Based on the mean wind profiles, we determine the average depth of the monsoon layer (Fig. 2, left). At both Kumasi and Savè, a distinct minimum in mean wind speed (≤1.5ms-1) and a shift in wind direction from south-west to east occur at about 1.9 km above ground level (a.g.l.). This wind shear zone is often used to define the height of the monsoon layer (e.g. Fink et al., 2017). This height shows great variability with time as indicated by the increase in the standard deviation of the wind direction with height above the ground. The standard deviation of the wind direction is calculated according to Yamartino (1984). In the monsoon layer, a mean wind speed maximum of around 6 m s-1 at Savè and 8 m s-1 at Kumasi occurs at about 400 m a.g.l. This wind maximum is related to the LLJ that is usually still present at 06:00 UTC. Possible reasons for the greater mean wind-speed maximum in Kumasi compared to Savè will be investigated in Sect. 4. No distinct AEJ is visible in the mean wind profiles. This is because the AEJ was further to the north in the second half of the campaign and due to averaging over opposing winds; i.e. normally winds are easterly but westerly during the vortex period (Fig. 2, right). The standard deviation of the wind speed, which is more than ±3 m s-1 in most of the layers at both sites, reflects strong day-to-day variations.

In the monsoon layer, the mean RH varies between 80 and 100 % at both sites (Fig. 3, left). Above the monsoon layer, RH is about 7 % at Kumasi. At Savè, RH has a minimum of 75  % at about 3 km a.g.l. and increases to about 85 % above. That means, on average, the atmosphere between 3 and 5 km a.g.l. at Kumasi is somewhat drier than Savè. This could be caused by convective activity upstream of Savè in Nigeria. Note that the standard deviation of RH, especially above 1 km, is quite high (more than ±10 %). The mean temperature profiles are rather similar at both sites, but the atmosphere at Savè is slightly warmer than at Kumasi. The mean gradient of the potential temperature is about 6 K km-1 between 0.5 and 2 km and about 5 K km-1 between 2 and 5 km. As at the same time the mean gradient of the saturated equivalent potential temperature in these two layers is negative (not shown), the lower troposphere is conditionally unstable. At both sites, the standard deviation of the potential temperature is on the order of ±1 K.

The large standard deviation of wind speed and direction can be explained by the atmospheric conditions on individual days (Fig. 2, right). In the harmattan layer above the monsoon layer, the AEJ is strong at both sites (mainly > 10 m s-1) during the first half of the observation period. On 11 July, a cyclonic vortex slowly propagates west–north-westwards from Gabon across the Gulf of Guinea and reaches Sierra Leone on 14 July (Knippertz et al., 2017). This causes large-scale westerly winds up to about 3 km a.g.l. associated with a weakening and northward shift of the AEJ. This is also visible in the wind profiles at Savè (11–13 July) and Kumasi (11–14 July). After the vortex period, the mid-level easterly winds remain weak in the second half of the measurement period (15–30 July) at both sites. The large standard deviation in the monsoon layer partly arises from days that are affected by mesoscale convective systems (MCSs), e.g. on 19 June in Savè. As the wind profiles at 06:00 UTC result from the monsoon flow with the embedded LLJ, variations are caused by large-scale conditions and the characteristics of the LLJ. For example, a deep layer with high wind speed can be found on 21 June and 11 July at both sites.

The classical concept is that the monsoon layer is associated with south-westerly winds and moist air. This means that in general these qualities can be used to identify the monsoon layer. Using RH as an indicator for the moist layer, Fig. 3 shows that the day-to-day variability of the depth of the moist layer in the lower atmosphere varies by a factor of 2 or more and does not always coincide with the depth of the south-westerly wind layer. On the one hand, periods occurred when the moist layer was much deeper than the monsoon layer. These periods can often be assigned to MCSs passing the sites, accompanied by deep vertical mixing. For example, on 16 July a long-lived MCS (associated with cyclonic system “C” in Knippertz et al., 2017) passed Savè and on 26 July another MCS (associated with cyclonic system “J”) affected Kumasi. On the other hand, the moist layer is significantly shallower than the monsoon layer during some periods. For example, the moist layer is only some hundred metres deep at Savè during the end of the pre-monsoon phase. During the last part of the post-onset phase, the moist layer at Savè is somewhat drier than the weeks before. This period is roughly related to the vortex period when dry air was transported northwards.

