Nocturnal low-level clouds in the atmospheric boundary layer over southern West Africa: an observation-based analysis of conditions and processes

During the West African summer Monsoon season, extended nocturnal stratiform low-level clouds (LLC) frequently form in the atmospheric boundary layer over southern West Africa and persist long into the following day affecting the regional climate. A unique data set was gathered within the framework of the Dynamics-Aerosol-Chemistry-Cloud-Interactions in West Africa (DACCIWA) project, which allows, for the first time, for an observational analysis of the processes and parameters crucial for LLC formation. In this study, in situ and remote sensing measurements from radiosondes, ceilometer, cloud radar 5 and energy balance stations from a measurement site near Savè in Benin are analyzed amongst others for 11 nights. The aim is to study LLC characteristics, the intra-night variability of boundary layer conditions and physical processes relevant for LLC formation, as well as to assess the importance of these processes. Based on the dynamic and thermodynamic conditions in the atmospheric boundary layer we distinguish typical nocturnal phases and calculate mean profiles for the individual phases. A stable surface inversion, which forms after sunset, is eroded by differential horizontal cold air advection with the Gulf 10 of Guinea maritime inflow, a cool air mass propagating northwards from the coast in the late afternoon and the evening, and shear-generated turbulence related to a nocturnal low-level jet. The analysis of the contributions to the relative humidity changes before the LLC formation reveals that cooling in the atmospheric boundary layer is crucial to reach saturation, while specific humidity changes play a minor role. We quantify the heat budget terms and find that about 50 % of the cooling prior to LLC formation is caused by horizontal cold air advection, roughly 20 % by radiative flux divergence and about 22 % by 15 sensible heat flux divergence in the presence of a low-level jet. The outcomes of this study contribute to the development of a conceptual model on LLC formation, maintenance and dissolution over southern West Africa. Copyright statement. TEXT


Introduction
Nocturnal stratiform low-level clouds (LLC) frequently form in the atmospheric boundary layer (ABL) over southern West Africa during the West African summer Monsoon season. These LLC have typical cloud-base heights (CBH) of only a few hundred meters above ground (Schrage and Fink, 2012;Kalthoff et al., 2018) and cover an area of about 800 000 km −2 (van der Linden et al., 2015). The LLC form during the night and persist long into the following day . They thus 5 affect the energy balance at the Earth's surface, the diurnal cycle of the ABL and the regional climate (Knippertz et al., 2011;Hannak et al., 2017). Numerical weather prediction and climate models still struggle to correctly represent the West African monsoon (Hannak et al., 2017), which may be related to the erroneous representation of the LLC. A profound and accurate understanding of the processes relevant for the formation, maintenance and dissolution of the LLC might help to identify flaws in these models. 10 Most of our knowledge on the processes relevant for the LLC is based on numerical simulations, as high-quality observational data are scarce in this region. Studies by Schrage and Fink (2012), Schuster et al. (2013), Adler et al. (2017) and Deetz et al. (2018) suggest that processes spanning from the microscale to the synoptic scale are important such as horizontal cold air advection, orographic lifting, lifting related to gravity waves and shear-generated vertical mixing underneath the axis of a nocturnal low-level jet (LLJ). There is evidence that the horizontal cold air advection is related to the south-westerly Monsoon 15 flow which transports maritime air from the Gulf of Guinea northwards over land. The daytime conditions at the coast are characterized by a superposition of the Monsoon flow and a sea breeze. During the Monsoon season, the strong south-westerly Monsoon flow dominates and makes it difficult to distinguish between both (Bajamgnigni Gbambie and Steyn, 2013). During the day, the south-westerly flow is decelerated over land leading to convergence along a line parallel to the coast often associated with moist convection Parker et al., 2017) and rainfall (Maranan et al., 2018). Model studies 20 indicate that the convergence zone which separates the cool maritime air in the south from the warmer air in the convective ABL over land is rather stationary and located at several tens of kilometers distance from the coast until the late afternoon (Adler et al., 2017;Deetz et al., 2018). In the late afternoon and early evening, the cool maritime air starts to propagate inland and reaches distances of more than 100 km. Grams et al. (2010) investigate a sea breeze front which is stationary at the coast of Mauritania during daytime and propagates inland in the evening. These authors relate the stationarity to a balance between 25 horizontal advection of cool maritime air and turbulence in the convective ABL over land and the inland propagation to the decay of turbulence in the late afternoon. As the maritime air which propagates northwards in our investigation area in southern West Africa has its origin over the Gulf of Guinea, we call this feature Gulf of Guinea maritime inflow and refer to it as maritime inflow hereafter for the sake of brevity.
To evaluate the hypotheses from previous studies and to enhance our understanding of the physical processes relevant for 30 LLC, high-quality comprehensive observations in the coastal region of southern West Africa were urgently needed. Therefore, a concerted measurement campaign was conducted during the summer Monsoon season in June and July 2016 within the framework of the Dynamics-Aerosol-Chemistry-Cloud Interactions in West Africa (DACCIWA) project (Knippertz et al., 2015). The meteorological measurements were airborne (Flamant et al., 2018) and ground-based at three supersites in Ghana, Benin and Nigeria  and depict the most comprehensive data set for this region so far. With respect to climatological conditions, the period of the field campaign was characterized by Pacific La Niña and Atlantic El Niño and overall average rainfall across the whole of West Africa .
Based on the observational data gathered at the ground-based supersite in Benin (Savè, location in Fig. 1), which is the supersite with the most comprehensive instrumentation, a series of analysis have been conducted on the LLC: Babić et al. 5 (2018) present -for the first time -a detailed analysis of the diurnal cycle of LLC for a case study of a typical night (7-8 July, intensive observation period (IOP) 8) and identify physical processes and factors which control the formation, maintenance and dissolution of LLC at Savè. Based on the dynamic and thermodynamic conditions in the ABL during IOP 8, these authors identify different phases, which are outlined in Fig. 2: the stable phase describes a period after sunset when the horizontal wind is weak and a surface inversion forms. With the arrival of the maritime inflow a few hours after sunset, dynamic and 10 thermodynamic conditions in the ABL change and are then characterized by a LLJ wind profile (jet phase). Eventually, LLC form (stratus phase). About 1 h after sunrise, the cloud base starts rising due to the evolution of the convective ABL (convective phase). Babić et al. (2018) determine the contributions of temperature and specific humidity changes to the relative humidity changes and quantify the heat budget terms for the different phases during IOP 8 using radiosoundings.  perform a detailed statistical analysis on the characteristics of the LLC and the low-level atmospheric dynamics using mainly 15 data from continuously running remote sensing instruments for a 41-day period. While the studies of Babić et al. (2018) and  either look at the diurnal cycle during one case study or at mean quantities during a longer measurement period, the present analysis focuses on 11 IOP nights. As radiosoundings were performed in short temporal intervals of 1 to 1.5 h throughout the IOP nights, high-quality profile information on temperature, specific humidity and horizontal wind are available, which allows us to perform an analysis for these nights in a manner consistent with the methods used by Babić et al. 20 (2018). By analyzing several nights -instead of one -we are able to address the representativeness of the results obtained from the single case study and to generalize some of the findings. Besides the generalization of process relevance for the formation of LLC, we aim to characterize the LLC and to investigate the intra-night variability of ABL conditions. The research questions to be answered are: (i) What are the temporal and spatial characteristics of LLC? (ii) How do ABL conditions change during the different nocturnal phases? (iii) What dominates the relative humidity changes and heat budget? (iv) How do the processes 25 involved in (iii) vary with height and from night to night?
In Sect. 2, the observational data used and the methods applied to derive different LLC characteristics and to estimate the heat budget terms are described. Section 3 includes LLC characteristics and Sect. 4 the conditions in the nocturnal ABL for the different phases. In Sect. 5, relative humidity changes and heat budgets terms are investigated and processes resulting in LLC formation are assessed. In Sect. 6, the observed processes are discussed in comparison to recent studies, followed by a 30 summary and conclusions in Sect. 7.
3 2 Data used and methods

