ERA5 composites at 05:00 UTC were computed by calculating the median pixel value of fog or low-cloud events, respectively, over southern Africa and adjacent parts of the Atlantic Ocean (median composites). Difference maps were created by subtracting the fog from the low-cloud composite. As fog and low clouds at CM feature a contrasting seasonality, the resulting meteorological composites can be expected to have a strong signature of the peak seasons. The composites can therefore not be directly interpreted as being mechanistically linked to fog vs. low clouds. Hence, regions are identified where the composite differences have the same sign in all 4 seasons DJF – MAM – JJA – SON to facilitate interpretation. The heatflux composites were computed additionally at 03:00 UTC, which is before sunrise in all seasons, to check for radiation influence, and showed no relevant differences to 05:00 UTC. To examine the specific interplay of the identified general drivers, the fog event of the night of 22/23 September 2017 was analysed in more detail and supplemented with the additional ceilometer and SODAR data from the NaFoLiCA campaign. Anomalies from typical September conditions were calculated in units of the standard deviation from hourly mean values of all September events (including clear days).

A simple logistic regression is used to classify FLC-covered days at CM into fog and low-cloud days using ERA5 data of large-scale features associated with fog vs. low-cloud appearance. The logistic regression resembles a linear regression but uses a sigmoid function that is fitted by maximum likelihood estimation to predict class probabilities . We use the default settings of the Python library scikit-learn but without L2-regularisation. The L2-regularisation reduces overfitting in the case of many input variables , which does not apply to our study since we only use three variables.

To estimate the model's skill and transferability, we use cross-validation. Because the data is a non-continuous time series (only FLC-covered days), a custom cross-validation algorithm is implemented to achieve (1) minimum temporal correlation between training and test set, and (2) representation of all seasons in the test set. For that purpose, all events within a window of 365 consecutive days are used as a test set, and the remaining FLC days are used for training, discarding the 7 d before and after the test window. During the iteration process, the window is moved over the time series in 3-month steps. This results in 23 cross-validation iterations (6 non-overlapping). Across iterations, this corresponds on average to a training-to-test ratio broadly equivalent to a standard 5-fold cross-validation split.

As suitable input variables for the logistic regression, the observed patterns from the composite analysis were captured in a range of single-value variables and refined to three meaningful features (Table ). The features were standardized separately in each iteration of the cross-validation using a pipeline to mitigate the effects of different value ranges.

The performance of the logistic regression is compared to a climatological baseline. Instead of using model predictions, the cross-validation procedure is performed using random draws from the monthly probabilities of fog/low-cloud occurrence. Full years serve as folds. To ensure reliable performance estimates, the random draw predictions for each excluded test year are repeated 1000 times and averaged. Model skill is evaluated using two complementary metrics: classification accuracy, defined as the fraction of correctly classified events, and the F1-score, defined as the harmonic mean of precision and recall for the fog class. The F1-score is reported in addition to accuracy because it is less sensitive to changes in class balance between cross-validation folds and rewards a model only when both classes are correctly identified.

Fog and low clouds at CM show a pronounced alternating seasonality (Fig. ). The months April-May-June (AMJ, indicated by the red bars at the top of the figure) capture the peak fog season, while the months September-October-November (SON, blue bars) are dominated by low clouds. This seasonality agrees roughly with , who find a deeper MBL during austral summer and a thinner MBL during the fog peak season in early winter. The peak occurrence time can vary considerably between years, as 2017 shows a late fog peak in August and September compared to the expected peak time in May and June (e.g., in 2018). The usual low-cloud peak is missing in November/December 2019, followed by irregular fog peaks during 2020, which makes these two years show unusual FLC occurrence.

To examine the key drivers beyond peak season characteristics, we compare fog events during the whole year to low-cloud events during the whole year using median composites of meteorological variables (Fig. ). A pattern of dipolar pressure differences appears in the composites of the fog vs. low-cloud events (Fig. a–c), resembling the peak season conditions (Fig. a, b). The outlined differences are consistent throughout the year (all four mean differences during DJF, MAM, JJA, SON have the same sign, e.g., MSLP is greater in all four seasons) and may thus point to a physical link to stratus altitudes and, thereby, to spatial fog occurrence patterns. In the following, the consistent differences over the ocean are described and discussed first, followed by the consistent differences over the southern African continent.

