Precipitation information from two different sources is considered in this study. The first is the Tropical Rainfall Measuring Mission (TRMM) 3B42 version 7 dataset. TRMM is a joint mission of the National Aeronautics and Space Administration and the Japan Aerospace Exploration Agency covering the tropical and subtropical regions of the Earth during 1997–2015. This dataset is created with the TRMM Multisatellite Precipitation Analysis method (Huffman et al., 2007), combining the TRMM precipitation radar with measurements from microwave and infrared sensors on several low Earth orbiting and geostationary satellites, and is calibrated with rain gauge data on a monthly basis. The rainfall data used in this study were aggregated from 3-hourly measurements on a 0.25∘ × 0.25∘ grid.

In addition to TRMM, rainfall from the Global Precipitation Climatology Project (GPCP) was used. GPCP combines several sources of rainfall measurements into one global dataset with a high data density and accuracy. It was established by the World Climate Research Programme to quantify the distribution of precipitation around the globe on climatological timescales (Adler et al., 2003). In GPCP, ground-based rain gauge measurements and satellite-based precipitation estimates are combined to give a merged product. The rain gauge measurements stem from the Global Precipitation Climatology Centre monitoring product of the German Weather Service (DWD). The satellite data consist of infrared and microwave-radiance-derived rainfall estimates from geostationary and polar orbiting satellites. We used daily data in 1.0∘ × 1.0∘ horizontal resolution.

SSI measurements stem from the climate data record SARAH (Surface Solar Radiation Data Set Heliosat) version 2. It was created by the Satellite Application Facility on Climate Monitoring (CM SAF) based on Meteosat Visible and Infrared Imager (MVIRI) and Spinning Enhanced Visible and InfraRed Imager (SEVIRI) measurements on the geostationary Meteosat satellites (Müller et al., 2015). From MVIRI, the broadband visible channel and from SEVIRI the channels 0.6 and 0.8 µm are used. SARAH was produced using a retrieval system based on the Heliosat method and an efficient clear-sky surface solar radiation transfer model (Müller et al., 2009; Posselt et al., 2012). For this study we use the monthly mean products of the dataset with a horizontal resolution of 0.05∘ × 0.05∘. In addition, we employ the much coarser EBAF-Surface Ed4.0 dataset (Energy Balanced And Filled) containing monthly averaged SSI fields with a horizontal resolution of 1∘ × 1∘. This product is based on the CERES (Clouds and the Earth's Radiant Energy System) algorithm (Loeb et al., 2009; Young et al., 1998), which uses information from the CERES shortwave broadband radiometers but also from instruments on geostationary satellites to account for the diurnal variability in the data. Several CERES instruments are mounted on polar orbiting satellites such as TRMM, Terra, Aqua and NPP (Suomi National Polar-orbiting Partnership). To derive the radiative fluxes at the surface, cloud imager data for scene classification, cloud physical properties, temperature, water vapour, ozone and aerosol data as well as a broadband radiative transfer model are needed.

The satellite-derived SSI fields are complemented with a small set of surface measurements. Unfortunately, there are very few ground-based measurements of SSI available in the DACCIWA study region during July 2006. South of 10∘ N, only the stations Lamto (Ivory Coast; 6.22∘ N, 5.03∘ W), Cotonou and Parakou (Benin; 6.35∘ N, 2.43∘ E and 9.33∘ N, 2.62∘ E, respectively) delivered gap-free measurements from standard instruments, i.e. a Gunn-Bellani radiometer (Lamto) and CNR1 radiometers from Kipp & Zonen (Parakou and Cotonou).

For OSR at TOA, monthly mean averages from the dataset GERB/SEVIRI ed. 2.0 from CM SAF (Clerbaux et al., 2017) were used. GERB is the geostationary Earth radiation budget instrument onboard Meteosat Second Generation satellites (Harries et al., 2005). This broadband radiometer is designed to measure the Earth's total emitted longwave and solar reflected radiances with high temporal resolution (5 min) and 50 km grid spacing. It is available as TOA reflected shortwave and TOA emitted thermal fluxes. In the present study, we consider only the shortwave flux. For this dataset, SEVIRI measurements are employed to refine GERB's original spatial resolution to a 0.1∘ × 0.1∘ grid. In addition, the 1∘ × 1∘ monthly EBAF-TOA Ed4.0 dataset for shortwave radiation is used that has been derived using the same CERES algorithm as for the surface.

