The GAW Global station Mt. Kenya (MKN), located at 0.062° S, 37.297° E (WMO Integrated Global Observing System (WIGOS) Station Identifier (WSI) 0-20008-0-MKN), is situated on the north-western slope of Mt. Kenya at an altitude of 3678 m a.s.l. and is surrounded by alpine grassland and shrubs. As part of the Mt. Kenya National Park, the whole mountain area is protected and there are minimal local anthropogenic emissions, making the site suitable for continuous observations of the tropospheric background composition. Access by car is possible to the Old Moses Camp, located 300 m below and approximately 1.9 km northwest of the station. The closest settlement is situated 17 km in north-westerly direction, the closest city (Nanyuki) lies 27 km westwards at 1900 m a.s.l .

MKN was established under the World Meteorological Organization (WMO)/United Nations Development Programme (UNDP)/Global Environment Facility (GEF) programme GLO/91/G32 in the mid-1990s and was officially inaugurated in October 1999. Due to limited power supply, it has operated on an irregular basis until 2014, when the power line was completely renewed. Nowadays, Mount Kenya is one of the very few high-altitude observatories in East-Africa and is among the best-equipped mountain stations in terms of atmospheric composition in the region.

The climate in Kenya is governed by the seasonally varying position of the ITCZ. The ITCZ lies North of MKN in boreal summer and south of MKN in boreal winter. The recurrent ITCZ displacement leads to different advection regimes in Equatorial Africa (see Fig. ) with two rainy seasons from March to May (MAM) and October to December (OND), when the ITCZ crosses the equator . The months in-between are characterized by a short dry season from January to February (JF) and a long dry season from June to August. September is a transitional month, with remaining dry conditions but higher temperature than in previous months. In this study, we count September as part of the long dry season (JJAS). Rainfall at MKN is governed by the mentioned rainy seasons, with maximum precipitation in April (see Fig. ). However, precipitation measurements at MKN reveal larger values in August and September than expected from the regional seasonal classification (Fig. ). This local precipitation is mainly governed by advection of Congo air masses (e.g. ), which converge around northern and western parts of MKN, leading to lifting air masses with showery precipitation. Nevertheless, we use the more regional classification of JJAS as “dry season” in our study. Daily minimum and maximum temperature at MKN vary on average between 4 and 11 °C. For more insights on the meteorological and climatological conditions and a comprehensive overview of the station characterization, refer to . Statistical data analysis of early carbon monoxide and ozone measurements (2002–2006) and air mass trajectory clustering confirmed the large-scale representativeness of the station and suitability of the location for background observations, specifically during the night .

In-situ observations of greenhouse gases and air pollutants have been performed at the MKN station since the early 2000s, providing a unique long-term dataset for equatorial Africa. In 2006, issues with the station's power supply resulted in an interruption to the early CO and O₃ measurements. In 2008, there was an attempt to implement continuous GHG measurements, but it was unsuccessful due to severe damage to the power line caused by a wildfire in 2009. In 2010, the power supply was completely disconnected, but a new power line was installed in 2014, and continuous gas measurements were resumed in December 2019. In the following analysis, we therefore focus mainly on the continuous data from 2020 to 2024. The measurements are usually level 0 data stored as 1 min averages. Quality-controlled data are provided with 1-hourly temporal resolution.

For a better visualization and to obtain smooth seasonal cycles, we applied a curve fitting method provided by NOAA (https://gml.noaa.gov/ccgg/mbl/crvfit/crvfit.html, last access: 9 September 2025). The corresponding python code is publicly available (https://gml.noaa.gov/aftp/user/thoning/ccgcrv/, last access: 9 September 2025). The curve fitting method, also called CCGCRV (Carbon Cycle Group, Earth System Research Laboratory (CCG/ESRL), NOAA, USA), is a slight modification of the filtering approach presented by and has been described and used in multiple studies (e.g. ). The method first fits a 3rd order polynomial and 4th order harmonic function. Afterwards, the data is smoothed with a fast Fourier transform (FFT), and filtered with a low-pass filter, for which we use a short-term cut-off frequency of 80 d and a long-term cut-off of 667 d to remove remaining oscillations. To obtain the unbiased seasonal variability, any linear trend was removed.

Due to the sparse availability of continuous measurements in the tropical region, the MKN measurements are highly valuable for validation of atmospheric models. We compared the MKN measurements with model data from CAMS . For comparison of the model output with the MKN measurements, we selected for each species the CAMS grid for which CAMS data correlates best with the corresponding measurements over the entire time period. In the following, we describe the CAMS products used in our analyses.