Low-cloud situations show higher pressure (red) and about 27 % stronger anticyclonic winds on the northeastern flank of the AH, at ≈30°S (Fig. c). In the region of increased surface winds, increased upward latent and sensible heat fluxes are apparent during low-cloud situations (blue in Fig. f, i, as fluxes are negative upward and generally upward over the ocean during both fog and low clouds in Fig. d, e, and g, h).

The differences in geopotential at 850 hPa (Fig. l) appear similar to the MSLP differences (Fig. c) with lower values over the continent (blue) and higher values (red) in the AH region in low-cloud situations. However, the region of seasonally consistent lower geopotential extends from the continent over the ocean (Fig. l). This difference persists up to 500 hPa (not shown) and is tilted towards the coast, suggesting a slightly stronger subsidence regime in fog cases. In this marine region and extending offshore, the MBL is consistently shallower in fog situations (Fig. o), particularly near the Namibian coastline northwest of CM, where the median difference in MBL depth is up to 130m. While it is likely that the BLH is generally biased low (see Sect. 2), it is not expected that this bias would strongly influence this composite difference.

We interpret the composites and their differences in terms of AH-related processes as follows. The more pronounced AH in low-cloud situations produces the observed stronger coast-parallel southeasterly winds at ≈30° S. This causes increased cold-air advection from the south, which likely drives the intensified sensible and latent heat fluxes into the MBL in low-cloud situations. Increased upward heat fluxes indicate a stronger contrast between cool air over warm ocean water, leading to unstable conditions, stronger evaporation and heat uptake from the ocean, and have been found to result in a deeper MBL by increasing the buoyancy and the vertical mixing in the MBL . Indeed, found that specific humidity in MBL air masses is higher on FLC days than on clear days already at least 24 h before the FLC event occurs in the Namib. This suggests that the development of the MBL over multiple days critically determines its moisture content, depth, and the resulting stratus altitude.

The AH also drives a coastal jet centered at around ≈20° S , which is weakest in May and June during the fog peak season and most pronounced during the low-cloud peak season. The jet evolves when winds from the AH are deflected into a coast-parallel direction by the elevated terrain, and it advects cold air into the marine stratocumulus region, thereby enhancing stability and cloud formation . While our results do not show local differences in wind speeds at this latitude between fog and low-cloud days, a link to stratus altitude is likely and warrants further attention in the future.

Over the southern African continent, low-cloud situations feature lower MSLP (blue), and a less pronounced influence of continental airmasses at the coast. One should note that MSLP is likely affected by the elevation of the topography in southern Africa. The difference in the wind patterns is even more pronounced at 850hPa than at the surface, featuring less continental easterly winds arriving at the coast in low-cloud situations (Fig. l).

In both fog and low-cloud situations, the cool and moist coastal MBL is capped by warm and dry air (Fig. ), forming an inversion that lowers and strengthens towards the continent. The vertical position of this inversion cap is lower during fog situations throughout the seasons and along the entire Namibian coast, confining the shallow BL. The inversion cap in fog situations is warmer than during low-cloud situations. It should be noted that the differences in temperature and moisture profiles is strongest below the inversion base in low cloud conditions, i.e. in layers that already feature an inversion in fog cases but that are within the boundary layer in low cloud cases. Local temperature and moisture profiles at CM are shown in supplementary Fig. , and highlight that differences between fog and low cloud cases peak at 950 hPa. As the temperature cross sections show limited temperature differences at 1000 hPa over the ocean and above 850 hPa in the free troposphere, the overall stability of the lower troposphere, which is an important fog and low cloud control , is not found to differ substantially.