The highly scalable ICON model (Zängl et al., 2015) was recently developed by the Max Planck Institute for Meteorology and the DWD; it became DWD's new operational global NWP in January 2015. ICON's horizontal Arakawa C-type grid is based on triangles, which cover the globe with approximately equal area everywhere and allow easy nesting. The vertical coordinate is height-based and terrain-following in the lower levels but smoothed in the upper troposphere via the application of a SLEVE (smooth level vertical) coordinate (Leuenberger et al., 2010). For the dynamical core the continuity equation is formulated in the flux form with density as the prognostic variable, enabling exact local mass conservation. The equations are solved non-hydrostatically and the time integration is performed with a two-time-level predictor–corrector scheme. Apart from the sound wave propagation, this scheme is fully explicit. The fast physics packages are inherited from the Consortium for Small-scale Modelling (COSMO) model (Doms and Schättler, 2004) but are partly reformulated for ICON. The cloud microphysics scheme is the COSMO-EU five-category prognostic scheme (Doms and Schättler, 2004; Seifert, 2008) with the extension of ice sedimentation. The turbulence scheme by Raschendorfer (2001) solves the prognostic equation for turbulent kinetic energy (TKE) and for the land–surface interaction TERRA (Heise et al., 2006) is used in an updated version. The slow physics parameterizations correspond to those from the Integrated Forecasting System (IFS) of the European Centre for Medium-Range Weather Forecasts (ECMWF): the Bechtold et al. (2008) convection scheme, the Lott and Miller (1997) sub-grid-scale orography scheme and the Orr et al. (2010) non-orographic gravity-wave drag scheme. Radiative transfer is solved with a rapid radiation transfer model (RRTM; Mlawer et al., 1997), whereby a Green's function approach is applied for solar bands with approximated diffuse radiation (Barker et al., 2002).

Currently ICON is only configured as a global model but its high flexibility in terms of one- and two-way nesting allows a regional focus without the undesirable boundary effects sometimes observed for traditional limited-area models. It performed well compared to ERA-Interim (ERA-I hereafter) in several test cases we ran for DACCIWA. The comparison can be found in the Supplement. The simulation period was not so much limited by computational cost but by the large amount of output, since many different state variables had to be saved for the analysis.

All simulations in this paper were initialized with ERA-I, ECMWF's global atmospheric reanalysis (Dee et al., 2011), and do not use data assimilation. ERA-I is created by assimilating all available measurements into a single forecast model environment, resulting in a multivariate, spatially complete and coherent record of the global atmospheric state. ERA-I data are used in the highest possible horizontal resolution of about 80 km and with 60 vertical levels up to 0.1 hPa. Typically ERA-I contains most observations at 12:00 and 00:00 UTC. Initializing ICON runs at 00:00 UTC would mean starting the runs during the development phase of the low-level clouds, and therefore the runs were initialized at 12:00 UTC.

To assess the impact of variations in cloudiness in the ICON model, a series of experiments was designed. In these, the original cloud liquid water content qc in the DACCIWA study region and below 700 hPa (see Fig. ) is manipulated immediately before the call of the radiation scheme by multiplying it with an opacity factor fop to mimic an increase or decrease in the low clouds' optical thickness. After that, qc is set back to the original value and the model is allowed to run freely until the next call of the radiation routine. In this way it is ensured that only the radiation can impact the dynamics and thermodynamics, creating changes in temperature T, relative humidity RH and winds, which in turn can influence the development of cloud itself. fop is varied from 0.1 to 10, where fop=1 corresponds to the control experiment. The low values are at the extreme end of cloud underrepresentation found in Hannak et al. (2017), while fop=10 should be regarded as a somewhat unrealistic sensitivity test.

Two sets of experiments were performed with ICON.

PARAM. For this set, ICON was run in the current operational global setting with a horizontal grid spacing of 13.2 km and 91 vertical levels. Integration time is 5 days. fop is varied in eight steps from 0.1 to 10.0 to systematically analyse the effect of low-level clouds. Due to the relatively high computational costs, runs are restricted to July 2006 and only started every fourth day in order to have 1 day of overlap between the simulations. All in all 8×8 5-day simulations were performed for this set.