Measurement site and instrumentation
The supersite at Savè (166 meters above mean sea level (m m.s.l.)) was located in Benin about 185 km inland from the coast (Fig. 1b). The terrain in the immediate surrounding of the site is rather flat, while higher terrain up to 500 m m.s.l. is found to the north and east of it. During the 7 weeks of the ground-based measurement campaign comprehensive in situ and remote 5 sensing measurements were conducted. A detailed overview of the atmospheric conditions during the whole campaign period is given by Kalthoff et al. (2018) and the complete instrumentation including information on the manufacturers is described by Bessardon et al. (2018). In this study, we use data from the continuously running ceilometer (measuring backscatter profiles from which information on CBH is derived) and cloud radar (measuring radar reflectivity profiles from which information on cloud-top heights (CTH) is derived) to obtain information on the LLC characteristics at Savè. Spatio-temporal information on 10 LLC in a larger area around Savè (0-4°E and 5.5-10°N, dashed box in Fig. 1b) is obtained from the Spinning Enhanced Visible and Infrared Imager (SEVIRI) sensor (Schmetz et al., 2002). Two energy balance stations provide the near-surface radiation and energy balances and meteorological parameters over two types of land use (fallow and corn) and an ultra-high-frequency (UHF) wind profiler and a Doppler lidar are used to obtain information on the horizontal and vertical wind. Ceilometer, cloud radar, one of the energy balance stations and the Doppler lidar are part of the mobile integrated atmospheric observation system 15 "KITcube" deployed by the Karlsruhe Institute of Technology (Kalthoff et al., 2013).
During the 7 weeks of the campaign, 15 IOPs in total were conducted (Table 1). Every IOP lasted from the late afternoon throughout the night until the afternoon of the following day to capture the whole diurnal cycle of the LLC. Although we aimed to perform IOPs during synoptically undisturbed nights without any mesoscale convective systems, the conditions during 3 of the 15 IOPs (IOPs 2, 12 and 13) were disturbed by rain at or near the supersite, preventing the evolution of "typical" LLC. This 20 is why we exclude these IOPs from this analysis. As no LLC existed during IOP 10 at Savè, this leaves 11 IOPs for the analysis of the relationship between ABL conditions and LLC (the used IOPs are indicated in Table 1). IOP 1 falls into the pre-onset phase of the monoon and the other ten IOPs into phases during which the monsoon conditions are relatively undisturbed and the strength and position of the Saharan heat low and African easterly jet are close to the climatological average . During IOPs, radiosondes were released at Savè one hour before the nominal times at 0000, 0600, 1200 and 25 1800 UTC. The first radiosounding was released at 1700 UTC (the local standard time in Benin is UTC+1), in order to be synchronous with the soundings at operational radiosonde stations. In between, so-called frequent radiosondes were launched reaching maximum heights of around 1500 meters above ground level (m a.g.l.) to get a higher temporal resolution of the ABL conditions. These sondes are attached to two ballons of different volume, whereas the line to the larger balloon is cut after a preset time of ascent (corresponding to around 1500 m a.g.l.) initiating a controlled descent of the sonde (Legain et al., 2013).

30
This method allows a re-use of the sonde and short sounding intervals. During IOP 1-6, frequent radiosondes were launched in hourly intervals starting at 2100 UTC. After IOP 7, the sounding schedule was changed to 1.5-h intervals starting at 1830 UTC to better resolve the early evolution of the nocturnal ABL. An impression of the radiosonde schedule during individual IOPs can be obtained from Fig. 3 with each coloured column indicating one sounding centered on the launch time.
In addition and as part of the DACCIWA radiosonde campaign (Flamant et al., 2018) radiosoundings in up to 6-h intervals were launched at three stations along the coast, at Abdijan (Ivory Coast), Accra (Ghana) and Cotonou (Benin) (locations are given in Fig. 1b).