There are also differences in the regional circulation. The zonal circulation pattern over the ocean is shifted vertically by one reanalysis pressure level (25hPa), with a turnaround from easterly winds near the surface to westerly winds at 850hPa (fog) vs. 825hPa (low cloud) where temperatures decrease again (Fig. a, b). This vertical shift appears consistently throughout the seasons with free troposphere (FT) winds between 750 and 900hPa having a consistently weaker eastward component in low-cloud situations, visible by the easterly direction of the residual wind vector in Fig. c. The subsidence over the Great Escarpment is consistently less pronounced in low-cloud situations and vice versa, more pronounced in fog situations (Fig. a–f). One should note the lack of a pronounced easterly wind at the Escarpment which would, if present, indicate a systematic synoptic forcing in the composite. In both fog and low-cloud situations, the subsiding air masses follow the topography towards the coast. In fog situations, this easterly flow is slightly stronger and persists to lower elevations, thus reaching closer to the coast. The depth of these easterly winds is comparable to the typical depth of the katabatic mountain-plain wind with ≈ 1000 m a.g.l., peaking at 250–500 m a.g.l. described by .

The described difference patterns likely work together to modulate the altitude of the advected stratus through the following sequence of processes. A stronger AH in low-cloud cases drives enhanced coast-parallel southeasterly winds and increased surface heat fluxes upstream (Fig. 3f, i), supporting a deeper, more vigorously developing MBL that arrives at the Namib coast deeper than in fog cases. A weaker AH in fog cases produces weaker upstream heat fluxes and an MBL that arrives shallower. The temperature signature of this depth difference at fixed pressure levels in the coastal MBL (cooler in low-cloud cases at 975–900 hPa, with near-surface temperatures and T at 850 hPa essentially unchanged) is likely dominated by the inversion sitting at higher altitudes in low-cloud cases, rather than by a uniform thermodynamic anomaly of the MBL air mass. Locally over the elevated interior, the higher continental pressure and stronger subsidence in fog cases is likely to suppress daytime convective cloud cover, which enhances overnight longwave radiative cooling of the elevated terrain, and thereby strengthens the shallow surface katabatic drainage component that can mix with, undercut, or block the marine air mass near the coast . Because this drainage flow is cool, it cannot cap the MBL from above and is therefore not expected to contribute meaningfully to the inversion-lowering signal of the fog days in the composites.

Lower continental pressures have already been related to inland-reaching FLC occurrence and longer persistence of FLC events . Warm easterly continental winds have been associated with lower cloud base height of Namibian FLCs during events with limited inland extent in simulations . These simulated low-lying fog events featured warmer capping air and a stronger inversion, in line with the satellite and ERA5 data. suggests that a merging of the continental NE wind with SE trade winds may result in a strengthening of the capping inversion. Easterly continental winds have also been found to cause a shallower MBL offshore the central Namib during the fog season March–August, with decreased FLC cover, especially in the region north of CM , where we observe the greatest BLH differences between fog and low-cloud situations. One should note that both the simulated low-cloud-base events of and the monthly-mean easterly-winds presented in are at least partly capturing bergwind-type conditions. In most such situations the warm, dry descending continental air entirely dissipates the marine stratus, consistent with the decreased FLC cover reported by and with the general expectation that strong descending continental winds and the associated adiabatic heating overwhelm the coastal stratus regime. When the dry descending continental air does not reach the surface at the coast, however, it can instead lower and confine the stratus into a narrow band of low-lying fog at the coast . It is possible that the warmer capping air and slightly stronger easterly advection in fog cases could be partially due to such bergwind-type conditions contained in the fog composite. We therefore present one such event as a case study in Sect. 4.2.