For precipitation, PARAM generally shows a distinct maximum at 15:00 UTC (corresponding to local time in our study region) and the lowest rainfall in the second half of the night (Fig. ). Consistent with Fig. c, a decrease in fop leads to a monotonic and smooth increase in precipitation at all times of day, apart from the early morning hours when the effect is weak. At the time of maximum precipitation, the rainfall from experiment fop=0.1 is 2.5 times larger than that for fop=10.0. The morning onset of rainfall is earlier for low fop, as the build-up of instability due to incoming solar radiation occurs faster after sunrise. EXPL shows some significant differences (blue lines in Fig. ). The diurnal peak is shifted to 18:00 UTC, as it takes more time to trigger convection without a parameterization (Marsham et al., 2013). This corresponds much better to the typical timing of precipitation observed in this area (Kalthoff et al., 2018) and to the TRMM observations included in Fig. , despite the overall large bias already discussed (Fig. ). The onset of precipitation is not strongly affected by fop in EXPL, but the cessation is, with convection persisting much longer into the night for the optically thinnest low-level clouds, suggesting a much higher degree of organization. The latter is reflected in a larger variance of qv throughout the lower and mid-troposphere in EXPL than in PARAM (not shown). We have no explanation for the kinks in the curves around 12:00 UTC in EXPL and therefore attribute those to insufficient sampling. In terms of the diurnal maxima, values for EXPL are systematically lower with 0.17 and 0.47 mm h-1 for fop=1.0 and fop=0.1, respectively, compared to 0.32 and 0.5 mm h-1 for PARAM.

Figure  shows the diurnal cycle in the vertical structure of cloud cover CLC, cloud water content qc and cloud ice content qi for PARAM and EXPL and for fop values of 0.1 and 1.0. PARAM shows a clear three-layer cloud structure at all times of day as documented for other tropical regions (e.g. Johnson et al., 1999). Low-level clouds are mostly confined to below 750 hPa with a relatively minor mid-level cloud layer around 500–600 hPa. While the former contain significant amounts of qc (middle column of Fig. ), the mid-level clouds also contain some cloud ice (right column of Fig. ). In addition, a substantial high-level cloud cover between 400 and 100 hPa containing significant amounts of cloud ice is simulated in PARAM. In particular, the low and high clouds show a distinct diurnal cycle. At 00:00 UTC the low-level cloud deck is beginning to form, reaching a sharp peak around 950 hPa at 06:00 UTC accompanied by a corresponding increase in qc (Fig. a and b). At midday (Fig. c), radiative heating lifts and dissolves the low-level cloud deck, shifting the maximum in CLC and qc to 850 hPa (Fig. c). Finally, by 18:00 UTC (Fig. d) daytime heating and mixing have reduced CLC and qc to create a diurnal minimum. This general diurnal behaviour in low-level cloudiness in PARAM resembles that found in ECMWF analysis data (see Hannak et al., 2017). Mid-level clouds do not show pronounced diurnal variations but also have a minimum in CLC and qc at 18:00 UTC, possibly suggesting similar mechanisms as for the low clouds. High-level CLC and qi are lowest at 12:00 UTC and highest at 00:00 UTC, when they respectively reach more than 30 % and almost 0.008 g kg-1. This indicates a relationship of high-level clouds with the diurnal cycle of convection (Fig. ), leading to an increase in the second half of the day.

Reducing the optical thickness of low clouds in PARAM (fop=0.1; dashed green lines in Fig. ) has hardly any impact on low-level CLC during night-time but leads to a small decrease at 12:00 UTC and an even lesser decrease at 18:00 UTC, possibly due to a deeper and/or drier PBL. Surprisingly, however, qc is decreased by of the order of 0.01 g kg-1 at all times and most strongly so at 00:00 UTC, indicating that for fop=0.1 a similar cover of clouds is achieved with less liquid water. This aspect will be further discussed in the following subsection. For high clouds, in contrast, both CLC and qi increase markedly for all times with absolute increases of the order of 7 % and 0.005 g kg-1 at the peak of the profile at about 250 hPa. This is likely a reflection of the increased daytime convection in the sensitivity experiment, leading to more precipitation (Fig. ) and generating substantially more cirrus. This also suggests that part of the effect of more solar radiation reaching the surface through the optically thinned low clouds is compensated for by an increase in high clouds. The comparison with the radiative transfer results by Hill et al. (2018) in the previous section, however, suggests that this is a relatively small effect overall.