Cloud-base and cloud-top heights estimated from RH profiles are also included in Fig. 3. Days with precipitation during the radiosounding were excluded from the analysis (Kumasi: 11 and 24 July; Savè: 19 June, 20 and 24 July). The results show more frequent occurrence of LLCs at Kumasi (88 %) than Savè (64 %) at 06:00 UTC (purple dots in right panels of Fig. 3). In Savè, the median height for cloud base is 227 m a.g.l. and for cloud top is 587 m a.g.l., while in Kumasi, the median heights are 137 and 692 m a.g.l. respectively. That means on average, LLCs in Kumasi are more often thicker than in Savè in the morning. As for the other parameters, there is evidence for a link between the monsoon phases and LLC occurrence and depth (see also Fig. 17 in Knippertz et al., 2017). During the pre-onset phase, LLCs are rather thin at both sites and less frequent at Savè. A higher LLC occurrence is observed during the post-onset monsoon phase, whereas dry air advection during the vortex period tends to reduce the LLC depth at Kumasi and even prevents LLC formation at Savè. For the remaining observation period, LLC formation is more sporadic, partly due to the increase in MCS events. Earlier work has already pointed to the fact the LLC forms in nights that lack deep convection, but farther north, cloud-free non-precipitative nights appear to occur more frequently (e.g. Schrage et al., 2007; Schrage and Fink, 2012). More information on the temporal evolution of cloud characteristics is given in Sect. 4.

The IWV, calculated from radiosoundings, shows higher values at Savè (median of 55 kg m-2) than at Kumasi (median of 50 kg m-2; Fig. 4c). This difference can be attributed to higher absolute moisture content at all levels. During the whole measurement period, no clear trends are visible for either Kumasi or Savè. A similar difference between the two sites exists in the IWV calculations based on microwave radiometer data (see Fig. 7). In Kumasi, the vortex period is accompanied by lower IWV values (Fig. 4c). This IWV decrease from about 50 to 34 kg m-2 could be attributed to dry air that is advected with the south-westerly winds (Fig. 2a) from the area of subsidence in the equatorial zone or even the southern hemisphere. Knippertz et al. (2017) report that the intrusion of dry air resulted in RH values as low as 10 % in the middle troposphere over Abidjan in Côte d'Ivoire during this vortex phase. Farther east in Savè, the dry-air advection is much less pronounced and does not show up in the IWV (Figs. 2b and 4c).

The convection-related parameters convective available potential energy (CAPE) and convective inhibition (CIN) are calculated from the soundings in Kumasi and Savè and shown in Fig. 4a and b. As the soundings are performed at 06:00 UTC, both quantities are calculated using the most unstable layer for lifting (Doswell III and Rasmussen, 1994), because surface-based parcels are inappropriate for lifting when a nocturnal surface inversion still exists. At both sites, CAPE shows strong day-to-day variability, ranging from about 100 to about 2000 J kg-1, with no clear dependence on the different monsoon phases (Fig. 4a). The median for the whole period for Kumasi is 434 and 508 J kg-1 for Savè. We attribute the large day-to-day variability of CAPE to convective precipitation events, which are typically followed by a decrease in CAPE as described by Schwendike et al. (2010) for the Sahel region and Schrage et al. (2006) for the Sudan region. High values of CIN (about 50 to 100 J kg-1) exist during the pre-onset phase (14–20 June) and at the end of the post-onset phase (about 14–21 July) of the monsoon (Fig. 4b). During these phases, stronger precipitation occurs (Fig. 4d). This link can be explained by a build-up of latent instability due to an accumulation of warm, moist air in lower layers in the presence of higher CIN values. This ultimately enhances the potential for stronger convection with heavy precipitation (e.g. Browning et al., 2007; Khodayar et al., 2010). The lowest CIN values (< 50 J kg-1) are mainly observed during the post-onset phase, associated with little precipitation (Fig. 4d), typically generated by local convection. Fink et al. (2006) describe different types of rainfall associated with different CAPE and CIN values and intensities of precipitation for Parakou, which is about 130 km north of Savè. It is conceivable that the observed variability in CAPE and CIN and the related rainfall events fit the types of rainfall proposed by Fink et al. (2006), but this is left for future studies. The total amount of precipitation over the whole period ranges from 217 mm in Savè to 258 mm in Kumasi to 271 mm in Ile-Ife. It has to be kept in mind that all precipitation data are based on local measurements, which are not necessarily representative of the amount of convective precipitation in the area. However, the temporal distribution of precipitation in Savè shown in Fig. 4d agrees with that of the Savè X-band rain radar, which covers a diameter of up to 200 km (not shown).