LLC characteristics
2.2.1 Cloud-base and cloud-top height detection 5 From the ceilometer backscatter profiles with 1-min resolution, up to three CBH are obtained using the manufacturer algorithm which is based on a threshold method (we only use the lowest CBH in this study). The cloud radar deployed at Savè is dual-polarized, i.e. it is possible to distinguish between hydrometeors and other targets. The classification uses the linear depolarization ratio and multi-peak moments and is described in detail in Bauer-Pfundstein and Görsdorf (2007). From the radar reflectivity of hydrometeors we estimate the CTH for 5-min averaged reflectivity profiles applying a threshold of -10 35 dBz, i.e. reflectivities larger than -35 dBz are considered as clouds. As the reflectivity changes abruptly near CTH, the determined CTH are not sensitive to the applied threshold value. Examples of ceilometer backscatter profiles and cloud radar reflectivity profiles with estimated CBH and CTH are shown in Fig. 3 in Babić et al. (2018). Beside information on dynamic and thermodynamic conditions, estimates of the vertical extent of clouds can be obtained from relative humidity profiles measured with radiosondes. The algorithm uses a threshold of 99 % and is described in Kalthoff et al. (2018). The CTH used in this study 15 are a combination of estimates from cloud radar and radiosonde measurements. During some soundings, discrepancies occur between CTH estimated from cloud radar and radiosonde measurements which are likely related to condensation of water vapour on the sensors when the radiosonde flies through a cloud (more details on this are given in Babić et al. (2018)).

Estimation of LLC onset time
Before estimating onset times of LLC from ceilometer a clear definition of LLC must be stated. Do we require a constant CBH 20 and a complete coverage of the sky or do we allow some variability in CBH and coverage? From the overview of CBH in Kalthoff et al. (2018), we expect LLC bases mainly to occur in the lower 600 m and consequently choose this layer to estimate cloud-base fraction from ceilometer. As CBH are available every minute from the ceilometer we calculate the cloud-base fraction every minute for the subsequent 60 min. The onset time is then defined as the first point in time when a cloud-base is detected in the lower 600 m and the subsequent cloud-base fraction is higher than a respective threshold. During all IOP 25 nights cloud-base fractions reach 100 %. During some nights, cloud-base fraction shifts almost instantly from 0 % to 100 %, while during other nights, lower cloud-base fractions precede the complete coverage by several hours. To take this into account, we introduce the stratus fractus phase before the stratus phase, in addition to the phases identified by Babić et al. (2018) for IOP 8 (Fig. 2). We choose two thresholds, 50 and 95 %, to detect the onset of stratus fractus and stratus, respectively. For the estimation of tendencies and contributions we look at the period before the onset of stratus fractus, to avoid the impact of phase 30 changes, while we investigate the modification of the ABL conditions by the LLC for the stratus phase only, as we expect a clearer signal from this phase. Detecting LLC during the night from the geostationary SEVIRI sensor is challenging, as the temperature at the cloud top is very close to the surface temperature in cloud-free regions, making them nearly indistinguishable from the surface in the infrared channels. In the absence of high or mid-level clouds, which are obscuring the LLC from the view of the satellite-borne sensor, the brightness temperature difference of the thermal-infrared channel at 10.8 µm and the middle-infrared channel at 3.9 5 µm are used to visualize the LLC during the night. Higher-level clouds are masked out by applying a brightness temperature threshold of 283 K to the 10.8 µm channel, i.e. cloud tops above around 2500 m m.s.l. are masked. More details on the method are given in Babić et al. (2018).

Estimation of heat budget terms
The tendency of the mean potential temperature (TOT) is generally influenced by various processes such as horizontal (HADV) 10 and vertical (VADV) advection, radiative flux divergence (RAD), sensible heat flux divergence (TURB) and phase changes (SQ) (e.g. Stull, 1988): with the mean potential temperature, Θ, the mean wind components, u, v and w, the mean air density, ρ, the specific heat capacity at constant pressure, c p , the sensible heat flux, H, the latent heat of vaporization of water, L, and the evaporation 15 rate E. The estimation of the different heat budget terms from observations is challenging and requires some assumptions.
The easiest term to derive is TOT, which we calculate directly from radiosonde profiles. As we only consider periods without LLC, we can neglect SQ in the budget. Sun et al. (2003) calculate radiative and sensible heat flux divergence from tower measurements for the nocturnal ABL and find that even small vertical or horizontal temperature difference can contribute significantly to TOT. The measurement accuracy of mean vertical velocity required to estimate VADV is hardly achieved by in 20 situ wind measurements and even less by remote sensing instruments. This is why we cannot estimate VADV in this study. This leaves us with HADV, RAD and TURB, for which we now describe the methods we used for their estimations. The estimations are done for different periods (Fig. 2). The period for HADV comprises the stable and jet phases, while TURB is calculated during the jet phase only and RAD and TOT are estimated for both time periods.