Comparing the results for PARAM with those for EXPL reveals some substantial differences. Low clouds are more abundant in EXPL at all times, as already suspected in Sect. , contain substantially more liquid water and peak at 12:00 rather than 06:00 UTC as in PARAM. qc can be up to 0.09 g kg-1 higher for EXPL. The sensitivity of qc to fop has a much stronger diurnal cycle with little effect at 00:00 UTC, a small increase at 06:00 UTC, and a large increase and deepening at 12:00 UTC, followed by a decrease at 18:00 UTC (middle panels in Fig. ). Consequently, the signals at 06:00 and 12:00 UTC go in the opposite direction in EXPL and in PARAM. This rather unexpected result will be discussed in more detail in the following subsection. In addition, there is a small increase in mid-level qc at all times. In stark contrast, high-level clouds are significantly reduced relative to PARAM in both amount and qi at all times with values up to 0.008 g kg-1 lower in EXPL. However, the general sensitivity is similar for high clouds with an increase for fop=0.1 for all times. The magnitude again appears to be related to the diurnal cycle of convection, which is delayed in EXPL relative to PARAM (see Fig. ). This comparison reveals that in many aspects the variations between EXPL and PARAM are larger than the differences between fop of 0.1 and 1.0 for each experiment. To first order, the convective parameterization appears to transport moisture more efficiently out of the low and mid-levels to deposit it into the convection-fed cirrus layer compared to explicit convection. This creates overall less sensitivity to our modifications of low clouds as already discussed in the context of Fig.  but also a weaker diurnal cycle in the sensitivities.

Given the overall higher sensitivities and likely more realistic diurnal cycle in EXPL, we will begin the following discussion of thermodynamic changes with this experiment instead of PARAM. This discussion will help shed more light onto the low-cloud behaviour and sensitivities discussed in previous sections. Figure  shows DACCIWA box-averaged profiles of differences between the fop=0.1 sensitivity experiment and the fop=1.0 control run for T, specific humidity qv, RH, TKE, qc and horizontal wind speed vhoriz. The coloured lines show eight different times of day.

With respect to T a relatively complicated vertical profile and diurnal cycle are found. Below 900 hPa, as expected, the reduced optical thickness of low clouds leads to more solar heating during the day and consequently an overall warming peaking at 15:00 UTC with a slight cooling at 06:00 UTC (Fig. a). Immediately above that, there are indications of enhanced latent heat release within the low-level cloud deck, at least for some times of day when CLC and qc increase (see Fig. ), but during the day this effect is not clearly separable from the sensible heat fluxes in the PBL. Above that, around 725 hPa is a shallow layer with a slight cooling, most pronounced during the day and possibly due to radiative effects at the low-level cloud tops. The increases in mid-level cloud and cloud water around 550 hPa (see Fig. ) also lead to a warming below (latent plus radiative heating) and radiative cooling above, with the latter most pronounced at night-time. Finally, the cirrus layer peaking around 250 hPa also produces such a dipole pattern but with a much smaller diurnal cycle consistent with Fig. .

Signals in qv, in contrast, are much simpler and show a deep atmospheric moistening at all times (Fig. b). The only drying occurs in the lowest few hundred metres at 12:00 and 15:00 UTC, when substantial amounts of moisture are pumped into the elevated low-level cloud layer where qv maximizes. An interesting time is 09:00 UTC, when qv is markedly enhanced near the surface. This may be related to an earlier start of the diurnal PBL growth (see the discussion on TKE below) or possibly also due to higher evapotranspiration in response to the increased precipitation (see Fig. ). The second maximum in qv increase is found in the area of the mid-level cloud layer around 550 hPa. Due to generally low values in the cold upper troposphere, changes in the cirrus layer are less evident in Fig. b. The net increase in column moisture and precipitation (Fig. ) suggests a substantial increase in moisture convergence into our study region. This will be further discussed in the next subsection. The signal in RH (Fig. c) is a combination of the signals in T and qv. Given the large increases in qv, RH increases everywhere above 800 hPa at all times of day, with the profile reflecting some of the modulations in the area of the mid- and high-level cloud decks already discussed. The highest RH increases of up to 5.5 % are found in the early morning at the end of a period with convective moisture transport and radiative cooling. At the very lowest layers, the large increase in T, particularly during the day, leads to a decrease in RH. The level with zero difference descends at night and ascends during daytime. It is lowest at 06:00 UTC, which facilitates the nocturnal low-level cloud formation for fop=0.1, leading to a slight increase in CLC and qc (Fig. ). At 12:00 UTC RH near the surface is reduced but the higher values above 900 hPa help expand the cloud deck upwards, while at 18:00 UTC the drying is so deep that clouds are reduced (Fig. ).