The energy exchange at the surface is analysed using the evaporative fraction EF=LE0/(LE0+H0), i.e. the ratio of the latent heat flux to the sum of latent and sensible heat fluxes or available energy (Fig. 4d). This quantity was calculated from H0 and LE0 values, which were averaged for the time period from 09:00 to 15:00 UTC, i.e. when the fluxes are sufficiently high. A value of 1 indicates that all available energy goes into the latent heat flux, a value close to 0 means that the sensible heat flux dominates and at a value of 0.5 energy is equally distributed between sensible and latent heat fluxes. The median of EF is 0.58 in Kumasi, 0.64 in Savè and 0.69 in Ile-Ife; i.e. on average the majority of the available energy is transformed into latent heat flux – less in Kumasi and more in Ile-Ife, although both are grassland sites. Although the EF at the different sites is rather constant over the campaign, some of the changes of EF are related to rain events; for example, the increases of EF in Savè from about 0.5 on 12 July to 0.75 on 13 July and from 0.6 on 19 July to 0.8 on 20 July. In Kumasi, a precipitation-generated EF increase from 0.4 to 0.7 is observed from 16 to 17 July. Similar behaviour of the evaporative fraction after precipitation events, when the soil is not saturated, is also reported by Kohler et al. (2010) and Lohou et al. (2014) for the AMMA field campaign.

From the 1 min cloud-base heights obtained from ceilometer measurements, we calculate the cloud-base fraction for 30 min intervals for the lowest 1000 m above ground (Fig. 5a, b). Cloud-base fraction gives information on the number of cloud bases in a certain time period and height layer; i.e. a cloud-base fraction of 100 % indicates that at all times in the 30 min interval is at least one cloud base detected in the lowest 1000 m a.g.l. Using a threshold of 100 % for LLC detection, it is evident that LLCs develop during many nights at both sites, Kumasi and Savè. At 06:00 UTC, i.e. at the same time when daily radiosoundings were performed, LLCs occur less often in Kumasi (approximately 57 %) than in Savè (approximately 64 %), which differs from the finding based on the radiosonde data (LLCs occur more often in Kumasi than in Savè, Sect. 3.2). However, in Kumasi the results are very sensitive to the selected threshold for cloud-base fraction. Applying a threshold of 80 %, LLC occurrence increases to approximately 80 % in Kumasi while it remains approximately the same in Savè. Nights without LLCs are found between 14 and 17 June at Savè, and around 23 and 24 June and 14–16 July at both sites. The latter period is roughly associated with the vortex occurrence, during which dry air masses are transported into the investigation area (see Knippertz et al., 2017). Using a median cloud-base fraction of 100 % as an indicator for the average onset of LLCs, LLCs approximately form at 00:00 UTC in Kumasi and 03:00 UTC in Savè (Fig. 5b). The onset of LLCs is accompanied by an increase in the median LWP to about 40 g m-2 at Savè and to about 50 g m-2 in Kumasi (Fig. 5d) and the net longwave radiation, Lnet=L↓-L↑, increases to about Lnet≈-10Wm-2 at both sites (Fig. 5e). Using a threshold of Lnet≈-10Wm-2 as a proxy for the existence of LLCs, as indicated from the comparison of net longwave radiation with cloud-base fraction in Kumasi and Savè, LLCs in Ile-Ife already form at around 21:00 UTC.