Horizontal temperature advection 25
The large-scale dynamic and thermodynamic conditions in the investigation area in southern West Africa are characterized by the south-westerly Monsoon flow and a meridional temperature gradient with lower temperatures over the Gulf of Guinea.
As outlined in the introduction, there is evidence that cool maritime air is transported further inland with the Monsoon flow during the late afternoon and night -a feature which we call maritime inflow. During the IOP days studied here, we detect an concurrent increase of wind speed with the profile showing a LLJ structure and a decrease of temperature in a layer of several 30 hundred metres depth during the first half of the night, which we interpret as the arrival of the maritime inflow at Savè. The changes in atmospheric conditions during this arrival are nicely illustrated for IOP 8 in Babić et al. (2018). In the present study, we use radiosoundings to detect the changes in temperature and wind profiles. We manually distinguish between the different phases and allocate the radiosonde profiles to the phases. Note that  find some differences in the onset times of the LLJ and the maritime inflow when analysing the whole campaign period. These differences may depend on the higher 5 temporal resolution of the data from continuous remote sensing instruments and on the criteria applied by these authors to detect the onset times.
In order to estimate horizontal cold air advection related to the maritime inflow from the available observations, several assumptions are necessary, which are illustrated in the schematic diagram in Fig. 4. To estimate the meridional temperature difference, we use radiosoundings performed at the three coastal stations and at Savè in the late afternoon (station locations in 10 Fig. 1b). We assume that the temperature distribution is homogeneous along the coast and that the zonal temperature gradient and wind component are small and make the following estimates for a meridional cross section through Savè. The difference of the mean temperature in the late afternoon between the coast and Savè is more than 3 K on the average (Table 1). Information on the daytime meridional temperature distribution between Savè and the coast are obtained from aircraft flights, when horizontal legs were flown in the ABL at some angle to the coast line in the afternoon (not shown). These measurements indicate that 15 the aircraft passes through the maritime inflow and through the convective ABL over the land during these flights. In the convective ABL, the temperature is rather horizontally homogeneous, while it decreases gradually towards the coast within a certain distance from the coast. We interpret this distance as the maximum inland propagation of the maritime inflow in the late afternoon, i.e. as the location of the convergence zone which separates the cool maritime air in the south from the warmer air in the convective ABL, (right edge of yellow box in Fig. 4). This means that the temperature decreases gradually within the 20 maritime inflow and we assume a linear increase of temperature in south-north direction in the maritime inflow and a constant temperature in the convective ABL (red curve in Fig. 4). Motivated by model studies by Adler et al. (2017) and Deetz et al. (2018), we expect the maximum inland propagation to be somewhere between 50 and 125 km inland from the coast in the late afternoon.
For the cooler air of the maritime inflow to be able to produce a cooling at Savè, the maritime inflow has to propagate far 25 enough inland to reach the site. To estimate the maximum distance from which the air mass measured at Savè before the LLC onset may originate (left edge of the dashed area in Fig. 4), we estimate the propagation speed from radiosoundings. Therefore, we average the meridional component of the coastal wind profile in the afternoon and of the wind profile at Savè after the maritime inflow arrives, v, and assume that the maritime inflow propagates with the maximum southerly wind component of the averaged vertical profile. If the propagation speed is high enough and the LLC onset late enough, the air in the maritime 30 inflow is able to reach Savè and to contribute to the cooling before the LLC onset. For example, assuming a propagation speed of 5 m s −1 , a LLC onset at Savè at midnight and a start time of the further inland propagation of the maritime inflow at 1600 UTC, the air mass passing Savè originates at a maximum distance of 41 km inland from the coast. With an assumed maximum inland propagation of the maritime inflow of 75 km during the day, parts of the air mass passing Savè before the LLC onset are of maritime origin and have a ∆Θ lower temperature than the air in the convective ABL. Based on the assumptions for the 35 maximum distance and the linear temperature change in the maritime inflow ( Fig. 4), HADV is calculated as −v∆Θ ∆y −1 using data from all three coastal stations (Abdijan, Accra and Cotonou), if available, and four different maximum propagation distances (50, 75, 100 and 125 km).

Radiative flux divergence
To estimate the contribution of RAD to the temperature tendency at Savè, we apply the Santa Barbara DISORT Atmospheric 5 Radiative Transfer (SBDART) model (Ricchiazzi et al., 1998) to the mean profile averaged for the time period between 1700 UTC and the formation of clouds and for the jet phase (Fig. 2) using a mean aerosol optical depth measured with the sun photometer at Savè and averaged for the months June and July 2016. More details on the radiative transfer model and on the used input parameters can be found in Babić et al. (2018).

Turbulent heat flux divergence
10 During the night, the absolute value of the sensible heat flux is usually at its maximum at the surface and decreases with height (e.g. Sun et al., 2003). To estimate TURB from the existing measurements we need to make an assumption for the height where the sensible heat flux vanishes. In the presence of a surface inversion, this height is usually assumed to be the top of the surface inversion. If a LLJ is present, we assume that the sensible heat flux vanishes at the height of the LLJ axis (as vertical wind shear vanishes). Both cases are present at Savè during the stable phase and jet phase, respectively. Due to the insufficient number 15 of radiosonde profiles during the stable phase (in particular during the IOPs 1-6), we estimate TURB for the jet phase only (Fig. 2). For this estimation we use measurements of the surface sensible heat flux at both energy balance stations and vertical profiles of horizontal wind to detect the height of the LLJ axis. The surface sensible heat flux is averaged over the considered time period and the bulk cooling by TURB is then estimated for the layer below the LLJ axis.