The discussion so far has illustrated the paramount importance of vertical mixing. To reveal the impact of low-cloud shielding on turbulence, Fig. d shows the vertical profile of differences between fop=0.1 and fop=1.0 for TKE, which is increased at all levels and all times. Below 700 hPa turbulence gradually dies down from 18:00 to 06:00 UTC. Due to the missing effect of low clouds in fop=0.1, TKE differences increase markedly from 09:00 to 15:00 UTC and rise upwards. 12:00 and 15:00 UTC show a secondary peak between 850 and 750 hPa, which is probably related to turbulence within the low-level cloud deck. Above 700 hPa, there is rapid increase from low values at 09:00 and 12:00 UTC to a maximum at 18:00 UTC, followed by a gradual decay. This behaviour clearly illustrates how deep convection communicates the – at first surface-based – signals into the entire troposphere. Finally, the localized maximum in TKE differences around 900 hPa at night is an indication of a slightly enhanced NLLJ creating turbulence through shear (see Fig. f), which in turn helps the cloud formation.

Figure e shows the effect of the discussed changes in RH and TKE on qc, shedding more light onto the absolute values already discussed above (solid and dashed red lines in middle panels of Fig. ). A good starting point to discuss the diurnal cycle of this signal is 18:00 UTC, when the increase in deep convection is largest (Fig. ) and creates more clouds above 750 hPa and less in the main low-level cloud deck (Fig. d), as the deeper mixing reduces RH (Fig. c). At 21:00 UTC the convective signal weakens and there are some first indications of increased qc in the nocturnal stratus deck around 925 hPa. As area mean RH is still negative at this level at this time (Fig. c), this is likely related to a greater variability within the box. The enhancement in qc in the low-level cloud deck increases and rises until 09:00 UTC. After 09:00 UTC the more dynamic evolution of the daytime PBL in fop=0.1 leads to a more elevated low-level cloud deck containing more qc in the vertical column. This denotes a negative feedback mechanism, as a (here enforced) reduction of low cloud opacity leads to more cloud production, at least in the early part of the day. Recall that the modification was only applied to the cloud optical thickness, as seen by the radiation scheme.

Finally, Fig. f shows impacts on horizontal winds. As already mentioned above, the fop=0.1 experiment has a stronger NLLJ developing around 18:00 UTC and lasting through the night. Only 12:00 and 15:00 UTC, when mixing is strongly increased (Fig. d), show a reduction of low-level wind speed. Above that, at the level of the African easterly jet (750–450 hPa) and at the level of the tropical easterly jet (300–150 hPa), vhoriz is markedly decreased to a signal with a relatively small diurnal cycle. One possible explanation for this finding is a reduction of wind peaks through increased convective mixing, depositing more momentum in the layer of lower background winds at 400 hPa.

Figure  shows the corresponding profiles for PARAM. Despite the overall consistent signal in rainfall and radiation as documented in Fig. , there are many substantial differences between the two sets of experiments.

Despite a larger SSI (see Fig. a), PARAM has a lower daytime increase in near-surface temperature, particularly at 15:00 and 18:00 UTC, suggesting a possible impact of the earlier triggering of convection in PARAM (see Fig. ). Near-surface qv (Fig. b) is strongly decreased at 09:00 UTC, probably due to the earlier onset of PBL mixing with transparent clouds, and then strongly increased at 12:00 and 15:00 UTC, possibly due to the lack of deep mixing as in EXPL, leading to very large differences between the two sets of experiments. Combined, the changes in temperature and moisture lead to overall less pronounced changes in RH at low levels (both negative near the surface and positive above; Fig. c), associated with mostly negative changes in qc (Fig. e) except for 09:00 UTC. These explain the somewhat unexpected results for qc discussed in the context of Figs.  and . In contrast to EXPL, PARAM operates a positive feedback mechanism through which a reduction in low cloud leads to a further reduction. This may clarify why so many climate models show very large negative biases in cloud cover (Hannak et al., 2017).

Increased vertical mixing can be observed via TKE (Fig. d). Positive signals are restricted to the low levels during the day (09:00, 12:00 and 15:00 UTC), with the latter time showing indications for increased mixing reaching mid-levels. All hours from 18:00 to 09:00 UTC show decreased TKE below 600 hPa and hardly any change at all above that. One needs to bear in mind, however, that mixing through convection is not reflected in TKE fields in PARAM. Nevertheless, the PARAM signals, at least at low levels, are in clear contrast to EXPL (Fig. d) for which TKE increases everywhere. These differences are strong indicators that the interplay between PBL turbulence and shallow and deep convection fundamentally differs between the two model configurations. Particularly during night-time, PARAM shows a slight stabilization in the temperature profile (Fig. a) above 925 hPa that appears to suppress turbulence generation in this layer. This cooling may be related to the enhanced NLLJ (Fig. f), but it is not clear why this effect does not work in EXPL, for which an even more enhanced NLLJ and also a stabilization is observed (Fig. a and f). The changes in mixing have profound impacts on many low-level fields, whereas more agreement between EXPL and PARAM is found at middle and upper levels, except for some changes in the diurnal cycle.