A note on scale separation. The discussion above implies that the regional easterly circulation in the central Namib is the product of two physically coupled but qualitatively distinct contributions: a synoptically forced subsiding easterly (bergwind), driven by the horizontal pressure gradient between the elevated continental interior and the coast, and a cold surface-based katabatic drainage component driven by local nocturnal radiative cooling. The two are not independent as a similar high-pressure regime that establishes the synoptic pressure gradient also facilitates the formation of the local katabatic flow, but they cannot occur simultaneously and interact with the marine boundary layer in qualitatively different ways. The synoptically forced flow arrives at the coast warmed by adiabatic descent and would act on the MBL from above, while the shallow cool surface katabatic drainage is more likely to mix with or undercut/block it near the surface than to cap it . An inversion-strengthening or lowering signal in our composites can though only be caused by the synoptically forced component, since only the warmed-by-descent synoptic flow can cap the marine boundary layer from above. However, this is not thought to be a dominant process as no clear evidence of a systematic presence of synoptically forced easterlies is present in the fog composite (e.g. no easterly component of winds above the Escarpment). The same coupling affects the interpretation of regional pressure-gradient diagnostics such as Pdiff (Table 1), which captures the local east–west pressure gradient driving the near-coastal flow but cannot be uniquely attributed to a single scale of forcing.

The relationships between the continental synoptic pressure (CH), the regional pressure gradient Pdiff, the near-coastal flow and the offshore marine BLH are shown in Fig. 5. The strength of the CH is moderately and negatively correlated with Pdiff (Pearson r=−0.43, Fig. 5a), confirming that a stronger CH is reflected in a more pronounced east-west pressure gradient toward the coast, and that the synoptic and regional pressure signals are physically coupled. Pdiff in turn is moderately correlated with the 900 hPa zonal wind close to CM (r=0.51, Fig. 5b), showing that the regional pressure gradient is associated with a stronger easterly component of the near-coastal flow. The strongest relationship is between Pdiff and the offshore BLH at the same latitude (r=0.62, Fig. 5c), with a more negative Pdiff associated with a shallower offshore MBL. This last relationship connects the CH to the depth of the marine boundary layer that ultimately controls the stratus altitude. One should note that the Pdiff–BLH scatter (Fig. 5c) is triangular rather than elliptical: a deep MBL is observed only when Pdiff is weakly negative or positive, while strongly negative Pdiff values are systematically associated with a shallow MBL. This asymmetric distribution suggests that the bergwind regime contained in the lower-tail of the Pdiff distribution imposes a direct local upper bound on the offshore MBL depth by mechanically suppressing it, while in the bulk of the distribution the Pdiff–BLH relationship does not seem to be as tightly coupled and is more representative for the synoptically unforced conditions. The fog (red) and low-cloud (blue) events are visibly displaced from each other in all three panels, illustrating that these synoptic and regional variables jointly carry information that discriminates between fog and low-cloud events. This motivates their use as features in the classification of Sect. 4.3.

From the composite analysis, we summarize the following seasonally consistent large-scale characteristics of low clouds compared to fog (and vice versa): (1) A stronger AH produces stronger MBL winds, cold air advection, and enhanced sensible and latent heat fluxes upstream of the study region, in combination leading to a deeper marine BLH during low cloud cases than in fog cases. (2) Inland over the elevated terrain, a weaker continental high reduces the production of near-surface katabatic drainage . The composites do not show marked differences in local subsidence or synoptically forced warm and dry easterlies between fog and low-cloud cases, so the majority of the inversion-lowering signal in the composite is most plausibly attributed to the upstream processes rather than to anomalously strong local downward forcing.

In the following, an anomalous fog event that occurred during the low-cloud season is used to assess how the identified meteorological differences may contribute to day-to-day variability in stratus altitude and out-of-season occurrence of fog at CM. The event occurred during the NaFoLiCA campaign and is partially documented in .

During the night of 22/23 September 2017, a stratus is advected over CM with a cloudbase height consistently below 100m a.g.l., appearing as fog from about 23:00 to 06:30 UTC at CM with almost no lowering and lifting (Fig. a). During the nights before and after, a low stratus is advected at 300–500 m a.g.l. that either stays lifted as a low cloud (leading to fog further inland, 20/21 September) or lowers towards the surface, leading to fog occurrence at CM, and then strongly lifts and dissipates after sunrise (19/20, 21/22, 23/24 September). The sudden appearance of a very low cloud base height during the case study possibly points to a generally low stratus layer and thus to a shallow MBL. A limited stratus extent is apparent on the available MODIS Terra image at 09:00 UTC (Fig. b), showing a narrow 20–30 km wide coast-parallel cloud strip that has, at that time, already been shifted back towards the ocean.