LLC characteristics 20
This section overviews the vertical distribution and temporal evolution of LLC at Savè using ground-based remote sensing instruments and the horizontal distribution of LLC using satellite images.
The combination of ceilometer and cloud-radar measurements allows us to obtain information on the vertical extent of the LLC. In Fig. 3, cloud base from ceilometer and cloud top from cloud-radar measurments are shown as red dots and circles, respectively, and Fig. 5a contains statistical information on CBH and CTH during the stratus phase for the individual IOPs. 25 The median CBH ranges from 70 m to 450 m a.g.l. and the median CTH from 370 to 870 m a.g.l. When averaging the median heights and the vertical extents for all IOPs, CBH = 250±120 m a.g.l., CTH = 590± 170 m a.g.l. and a vertical extent of 340±80 m result. This confirms the vertical extent of simulated LLC (Schuster et al., 2013;Adler et al., 2017) and agrees with the median CBH and CTH pointed out by  using all days of the DACCIWA campaign. During IOP 1, 5 and 6, the median CBH is below 130 m a.g.l., while it is above 200 m a.g.l. for the other IOPs. This is in agreement with the two 30 layers found favourable for CBH occurrence at Savè by Kalthoff et al. (2018)  i. During IOP 3, LLC form already at 2200 UTC over the higher terrain north and north-east of Savè (Fig. 1b) and remain 20 more or less stationary until around 0330 UTC expanding only a little (Fig. 6a). After that, the LLC suddenly start to expand to the south-west until they cover most of the domain including Savè. This is in agreement with the onset of LLC at around 0400 UTC at Savè (Figs. 3b and 5b).
ii. In the evening of IOP 4, some mid-and high-level clouds obscure the lower levels in many parts of the domain, but it is nevertheless evident that LLC exist already at 2100 UTC east of Savè over the higher terrain (Fig. 6b). In the following 25 hours the mid-and high-level clouds move westwards allowing for the detection of the growth of the LLC towards the west affecting Savè. These LLC have quite scattered cloud bases (Fig. 3c). Between around 0030 and 0230 UTC a spatial gap occurs in the LLC deck right above Savè (Figs. 3c and 6b). After that the LLC cover most of the domain for the rest of the night (Fig. 6b). During this period, the CBH are rather homogeneous, which is visible in the ceilometer measurements at Savè (Fig. 3c).

30
iii. During IOP 7, the domain is mostly cloud-free until 2300 UTC (Fig. 6c). Then, LLC start forming at several locations in the domain and grow during subsequent hours, also occuring at Savè after 0000 UTC (Fig. 3f). After around 0100 UTC, high-level clouds in a layer between around 11 and 13 km a.g.l. (as visible in cloud-radar measurements at Savè) gradually move in from the east and cover the domain preventing the analysis of the further evolution of the LLC.
iv. During IOP 8, LLC first form south-west and east of Savè after around 2200 UTC (Fig. 6d). In the following hours both patches of LLC expand, the one in the south-west reaching Savè at midnight (Fig. 3g). After 0030 UTC both patches grow together. Subsequently, the LLC expand in all directions and they cover most of the domain at sunrise.
Overall, we can identify two types of horizontal LLC expansion: during IOPs 3, 4 and 8, the LLC grow to the upstream side, i.e. towards the direction of the mean south-westerly flow, and to the downstream side, i.e. away from the direction of the mean 5 flow. During IOP 7, LLC form and grow at many locations at the same time with no clear direction being distinguishable. This suggests that different mechanisms are involved, some of those are assessed in Sect. 5.3.

Intra-night variability of ABL conditions
This section characterizes the mean nocturnal ABL conditions during different phases of the night -these are the LLC-free stable and jet phases and the stratus phase (Fig. 2). The profiles which we use for the averaging during the individual phases 10 are indicated at the bottom of each plot in Fig. 3 by the squares. To take into account the large variability of CBH during the individual IOPs (Fig. 5a), we normalize the profiles with the median CBH of each IOP (z/CBH). Despite the large day-to-day variability in the wind speed profiles (Fig. 3), there is a clear signal in the mean normalized profiles (Fig. 7a): during the stable phase, the mean wind speed is lowest with around 3 m s −1 and little variation with height, while the LLJ shape is clearly visible during the jet and stratus phases. During the jet phase, the LLJ axis with a maximum value of more than 7 m s −1 is near 15 the height where cloud bases exist later on. In the presence of LLC, the LLJ axis shifts upwards to around z/CBH=1.6 and the maximum decreases by about 1 m s −1 . The mean potential temperature decreases in the course of the night leading to an up to 4 K cooler atmosphere during the jet phase than during the stable phase and a decrease of about 1 K from the jet to the stratus phase (Fig. 7b). The strongest changes occur up to around z/CBH=2. Specific humidity changes little during the stable phase and jet phase, while it is about 1 g kg −1 lower during the stratus phase (Fig. 7c). The stable phase is characterized by a shallow 20 surface inversion and a weakly stably stratified residual layer above (Figs. 7b and d). During the jet phase the surface inversion is eroded and static stability becomes more constant with height but has quite a large standard deviation. During the stratus phase, static stability below and also above the CBH decreases compared to the jet phase. This is likely driven by longwave cooling at cloud top, which leads to mixing. The top of the layer with reduced stability coincides with the LLJ axis during the stratus phase. This suggests that the upward shift of the inversion due to the reduced stability in the presence of clouds causes 25 the LLJ axis to shift upwards, which agrees with the results of Babić et al. (2018) and .
5 Changes in atmospheric conditions and processes leading to LLC formation