Overall, this discussion demonstrates the enormous importance of vertical transport and mixing in a moist tropical environment where the PBL, low clouds and deep convection are closely coupled through radiative effects.

The previous sections have revealed how moderate reductions in low-level cloud opacity can profoundly change the diurnal cycle in many meteorological variables over southern West Africa, leading amongst other things to a substantial increase in precipitation. This raises the question of to what extent these modifications have an impact on neighbouring regions or even on the entire WAM system. Does the increased precipitation over the DACCIWA box suppress precipitation to the north and south? Does this enhance or weaken the monsoon circulation? To answer these questions, we expanded the analysis of the sensitivity experiments and included the Sahel zone up to about 25∘ N.

Figure a shows zonally averaged (8∘ W–8∘ E) south–north distributions of precipitation for the ICON EXPL and PARAM experiments with fop=0.1 and fop=1.0 together with the corresponding TRMM observation, while Fig. b displays the sensitivities in absolute and relative terms. Despite the differences in EXPL and PARAM, the response to reducing the cloud optical thickness is similar, with a large increase over the modification region itself (5–10∘ N) and immediately to the north, i.e. downstream with the monsoon flow, and rather small changes elsewhere. For PARAM, differences outside of the DACCIWA box are small in both an absolute and relative sense (Fig. b). The largest differences occur in the northern half of the box with an increase of almost 3 mm day-1 corresponding to about 80 %. For fop=3.0 and fop=10.0 (not shown in Fig. ) decreases in rainfall can be observed outside of the DACCIWA box, but are almost undistinguishable from the control experiment.

Changes in EXPL are generally more dramatic. Given the drier conditions over the Guinea coastal region in the control run, the increase of almost 4.5 mm day-1 in the northern half of the DACCIWA box corresponds to 560 %, while the southern half of the box and the 2.5∘ strip to the north of it still reach increases of the order of 100 %. To the north and south of that, small decreases in absolute values are found, most likely due to an immediate suppression by the enhanced convection in the box, but these are barely significant in a relative sense (Fig. b). Finally, to the north of 17.5∘ N there is a small increase in absolute values, which, given the increasingly dry conditions in this area, corresponds to considerable relative changes.

This may suggest that modulations to the WAM allow a slightly deeper penetration of rainfall into the continent. However, given that in the northern Sahel rainfall is usually caused by few distinct, intense convective systems and that soil moisture perturbations are becoming increasingly important, 5-day simulations during 1 month are probably insufficient to make any definite statements for this area.

In order to better understand these precipitation signals, Fig.  shows corresponding south–north distributions of differences between the two EXPL runs for various meteorological quantities and their diurnal variations. Despite the relatively small impacts on precipitation, it demonstrates that the influence of the low-cloud manipulation is not restricted to the manipulated area itself (dark grey lines) but is transported northwards with the mean flow as proposed by Zheng et al. (1999). This is evident, for example, for temperature at 975 hPa, T975 (Fig. a). The near-surface heating peaks at 15:00 UTC within the box, reaching values well above 1.0 ∘C apart from the southernmost part, where inflow from the ocean creates cooling. Until 06:00 UTC the T975 signal weakens in magnitude and drifts northward out of the DACCIWA box. This change in advection (possibly in addition to radiative changes) leads to an overall moderate warming of the 10–20∘ N strip with a maximum at the end of the night. Farther to the north, there is a moderate decrease in the afternoon, likely connected to the increase in rainfall in this area (see Fig. b). The very small T975 decrease over the ocean could come from enhanced sensible heat fluxes over the cool coastal waters caused by stronger winds (see Fig. c).