From the cloud-base fraction we are unable to determine whether the cloud bases are all at the same height or distributed over several layers within the 1000 m layer. To get an idea of the vertical distribution of cloud-bases we also calculate the frequency distribution of cloud-base height by counting the number of days with at least one cloud base within a respective bin (bin size is 10 min duration and 20 m vertical range, Fig. 5c). As the backscatter coefficient profiles at Kumasi show no realistic values below around 50 m a.g.l. this layer is masked by the black bar in Fig. 5c. While there is some variability in cloud-base height during the campaign, we are able to distinguish those layers particularly favourable for cloud base occurrence. In Savè, surprisingly two layers are evident: one is at around 100 m a.g.l. and the other one at around 300 m a.g.l. Determining the cause for this will require detailed investigation. In Kumasi, cloud-base height occurs mainly in a layer around 200 m a.g.l. Radiosonde profiles at Kumasi indicate that cloud base also occurs at lower layers (Fig. 3a). Unfortunately, this cannot be verified by ceilometer data due to missing measurements.

At both sites, the cloud base starts to rise at approximately 07:00 UTC, i.e. about 1 h after sunrise (time of sunrise see Table 1a). During the rising period the cloud-base fraction remains close to 100 % (Fig. 5b) and fluctuations in cloud-base height remain small for some time (Fig. 5c), indicating that the clouds are still stratiform. Eventually a transition to convective clouds occurs, indicated by a decrease in cloud-base fraction and a stronger fluctuation of cloud-base height. On average, the transition is approximately 2 h after the start of the rising period; however, we find large variability of several hours. Parallel to the cloud base rising around 07:00 UTC, the LWP at Savè reaches maximum values of up to 100–125 g m-2 and Lnet decreases at both sites (Fig. 5d). Around midday, when the convective boundary layer is well developed, the cloud-base fraction decreased (to approximately 40 % in Kumasi and 80 % in Savè), Lnet≈-40 W m-2 and the cloud-base height is at around 800 m a.g.l. In Ile-Ife, Lnet starts to decrease at 07:00 UTC indicating the dissolution of the LLCs at that site, too. Overall, the LLCs at the individual sites show considerable variability from day to day with respect to their time of formation and dissolution, cloud-base fraction, LWP and cloud-base height. The same finding holds when comparing the LLC conditions between the different supersites. Next, the diurnal cycles of quantities related to the LLC formation are presented.

Figure 6 presents composites of the diurnal cycles of the radiation and energy balance components for the three sites. Figure 6a indicates that the radiation balance components are quite similar at the three sites (Fig. 6a). The maximum median G reaches up to approximately 550 W m-2 with an interquartile range of approximately 250 W m-2 at 13:00 UTC. This is caused by the considerable day-to-day variability of cloud cover. The median R is about 110 W m-2, i.e. the albedo at the three supersites is approximately 0.2, which is a typical value for grasslands (e.g. Oke, 1987). The longwave radiation shows much less variability.

At all three sites, the median Q0 is of similar magnitude, at 400–450 W m-2 around midday (Fig. 6b). At this time, the median is accompanied by a significant interquartile range of 200 W m-2 (Fig. 6b), as is to be expected given the variability in the radiation components. The partitioning of the available energy (Q0-B0) between H0 and LE0, however, differs at the three sites. In Kumasi, the median LE0 is only slightly higher than H0, which is consistent with a median evaporative fraction 0.58 (Sect. 3). In Savè, LE0 clearly dominates H0; e.g. at noontime the median LE0 is about 200 W m-2, while the median H0 is about 120 W m-2. At Ile-Ife, the median LE0 is also about 200 W m-2 and the median H0 is only 100 W m-2 at noontime, resulting in the highest evaporative fraction (0.69) of all three sites. The accumulated daily evapotranspiration, E0, derived from the median LE0 values, amounts to 1.3 kg m-2 at Kumasi, to 2.1 kg m-2 at Savè and to 1.6 kg m-2 at Ile-Ife. The turbulent fluxes at all three sites during daytime are associated with large interquartile ranges, due to both the strong day-to-day variability of available energy and because precipitation events modify the evaporative fraction from day to day (Fig. 4d). At all three sites, around sunset the median Q0 becomes negative (around -10 W m-2). The absolute value of H0 is low at night (median between H0=-8.5 W m-2 in Savè and -0.3 W m-2 in Ile-Ife), but by contributing to the compensation of the negative Q0, H0 contributes to the development of a stably stratified nocturnal surface layer, the decoupling of the daytime mixed layer from the surface layer and the development of an LLJ.