Relative humidity changes
For LLC to form, the ABL has to be saturated, i.e. relative humidity has to reach 100 %. In order to understand why saturation is reached, i.e. due to cooling or an increase of specific humidity, we calculate the contribution of temperature and specific 30 humidity changes to the relative humidity changes between the late afternoon and just before the onset of LLC (the formula is given in Babić et al. (2018) and the time interval for which the changes are calculated is indicated in Fig. 2). For this we use the radiosonde profiles at 1700 UTC and the last profile before the LLC onset at Savè, i.e. the elapsed time varies between 4 h for IOPs 4, 9, 11 and 14 and 10 h during IOPs 1, 3 and 15. The mean profiles of relative humidity changes and the individual contributions for this period are shown in Fig. 8a. On the average, relative humidity increased by about 25 % near the surface decreasing linearly with height up to around 750 m a.g.l. The shape of this profile is mainly due to the linear increase of relative 5 humidity with height in the ABL in the late afternoon. Nearly 100 % of the relative humidity increase is related to cooling, while specific humidity changes contribute only little (< 3 %, Fig. 8a).
To investigate this further, we split the period in two parts indicated in Fig. 2: one period between the late afternoon and just after the arrival of the maritime inflow (P1) and one period after the arrival and before the LLC onset (P2). During P1, 5 % of the relative humidity increase (which is about one quarter of the total increase) in the lower layer is caused by an increase of 10 specific humidity (Fig. 8b), while a decrease of specific humidity lowers the relative humidity increase by about 5 % during P2 (Fig. 8c). This means that an increase of specific humidity contributes to the relative humidity increase before and at the arrival of the maritime inflow at Savè. Once Savè is within the maritime inflow air mass, specific humidity decreases working against the cooling with respect to the relative humidity change. Independent of the considered period, cooling is crucial to achieve saturation in the nocturnal ABL. This is in qualitative agreement with the results for IOP 8 in Babić et al. (2018). The 15 fact that specific humidity changes play a minor role might come a little unexpected. This is likely related to the relatively low sea surface temperature of the Gulf of Guinea limiting the specific humidity in the maritime ABL and high evapotranspiration from the dense vegetation over land leading to high specific humidity in the continental ABL.

Heat budget terms
From the analysis of relative humidity changes in the previous section, we know that cooling is the key factor for LLC forma-20 tion. Consequently, we continue with an analysis of the different terms of the heat budget at Savè (Eq. 1). We start with the estimation of TOT, RAD and HADV for the period from the late afternoon until the onset of LLC, followed by an analysis of TURB during the jet phase (different periods are filed in Fig. 2) .

Heat budget estimates for the period from the late afternoon until LLC onset
We calculate the profiles of TOT between the 1700 UTC radiosonde profile and the last profile before the LLC occur at Savè 25 and average this over all IOPs (black curve in Fig. 9a). The cooling is strongest near the surface and decreases more or less linearly up to around 750 m which is consistent with the profiles of relative humidity changes (Fig. 8a). The variability of RAD between the individual IOPs is rather small and the mean cooling rates are in the order of -0.15 K h −1 near the surface increasing with height to around -0.1 K h −1 (green line in Fig. 9a). The mean profile of HADV averaged over all IOPs (red line) resembles the horizontal wind profiles used for the calculations with maximum values between 200 and 400 m a.g.l., i.e. To investigate the night-to-night variability of TOT, RAD and HADV, we vertically average the different terms up to the level where TOT becomes larger than TOT max e −1 with e being Euler's number and TOT max the maximum cooling for each IOP (mean and median of this level are 530 and 475 m a.g.l., respectively). The absolute vertically averaged cooling rates are shown in Fig. 9b. TOT varies between around 0.4 and 0.8 K h −1 , while the variability of RAD is small and absolute values are around 0.1 K h −1 as expected from the profiles. HADV varies considerably and reaches values from around 0.1 to 0.4 K h −1 .

5
The standard deviation of HADV (indicated by the orange errorbars) which results from the usage of the three coastal stations and four different maximum propagation distances reaches up to around 60 % of the mean HADV. The relative contributions of RAD and HADV to TOT vary betwen 13-29 % and 26-76 % for individual IOPs. We calculate the mean and standard deviation of the relative contributions for all IOPs and find that RAD can explain about 21±4 % and HADV about 50±17 % of the observed cooling rates on the average. This means that about 30 % of the cooling between the late afternoon and the LLC onset 10 are caused by other processes. From the mean profiles in Fig. 9a it is evident that most of the missing cooling occurs below around 400 m a.g.l. We expect that a large part of the missing cooling is related to TURB, which is further investigated in the following section.

Sensible heat flux divergence during the jet phase
The period before the LLC form can roughly be divided into two phases, i.e. the stable phase and the jet phase (Fig. 2). These 15 phases are illustrated using the example of IOP 15 (Fig. 10): the stable phase lasting from around sunset to 2000 UTC is characterized by the evolution of a shallow surface inversion (Fig 10a), relatively weak winds (Fig. 10a, b), low TKE values near the surface (Fig. 10c), strong stability as indicated by large Flux and Bulk-Richardson numbers (Fig. 10d) and small negative sensible heat fluxes (Fig. 10e). The Bulk-Richardson number is calculated for the layer between the surface and 200 m a.g.l. as this depicts the height of the mean wind speed maximum (not shown). With the arrival of the maritime inflow 20 and the embedded LLJ at around 2000 UTC, differential cooling occurs below around 500 m a.g.l., which is strongest between around 100 to 300 m a.g.l. (Fig. 10a) and reduces the static stability. At the same time wind speed in the residual layer increases sharply with a LLJ axis near 250 m a.g.l., which increases the dynamically induced turbulence. Both processes lead to an abrupt decrease in Richardson number to values close to 0 (Fig. 10d). Simultaneously, near-surface TKE increases rapidly (Fig. 10c) and surface sensible heat flux decreases to values of around -20 W m −2 (Fig. 10e). This marks the onset of the jet phase. The 25 concurrent LLJ and high near-surface TKE values agrees with observations at Nangatchori in central Benin (Lothon et al., 2008). The jet phase lasts until LLC form at around 0330 UTC (Fig. 10a). After around midnight TKE decreases (Fig. 10c) and stability slightly increases (Fig. 10d) which is likely related to an upward shift of the LLJ axis to around 400 m a.g.l. (Fig.   10a), which reduces the vertical wind shear below the LLJ axis. The measurements during IOP 15 show a relationship between the Richardson number and near-surface TKE. This is also the case for the other IOPs: near-surface TKE is clearly related to 30 the Flux-Richardson number and the Bulk-Richardson number (Fig. 11). High TKE values (larger than 0.3 m 2 s −2 ) only occur when the Richardson numbers are below 0.1 indicating a regime where turbulence is dynamically generated. The correlation between low Bulk-Richardson numbers and high TKE values indicates that turbulence is generated up to at least 200 m a.g.l.
The comparison of the temporal evolution of the potential temperature tendency profiles for IOP 15 (Fig. 10a) with the mean temperature tendency between late afternoon and LLC onset (Fig. 9a) suggests that a large part of the cooling below the LLJ axis happens during the jet phase (except for the shallow surface inversion which forms during the stable phase). To estimate the heat budget terms, we use a time period within the jet phase confined by the first and last radiosounding. This requires at least two radiosoundings within the jet phase, which is not the case for IOP 4 and 14, excluding these IOPs from this analysis. 5 As the cooling due to HADV with the maritime inflow already starts before the considered time period (Fig. 2), we cannot apply the assumption for HADV from Sect. 2.3.1 to estimate HADV for this period and only calculate TOT, RAD and TURB.
As described in Sect. 2.3.3, we estimate TURB for the layer below the LLJ axis using surface sensible heat flux values from both energy balance stations. TOT and RAD are vertically averaged up to the height of the LLJ axis. When calculating mean and standard deviation of the relative contribution of RAD and TURB to TOT for all IOPs -like we did in Sect. 5.2.1 -we 10 find that each RAD and TURB explain about 22±10 % of TOT (Fig. 12). We speculate that a large part of the residual during this period is related to horizontal advection with the maritime inflow. Furthermore, we expect vertical advection (which we cannot estimate) and horizontal advection related to temperature difference on a regional scale (which we also cannot estimate from the existing measurements) to contribute to TOT as well.