The increase in low-level temperature and higher-level latent and radiative heating (see Fig. a) leads to a considerable decrease in surface pressure, psfc, peaking at 18:00 UTC with decreases of more than 0.6 hPa (Fig. b). This effect is clearly spreading downstream of the box as for T975 (Fig. a) but also upstream, likely due to upper-tropospheric flow. Given the overall north–south pressure difference of the monsoon, this signal leads to a sharpening of the gradient near the coast and a weakening towards the Sahel. The change in pressure creates a marked signal in low-level circulation, represented here by the meridional wind at 925 hPa, v925 (Fig. c). Southerly winds into and within the box are enhanced by 1 m s-1 and more, particularly leading to an increased NLLJ, while the export towards the Sahel is reduced. Assuming a similar behaviour in models with parameterized convection, these processes may explain why an underestimation of low clouds is often found together with an overestimation of NLLJ for many climate models (Knippertz et al., 2011; Hannak et al., 2017) but this needs further study. Wind signals generally tend to be smaller during the day when PBL turbulence creates a drag on the monsoon circulation (e.g. Parker et al., 2005; Marsham et al., 2013). These changes in circulation also explain the strong moisture convergence into the DACCIWA box discussed above. In addition to the meridional component shown here, there is also strongly enhanced moisture convergence in the zonal flow in response to the reduced pressure (not shown). Enhanced evaporation due to stronger winds over the ocean (Fig. c) may also make a contribution. The link between temperature, pressure and wind is further illustrated in Fig.  that shows extrema in the south–north profiles of Fig. a–c for each hour. T975 signals clearly lag the diurnal cycle of solar radiation and peak around 16:00 UTC. Due to the additional effect of latent heating by convection, the psfc minimum is reached with a delay of about 2 h. Finally, v925 is even further delayed, peaking around 22:00 UTC when the increase in pressure gradient is still large but when daytime turbulence has died down.

The response in low-level moisture, represented here by qv at 925 hPa (Fig. d), shows a relatively complicated pattern. Signals within the DACCIWA box are predominantly positive, as already discussed, showing some signs of nocturnal advection to the north similar to T975 (Fig. a). Upstream over the ocean, qv is almost unchanged but downstream values are reduced almost everywhere at all times of day with the largest differences during the night. This is unlikely a purely advective signal and is suspected to be partly caused by local vertical mixing. To further investigate this point, Fig.  shows vertical profiles of the qv signal at 00:00 and 12:00 UTC, with the DACCIWA box and the 925 hPa level indicated by grey lines. At midnight (Fig. a), when daytime convection dies down, a deep atmospheric moistening with values of up to 0.6 g kg-1 is found in the DACCIWA box and immediately to the north of it. The near-surface layer shows both positive and negative contributions. Upstream over the ocean moderate drying occurs in the 800 to 900 hPa layer, possibly related to enhanced subsidence in this area in response to the convective enhancement over land (this signal is clearly stronger at 00:00 than at 12:00 UTC). The area to the north of the DACCIWA box shows little signal above 700 hPa but an overall drying below with two local minima around 12 and 15∘ N. Where do these minima come from? A possible clue is provided by the signals at 12:00 UTC shown in Fig. b. The deeper mixing in the DACCIWA box with optically thinner low clouds creates an earlier PBL build-up, mixing moisture from lower to mid-levels, as already discussed (see Fig. b). While southern areas in the DACCIWA box receive “fresh” moisture from the ocean, the low-level dry air is advected northward with the monsoon flow and reaches 12∘ N by 00:00 UTC (Fig. a), subject to some vertical mixing. In the same way, the dry signal at 15∘ N at 00:00 UTC would originate in the DACCIWA box 36 h earlier and the dry signal at 17.5∘ N at 12:00 UTC 48 h earlier. An additional factor could be that the warmer low levels in this area (Fig. a) enhance vertical mixing and therefore entrainment of drier air into the PBL advected westward with the African easterly jet. Above this drier surface layer, the 12:00 UTC profile shows a moistening between 10 and 13∘ N that supports the idea of deeper mixing but possibly also some advection in the deep southerly monsoon flow. Through compensation, column moisture does not change much in this zone and rainfall even increases (Fig. ). From this discussion the observed small precipitation increase to the north of 17.5∘ N (Fig. ) is not clear but a more detailed investigation is beyond the scope of this paper.

Finally, we would like to address the question of how the low clouds over southern West Africa affect the overall monsoon circulation. As mentioned in the Introduction, a well-established conceptual model for this is the theoretical framework proposed by Eltahir and Gong (1996), Zheng et al. (1999), and others, which relates the strength of the circulation to the large-scale meridional gradient in equivalent potential temperature θe within the PBL, assuming sufficient deep mixing by convection (e.g. Emmanuel, 1995; Nie et al., 2010). In order to apply this idea to our sensitivity experiments, Fig. e shows south–north distributions of θe at 925 hPa as absolute values (left) and as differences (right). As described by many studies, the monsoon is related to a large θe difference of almost 20 ∘C between the Equator and about 12.5∘ N. Despite the large local impacts discussed so far, our low-cloud modifications do not perturb this large-scale gradient significantly. Upstream changes are practically negligible. In the area of the θe maximum, changes remain well below 0.5 ∘C, resulting from the increase in temperature but decrease in low-level moisture. This is considerably smaller than observed inter-annual variations of 1–2 ∘C (e.g. Hurley and Boos, 2013) and consistent with the relatively small impact on precipitation in the Sahel evident from Fig. . In the DACCIWA box itself and immediately downstream, however, the combined increase in temperature and moisture leads to θe changes of more than 1 ∘C and a strong local precipitation increase. This means that the reduction of the effective albedo over southern West Africa allows for the concentration of more energy and precipitation over land without the necessity of shifts between land areas. An interesting implication of this result is that whatever change in aerosol–radiation or aerosol–cloud interaction is caused through changes in anthropogenic emissions in the DACCIWA region, it will likely have measurable local impacts but probably no significant ramifications elsewhere.