The synoptic situation during this specific fog event (Fig. b) shows characteristics that were identified as typical for fog events and that differ from the climatological September conditions that typically produce the higher-altitude low clouds at CM during this time of year (and the days before and after the event, Fig. a, c). The first characteristic is the higher pressure over the southern African continent (Great Escarpment and Kalahari) compared to the nights before and after. It is here caused by the formation of a ridging high-pressure circulation pattern over the subcontinent following the passage of a mid-latitude cyclone which impacted the continent at about 30° S. The second characteristic is a southeasterly 850 hPa wind that originates from the AH and passes over the continent before reaching CM, in contrast to the calm local conditions observed on the nights before and after. As a third characteristic, the coast-parallel southeasterly winds are weakly developed because of the position of the AH, which is also the case one day before the event.

The vertical structure of the regional atmosphere is shown in a zonal cross-section along the station latitude 23°S (Fig. b, e, h). The air masses coming from the continental interior turn into a strong downward wind of up to about 2 Pa s⁻¹ vertical and 20 m s⁻¹ easterly wind speed, blowing over the Great Escarpment and down to the top of the coastal boundary layer. As outlined before, such strong winds co-occurring with ridging highs have been described as bergwinds and are often strong enough to disrupt the stable surface layer, typically leading to the dissipation of low stratus clouds. The wind may have caused the clear sky over most of the Namib and the ocean without dissipating the small coastal FLC strip. It may have also been deflected by the wind shear with the northerly coast-parallel surface wind that was measured at CM and that likely advected the stratus from the low-cloud deck over the ocean about 350 km north of CM (Fig. b).

The comparably shallow coastal boundary layer is capped by a strengthened inversion with temperatures up to 301.8K (28.7°C) (Fig. a–c), likely caused by adiabatic heating of the subsiding air masses. The wind seems to not penetrate the cool and moist boundary layer containing the stratus. Rather, the wind is deflected offshore at the top of the BL, at the same time likely pushing it down and towards the coast (Fig. b, e, h). The continental easterly wind thereby seems to limit the inland extension of the MBL and a further inland advection of the stratus. The actual advection of the stratus is not visible in these curtain plots because it happens with northerly winds (FLC extension from northerly FLC deck in Fig. b, nocturnal northerly wind (blue) during nights in CM in Fig. b). The inversion strengthening and subsidence are more extreme in this case than in the median composites of fog, underscoring the strong synoptic forcing.

The meteorological characteristics related to fog occurrence are extraordinarily pronounced during this specific event compared to mean September conditions (Fig. a), pointing to its exceptional nature. The anomalies are even more pronounced inland (GB) than at the coast (CM), highlighting that the mechanisms leading to fog near the coast are of regional to synoptic scale. Anomalously increased temperatures (1–2 σ) at the altitude of the inversion (T950) together with extraordinary inland easterly-downward wind (-4 σ) around 900hPa (U900, W875) indicate the warm-air advection from the continent and adiabatic heating. Ground-based SODAR measurements confirm the easterly wind component. At GB (station elevation 405m a.s.l.), the easterly extends down into the lowest SODAR level at ∼450 m a.s.l. and thus effectively reaches the surface. At CM (94m a.s.l.), by contrast, the easterly is observed only at ∼500 m a.s.l. – several hundred metres above the surface – and only after 10:00 UTC due to the limited SODAR range (Fig. b). The anomaly in the regional pressure difference (-1 σ) developing over the night (Fig. ) that approximates the thermotopographic winds (Pdiff) is less pronounced than in the occurring winds, indicating a strong synoptic forcing on the local winds in the free troposphere. This is further illustrated by the position of this case in the climatological Pdiff–U950 scatter (Fig. 5b) where the particularly pronounced easterly advection of this event departs clearly from the linear relationship of the two variables, as the deep synoptically forced subsiding easterly contributes momentum at the coast beyond what the local pressure gradient encodes.