15
While horizontal cold air advection contributes significantly to the cooling of the ABL and is thus a necessary pre-condition for LLC (Sect. 5.1 and 5.2), the LLC themselves are not necessarily advected as seen in the satellite images (Sect. 3), but also form locally. We assess the possibility of three mechanisms triggering the LLC: during some nights, LLC form over higher terrain, suggesting that orographic lifting constitutes the final trigger mechanism, which confirms the results of numerical simulations by Schuster et al. (2013) and Adler et al. (2017). 20 The second possible mechanism is related to shear in the nocturnal ABL. Zhu et al. (2001) study how the formation of sheardriven idealized clouds in the nocturnal ABL is related to land surface and ABL processes using a simple well-mixed boundary layer theory. In the presence of vertical wind shear, these authors divide the nocturnal ABL into three parts -a surface layer, a mixed layer and a transition zone at the top of the nocturnal ABL. As long as the well-mixed conditions are met, a relationship between the nocturnal clouds and the lifting condensation level (LCL) calculated from surface data exists, i.e. the cloud base 25 is around the LCL. Zhu et al. (2001) assume that the nocturnal clouds form due to processes related to the land surface and the ABL, when the LCL is lower than the nocturnal ABL height, i.e. when the LCL is within the mixed layer. We calculate the LCL from air temperature, T , and dew point temperature, T d , measurements at 2 m a.g.l. with the formula LCL= 125(T −T d ), and compare it to the observed CBH. We find two distinct types: during IOPs 1, 5, 6 and 8 the CBH agrees pretty well with the LCL, while during the other IOPs the CBH is up to several 100 m higher than the LCL (examples for both types are shown 30 in Figs. 13a and b). From radiosoundings performed during the stratus phase we calculate the Bulk-Richardson number for the sub-cloud layer as an indicator for turbulent mixing and find a relationship between the stability in the sub-cloud layer and the difference between CBH and LCL (Fig. 13c). When the CBH coincides with the LCL, Bulk-Richardson numbers are very small or even negative. This indicates that during nights when the CBH equals the LCL the sub-cloud layer is near-neutrally stratified and the sub-cloud layer is coupled to the surface, i.e. according to Zhu et al. (2001), the LLC are triggered by shearrelated ABL processes for these cases. Interestingly, all IOP nights with low CBH, i.e. below 130 m a.g.l. (Sect. 3), are also nights when the sub-cloud layer is coupled to the surface. One the other hand, when the CBH is larger than the LCL, there is no coupling and other processes must be responsible for the triggering of LLC.
The third mechanism relates to the expansion of LLC to the upstream side, i.e. when new LLC are triggered upstream of 5 existing LLC. This upstream expansion of LLC also occurs in the numerical simulations by Adler et al. (2017). These authors identify a growth mechanism related to a modification in the stratification and LLJ profile in the model: in areas covered with LLC, maximum static stability and the LLJ axis occur in the upper part of the LLC layer and are thus shifted upwards compared to cloud-free areas (where the LLJ axis is near CBH). This results in horizontal convergence upstream of existing clouds and enhanced upward motion which causes additional cooling and triggers new clouds. We find observational evidence 10 for this mechanism in the temperature and wind profiles from radiosoundings: the mean profiles reveal that static stability below and in the LLC decreases during the stratus phase compared to the jet phase (Figs. 7b, d), which we assume to represent the conditions in the cloud free areas. When looking at individual nights it is also evident that static stability increases around Although we find observational evidence for all three possible mechanisms, their further analysis requires spatial information 20 on the dynamic and thermodynamic conditions, which are not available from the observations at Savè, but could be investigated by combining the observations with numerical simulations.