This quite noticeable impact was found in the simulations with explicitly simulated convection. The influence of the parameterization of convection in this experiment will be discussed next. For comparison, Fig.  shows the same fields as displayed in Fig.  but for PARAM. It clearly demonstrates the substantially smaller impact of reducing the optical thickness of low clouds on low-level fields within and beyond the DACCIWA box and the substantial changes to the diurnal cycle of the differences. The temperature signal (Fig. a) has a much smaller amplitude than in EXPL and peaks earlier in the day as discussed in the context of Fig. . Due to the main advection during the night, the impact on the Sahel is even further reduced and shows a slight cooling during daytime. A similar behaviour is found for psfc (Fig. b) with a smaller and earlier peak and less impact on the Sahel than in EXPL. Given the relation of pressure and wind, it is no surprise to find a significantly reduced (or even reversed) signal in v925 also (Fig. c). These differences are further illustrated in Fig. , showing the much flatter diurnal cycle in temperature with an earlier, less pronounced peak before midday (red curves). The pressure signal (blue curves) has a larger amplitude, as it is also related to latent heating at upper levels, but due to the different timings in precipitation (Fig. ), a shift of 3 h relative to EXPL is found. With the pressure signal already decreased around sunset, the wind response is weak (see also the discussion in Marsham et al., 2013) and shows very few diurnal variations (green curves).

With respect to qv at 925 hPa, differences between PARAM and EXPL are again more complicated. There is more consistent drying over the ocean in PARAM and moistening with a similar magnitude compared to EXPL in the DACCIWA box (Fig. d), but with a much different diurnal cycle as discussed in the context of Fig. . Over the Sahel, the increase in meridional wind (Fig. c) during the night leads to a clearer signal of northward moisture advection in stark contrast to EXPL in which substantial drying is found (Fig. d). Looking at the vertical structure of these signals (Fig. ) underlines the paramount importance of vertical transport and mixing of moisture. PARAM has generally weak signals everywhere to the north of the DACCIWA box apart from the stronger low-level moisture advection at night and does not show signs of the diurnal pulses of dry advection discussed for EXPL above (Fig. c). Over the ocean to the south there is some agreement between PARAM and EXPL on a general drying of low and mid-levels. In the box itself, contrasts are extremely large at 12:00 UTC. At this time, EXPL shows effects of enhanced night-time dry advection from the ocean at low levels and moisture left over from convective mixing from the previous day above 925 hPa (Fig. b). In PARAM, convection is already active at this time, effectively removing tropospheric surplus moisture and depositing it in the PBL (Fig. d). Due to the less effective vertical transport during the day in PARAM, the moisture signal at midnight is substantially weaker in the free troposphere (see also Figs. b and b). These changes lead to an overall smaller signal in θe at 925 hPa within the DACCIWA box and to the north of it, too (Fig. e), apart from the 10–12∘ N band where nocturnal moisture advection is enhanced, as discussed above.

Finally, we also tested the time needed for the atmosphere to return to a normal state after a switch-off of the induced cloud changes in the model using the EXPL configuration. These experiments show that low-level variables such as surface radiation and temperature react almost immediately to changes in low cloud during the day. Low-level cloud cover and rainfall respond after one full diurnal cycle, while impacts on higher levels and more remote regions can last days. However, the signals hardly stand out from the high level of background variations, indicating the chaotic nature of the atmosphere. More details can be found in the Supplement.

In conclusion, this discussion shows that the parameterized treatment of convection not only affects the diurnal timing of precipitation but also strongly impacts vertical mixing. Through a number of different mechanisms, these create substantial differences in thermodynamic environments and ultimately in the sensitivity to modifications of low-level clouds, which is generally higher in EXPL than PARAM. The differences also impact the propagation of signals to the Sahel in both magnitude and diurnal timings. Despite all this, precipitation signals are clearly dominated by the DACCIWA box itself with only minor impacts outside of the box.