As expected for a tropical region during the monsoon season, the near-surface temperature exhibits only a moderate diurnal cycle with a median diurnal temperature range of about 6 ∘C at the three sites (Fig. 7a). The median temperature maximum of 29 to 30 ∘C is reached in the early afternoon (about 15:00 UTC). The temperature's day-to-day variation is moderate, too; e.g. the interquartile range at 15:00 UTC is only about 2 ∘C. The diurnal amplitude of the median q is small as well, i.e. only about 1.0 g kg-1 in Kumasi, 1.2 g kg-1 in Ile-Ife and 1.8 g kg-1 in Savè (Fig. 7a). The median q is highest at Savè (about 19 g kg-1 between 14:00 and 20:00 UTC). This might be attributed to the higher evapotranspiration (Fig. 6b). The microwave-derived IWV values (Fig. 7d) do not show a pronounced diurnal cycle, while precipitation reveals a strong time dependence (Fig. 7d). A period with more precipitation occurs between about 15:00 and 23:00 UTC. This time period is typical for local convective precipitation. The occurrence of local and patchy precipitation in the late afternoon is confirmed by the X-band rain radar data at Savè (not shown).

When the surface layer becomes stably stratified around sunset, as indicated by the positive Rf at all sites (Fig. 7c), the near-surface wind speed and the TKE decrease (Fig. 7b and c). This decoupling of the surface layer from the mixed layer allows an LLJ to develop at all three sites (Fig. 8). The temporal cross section of median wind profiles is based on sodar measurements in Kumasi and Ile-Ife, while for Savè a combination of sodar and UHF wind profiler measurements are available. Affected by local noise production from a generator, the sodar at Kumasi did not reach high altitudes, especially at night (Fig. 8a). In Savè and Ile-Ife, the LLJ is well established at about 20:00 UTC (Fig. 8b and c). In Savè, the jet core can be found between 300 and 500 m a.g.l., and the maximum median wind speed is about 8 m s-1 with a decrease in wind speed in the second half of the night. In Ile-Ife, the top of the LLJ is not covered by the sodar. Mean hodographs of the wind speed, calculated for the 200–600 m layer, show a clockwise turning of the wind in the course of the night from south–south-westerly at sunset to south-westerly at sunrise. The evolution of the nocturnal LLJ at the different sites is similar to observations described by Lothon et al. (2010) and Abdou et al. (2010) from AMMA. In the upper and lower shear zones of the LLJ, turbulence is expected (e.g. Banta et al., 2006), and gravity waves, similar to those at the top of drainage flows (Viana et al., 2010), are likely. Both are expected to have an impact on the LLC formation (Adler et al., 2017) and will be investigated in subsequent case studies.

During the night, the median near-surface wind speed and TKE are higher in Kumasi (v≈1.5 m s-1, TKE up to 0.5 m2 s-2) than at the two other sites (v≈1.0 m s-1, TKE up to 0.2 m2 s-2; Fig. 7 and c). Due to higher TKE, the surface layer in Kumasi is less stably stratified, i.e. the positive values of Rf are lower than at the two other sites. We assume that differences in LLJ are likely responsible for the differences in the near-surface conditions. A weaker LLJ results in weaker downward mixing (lower TKE). To seek an explanation we inspect the large-scale pressure gradient as a driving force for the LLJ and the orography: (i) there is evidence that the meridional pressure gradient is larger in the western part of the investigation area, i.e. the geostrophic wind has a larger zonal component at Kumasi, and (ii) the 700 m high Atakora Mountains in Togo (Fig. 1) cause a reduction of wind speed on the downstream side (as evident in simulations by Schuster et al., 2013), which also affects Savè. Shortly after sunrise, when the surface layer becomes unstable again (Fig. 7c), convective vertical mixing starts and momentum is transferred down to the surface again, as indicated by a TKE increase (Fig. 7c), and the LLJ starts to dissolve (Fig. 8). Simultaneously, near the ground the wind speed increases and the wind direction becomes more westerly (Fig. 7b).