The case study highlights that mechanisms at the regional and particularly at the synoptic scale control the local boundary layer and the characteristics of this fog situation. This exceptional fog event presents an alternative mechanism by which synoptics can modify MBL and stratus height: the warm, dry, adiabatically heated bergwind character of the descending air mass demonstrates an inversion-lowering-from-above mechanism, separate from the upstream effects that seem to dominate in the composite. We also conclude that fog at CM can be connected to two different high fog situations: (1) a very low cloud base touching the ground when it is advected to CM (typical for AMJ, but can also be synoptically forced out of season), or (2) stratus base lowering after the advection of the stratus has already taken place (as shown by the ceilometer in the fog nights before and after the case night).

Taken together, the composite analysis (Sect. 4.1) and the case study presented above provide a consistent picture of the synoptic and regional-scale processes that modulate the stratus altitude in the central Namib which are schematically summarized in Fig. . Under typical low-cloud conditions, a strengthened Atlantic High drives enhanced coast-parallel southeasterly winds, increased latent and sensible heat fluxes upstream of the study region, and a correspondingly deeper marine boundary layer that supports an elevated stratus base. Under typical fog conditions these processes are less developed and the MBL is shallower, and elevated continental pressure favors the development of a nocturnal mountain-plain wind that descends the Great Escarpment, and can mix with, undercut or block the advection of marine air masses. The case study additionally shows that the regional katabatic flow can be completely overridden by a synoptically forced (here: ridging high-pressure system) easterly flow that is hot and dry and intensifies the inversion from above well beyond the composite mean and confines the stratus to a narrow coastal strip.

A logistic regression model was trained to classify FLC days into fog or low-cloud days with three features capturing the CH, AH, and the regional pressure gradient (described in Table ). With only these three meteorological features, the model achieves a cross-validated 69.1% accuracy and thus excels the mean baseline by 10.1% (F1-score of 0.622 vs. 0.545 for the baseline, an increase of 0.077; Fig.  bottom panel).

The mean baseline accuracies for the test years 2015–2020 are 0.594, 0.594, 0.613, 0.594, 0.578, and 0.567, respectively, illustrating that guessing fog/low cloud from seasonality is better than a random pick (accuracy = 0.5 to distinguish the event types for a given FLC event). The applied split method in the cross-validation allows for comparing the performance on different test sets of 365 d (Fig. ). The reduced classification skill during the 2019 and 2020 test windows (Fig. ) coincides with the Benguela Niño of October 2019–January 2020, which produced coastal SST anomalies exceeding 2°C and was accompanied by anomalous near-surface winds and surface heat fluxes . Under these anomalous oceanic conditions, the climatological coupling between the synoptic pressure pattern and the marine boundary layer state is weakened. Warmer SSTs enhance latent and sensible heat fluxes into the MBL and deepen it, so that synoptic conditions that would normally produce fog at CM may instead produce elevated low clouds. Because the model uses only pressure-based features (CH, AH, Pdiff) and does not carry direct information on SST or surface heat fluxes, this oceanically driven shift in the MBL response is not represented in the feature space. The simultaneous drop in baseline skill during the same period confirms that even the climatological seasonality is shifted, so that the lower model skill in 2019/2020 reflects a genuine departure from the climatological expectation rather than a methodological shortcoming. This is consistent with the findings of that Benguela Niños are a relevant driver of fog and low-cloud variability in the central Namib, and suggests that including SST patterns or MBL information could improve model robustness to anomalous oceanic states in future work.

As the model based on meteorological features outperforms the baseline, the features carry information related to the target variable fog or low cloud, consistent with their visible separation between fog and low-cloud events shown in Fig. 5. The results also highlight the potential to use reanalysis information to separate fog from low clouds in satellite-based FLC products, in particular, when local temperature and humidity profiles are additionally used.