generally wider rainband as can be seen from Fig. 4. In terms of northward shift of the monsoon rainband, the order of the PBL schemes is ACM2 < YSU < MYJ. The interquartile ranges of the ACM2 and MYJ parameterization groups do not intersect for any of the three regions, which underlines the opposing impact they have on the position of the rainband. However, the strongest northward shift does not necessarily coincide with largest precipitation amounts: YSU instead of MYJ shows largest values in the Sudano-Sahel, where the core of the rainband is situated.
The CU schemes show the largest interquartile-spreads and on average only weak dry/wet tendencies with respect to ENS. They also show the smallest mean difference in precipitation of only 0.6 mm day−1 at the Guinea Coast, 0.9 mm day−1 in the Sudano-Sahel and 0.4 mm day−1 in the Sahel. This suggests an inferior role for the generation of precipitation in the model. However, this depends on the region: In the Sahel, KF (GF) shows a neutral (restrictive) behavior, but a restrictive (neutral) behavior at the Guinea Coast. BMJ dampens the effect of other schemes in both regions, indicated by the consistently smaller inter-quartile spread.
Convective and non-convective precipitation
The ability of an atmospheric model to simulate convective processes is crucial and at the same time a limiting factor for the quality of model precipitation in the tropics and sub-tropics. In the WRF model, precipitation from unresolved deep convection is generated by the CU scheme, while the MP scheme produces grid-scale precipitation in case the air is still super-saturated after the instabilities are released. Thus, the convective fraction of the model is artificial and related to the model resolution. Nevertheless, the partitioning into convective and non-convective precipitation helps to identify the impact of either scheme on the representation of convection.
The amount of non-convective precipitation follows the same order we found earlier for the MP schemes (WSM3 < LIN < TH), with a total spread of 93 mm month−1 exceeding the spread of convective precipitation (68 mm month−1). There is no clear correlation between the CU schemes and the amount of convective precipitation. Model configurations with KF or BMJ generate convective precipitation from 40 mm month−1 to 110 mm month−1, depending on the choice of MP and PBL scheme, while GF generates less convective precipitation and never exceeds 80 mm month−1. In particular, the ACM2 scheme leads to small amounts of convective precipitation for all CU schemes and results in a very dry regime. The impact of a particular CU scheme depends on the chosen PBL scheme and vice versa. For example, for the MYJ PBL scheme, maximum convective precipitation is achieved in combination with KF, while for the YSU and ACM2 PBL schemes, BMJ produces almost consistently the largest amounts of convective precipitation. The mean convective fraction over the whole domain of the individual ensemble members varies between 24 and 63 %, with consistently lower values for ACM2 configurations and higher values for BMJ and YSU configurations.
Figure 6b illustrates more clearly the sensitivity of precipitation amounts with respect to each parameterization type. Each box consists of the nine precipitation spreads between ensemble members for which only the indicated parameterization type is rotated. For example, one of the nine spreads for CU is computed between (KF_LIN_YSU, GF_LIN_YSU, BMJ_LIN_YSU), another between (KF_TH_MYJ, GF_TH_MYJ, BMJ_TH_MYJ) and so forth. For non-convective precipitation, CU and PBL schemes show an average spread of about 20 mm month−1 compared to ~60 mm month−1 for MP, which indicates only minor influence. The average spread in convective precipitation is larger for configurations that differ in their PBL scheme than for those that differ in their CU scheme. However, the large inter-quartile spread of both illustrates their non-linear interplay for the production of convective rainfall. Here, the MP scheme is of minor importance. The sensitivity of total precipitation amounts to the MP and PBL choice is almost equal with average spreads of 67 ± 13 mm month−1 and 62 ± 11 mm month−1, respectively. The importance of the CU schemes is reduced and highly variable with a spread of 31 ± 21 mm month−1.
However, their impact becomes stronger on finer temporal scales: KF and GF have difficulties to reproduce the amplitude of the diurnal cycle (Fig. 7a), which results in a large overestimation of precipitation in the morning hours. BMJ produces the convective peak about 3 h too early, but is close to the amplitude of TRMM, especially in combination with the YSU PBL scheme. However, the phase of the diurnal cycle is somewhat better captured by GF and KF with the convective peak at 18 h for most configurations (Fig. 7b). Nikulin et al. (2012) report a shift of the phase of the diurnal cycle for almost all models in an ensemble of CORDEX-Africa regional climate simulations that includes the WRF model. They relate this deficiency to the formulation of the cumulus parameterizations. In accordance to our findings, their WRF-KF configuration captures the phase of the diurnal cycle reasonably well, but with a stronger amplitude. Figure 7b confirms that differences in phase and amplitude mainly arise from the convective precipitation fraction (CU scheme). Non-convective precipitation amounts mostly show a uniform phase and the amplitude is closely related to the respective CU scheme activity. BMJ is almost inactive during night hours, as was also found by Pohl et al. (2014), leading to an underestimation of precipitation compared to TRMM. According to Marsham et al. (2013) and further experiments that we conducted at convection-allowing scales (12–4 km, not shown), the explicit treatment of convection greatly improves the representation of the diurnal cycle and removes the phase shift.