Discussion
We discuss and relate the observed processes relevant for LLC formation with the processes postulated in former numerical studies. From the radiosonde data we estimate different terms of the heat budget and find that HADV with the maritime inflow   Schuster et al. (2013) found a cooling of around 5 K (12 h) −1 with advection being the largest contribution at some distance from the coast. This value is similar to the overall potential temperature change between the stable phase and stratus phase (Fig.   7b). The absolute values of RAD derived with the radiative transfer model are the same order of magnitude as the modelled contribution by Schuster et al. (2013). We estimate TURB to cause about 22 % of TOT during the jet phase. The relevance of turbulent mixing below the LLJ axis is supported by profiles of radial velocity variance estimated from azimuthal Doppler lidar scans at 15 degree elevation angle, which show higher values in the shear layers above and below the LLJ axis . This confirms findings by Schrage and Fink (2012), Schuster et al. (2013) and Adler et al. (2017). Unfortunately, no general quantification of the turbulence is possible from the observations, although two Doppler lidars were deployed at From the radiosonde observations at Savè during IOP nights we find a strong relationship between a decrease in temperature and a very sudden increase in horizontal wind speed with an LLJ-shaped profile, which is why we attribute both these changes to the arrival of the maritime inflow. Note that this relationship is less clear when using continuous remote sensing instruments 25 and different objective criteria to detect the onset times . We assume that the relaxation of friction force leads to the formation of the LLJ-shaped profile within the maritime inflow. As the maritime inflow generally dominates the ABL conditions a few hours after sunset, there is not enough time for a pronounced LLJ to form locally without the influence of the maritime inflow. Based on the observational evidence, we propose therefore that the LLJ we observe at Savè is mainly linked to the maritime inflow and does not form locally. That means that the circumstances leading to LLJ formation differ 30 from those relevant for regions further in the north of southern West Africa, i.e. for the Sahel or Sahara (e.g. Lothon et al., 2008). These authors attribute the formation of the LLJ to the relaxation of the friction force after sunset and to the temperature and pressure gradients related to the Saharan heat low. As the regions further north are affected much later if at all by the maritime inflow, the LLJ forms locally and is not linked to the maritime inflow.

15
Eleven IOP nights from the DACCIWA ground-based field campaign conducted in southern West Africa during the summer Monsoon season in 2016 are analyzed in order to characterize the spatial distribution and temporal evolution of LLC, to investigate the intra-night variability of ABL conditions and to assess the relevance of processes related to LLC formation.
We used comprehensive observational data obtained at a supersite near Savè (Benin) from a ceilometer and cloud radar for  ii. The mean profiles for each of the nocturnal phases reveal some characteristic features: the surface inversion, which is present during the stable phase, gets eroded during the jet phase by differential cold air advection and shear-generated turbulence. During the stratus phase, static stability below and within the LLC layer decreases compared to the jet phase.
A distinct LLJ profile is visible during the jet and stratus phases, with the LLJ axis being near CBH during the jet phase and shifting upwards during the stratus phase. 25 iii. We calculate the contributions of temperature and specific humidity changes to the relative humidity changes between the late afternoon and the LLC onset. The relative humidity increases by about 25 % near the surface and decreases linearly with height up to around 750 m a.g.l. This increase is for the most part caused by cooling. Specific humidity contributes to the relative humidity increase before and during the arrival of the maritime inflow at the site, but specific humidity decreases once Savè is within the maritime inflow. In total, the contribution of specific humidity increase to 30 reach saturation in the ABL is very small.
iv. We estimate different terms of the heat budget for the time period before the LLC onset from the observations to assess which processes contribute to cooling: the temperature tendency at Savè is quantified from radiosonde measurements; the contribution by radiative flux divergence is derived from a radiative transfer model; the contribution by horizontal advection with the maritime inflow is estimated from radiosonde measurements at the coast and at Savè; and the bulk contribution by sensible heat flux divergence below the LLJ axis is assessed based on the surface sensible heat flux measurements. Strong cooling occurs at Savè up to around 750 m a.g.l. with a large part being caused by horizontal advection. We vertically average the different contributions and find that around 50 % of the cooling before LLC forma-5 tion is caused by horizontal cold air advection, while sensible heat flux divergence contributes about 22 % during the jet phase. Radiative flux divergence contributes roughly 20 % in the cloud-free nocturnal ABL -independent of the phase.
v. While horizontal advection of cool air with the maritime inflow is a necessary precondition for LLC formation, the LLC are not necessarily advected, but rather triggered by other mechanisms. Besides orographic lifting, we find evidence for two more possible mechanisms leading to LLC formation: during some nights, the sub-cloud layer is characterized by 10 small Bulk-Richardson numbers and low static stability which indicates that the LLC are triggered by shear-related ABL processes. On the other hand, we find that the LLC impact the wind profile by shifting the LLJ axis upwards towards the cloud top -compared to cloud-free conditions. This supports the hypothesis of the modeling study by Adler et al. (2017) that horizontal convergence upstream of existing LLC due to the upward shift of the LLJ axis triggers new LLC.
By using observational data of 11 nights we are thus able to identify relevant processes for LLC formation, to quantify different 15 terms of the heat budget, to identify typical dynamic and thermodynamic profiles for the different nocturnal phases and to confirm hypotheses based on numerical simulations. The results can on the one hand serve for the validation of high-resolution or large-eddy simulations and on the other hand be used to identify flaws in global and climate models, which might add to the problems of these models to correctly simulate the LLC and the West African Monsoon system. Furthermore, this study adds to the development of a conceptual model explaining the evolution, maintenance and dissolution of the LLC, which will 20 be based on the observational and numerical analysis within the DACCIWA project.
Data availability. After the DACCIWA embargo period, the data used in this study will be available on the SEDOO database (Derrien et al., 2016;Handwerker et al., 2016;Kohler et al., 2016;Wieser et al., 2016) Competing interests. The authors declare that they have no conflict of interest.      Figure 13. Cloud-base height (CBH) measured by ceilometer and lifting condensation level (LCL) calculated from surface measurements at 2 m a.g.l. for IOPs 6 (a) and 7 (b). Relationship between the Bulk-Richardson number calculated from radiosoundings for the sub-cloud layer (RiB,0−CBH) and the difference between CBH and LCL (c).