To understand the origin and variability of trace gases and aerosols observed at MKN, it is essential to consider the influence of large-scale atmospheric transport and regional emission sources. The station's central location in equatorial Africa exposes it to contrasting air mass regimes from both hemispheres, resulting in a complex mixture of natural and anthropogenic signals. While in-situ measurements offer high-resolution temporal data, they do not provide direct information about the geographic or sectoral origin of the observed compounds. To address this, we use the Lagrangian particle dispersion model FLEXPART to simulate air mass transport to MKN. These simulations yield residence time distributions that describe the sensitivity of the station to air coming from different regions. By folding these residence times with spatially resolved emission inventories, we derive quantitative source contribution estimates to the measured concentrations. This approach enables us to disentangle the influence of biomass burning, wetland activity, and anthropogenic processes. Although the results depend on the transport simulations and cannot be used for quantitative verification of the emission inventories, they offer valuable insights into the processes driving variability at MKN and allow for a broad evaluation of the quality of current emission inventories for the African continent.

As emission inventories, we used satellite-based fire emissions from the CAMS Global Fire Assimilation System (GFAS), wetland methane emissions from WetCHARTs , and anthropogenic emissions from the Emissions Database for Global Atmospheric Research (EDGAR) . We concentrate on the regional emissions where sensitivities and emissions are largest and include emissions from most of Sub-Saharan Africa (0° W to 60° E and 20° N to 35° S, see Fig. a). The global WetCHARTs and EDGAR emission data were handled and prepared to fit our study area using the Python package emiproc . Averaged maps for CO and CH₄ emissions in 2020 from those datasets are shown in Fig. .

Using the FLEXPART model, we simulated 3-hourly footprints for artificial particle releases at MKN for the whole study period (2020–2024). The residence time τ averaged for each month is shown in Fig. . The simulations confirm the dominant role of the ITCZ, which acts as a dynamic boundary that creates seasonally distinct hemispheric atmospheric regimes at MKN . MKN is predominantly exposed to north-easterly advection from December to March, and to south-easterly advection from July to October, with the remaining months indicating the transition periods (Fig. ). The atmospheric transport simulations indicate that processes above Western Africa only have an indirect impact on the observations at MKN.

Continuous GHG measurements at MKN have become available since 2020, as detailed in Sect. , while earlier in-situ and flask measurements had been measured intermittently since 2002 (Fig. ). Trends for different periods as distinguished by different colours in Fig.  are listed in Table . Monthly aggregates of both the continuous measurements and the flask analyses were used for the determination of trends by linear regression. The CO₂ trend intensified over the 23-year record. While flask measurements for 2002–2011 indicated an increase of less than 2 ppm yr⁻¹, recent observations show a rise of nearly 2.4 ppm yr⁻¹. This rate is consistent with global increases reported by the NOAA network but is slightly below the most recent decadal global growth rate documented by WMO . CH₄ trends during 2002–2011 were characterized by the temporary global near-steady state in atmospheric CH₄, before growth resumed in 2007 . As a result, the early-period trend is comparatively weak, while recent observations show a significant increase of more than 12 ppb yr⁻¹, even exceeding the global mean growth rate of 10.6 ppb yr⁻¹ over the past decade . This acceleration aligns with recent assessments that the tropics dominate the latitudinal contribution of global CH₄ emissions and highlight Central Africa and tropical South America as the dominant contributors to the current net CH₄ flux to the atmosphere .

The combined O₃ time series for 2002–2024 is the most internally consistent among the four trace gases discussed here, as it is based on continuous in-situ measurements using the same technique throughout. A statistically significant positive trend is found for the full period (2002–2024) of 0.20 ± 0.07 ppb yr⁻¹. Comparison of sub-periods indicates a slowdown in the ozone growth rate, with a stronger increase during the early years (2002–2006: 0.42 ± 0.88 ppb yr⁻¹) and a smaller trend in the more recent period (2015–2024: 0.14 ± 0.21 ppb yr⁻¹) though both not statistically significant. This behaviour is consistent with SHADOZ observations showing that free-tropospheric and total tropospheric ozone above Nairobi exhibit weak, statistically insignificant long-term trends over 2000–2022 of 0.05 to 0.08 ppb yr⁻¹ depending on the trend calculation approach , with earlier positive tendencies levelling off in recent years . Our results also complement the conclusions of , who analysed long-term ozone trends from 27 remote ground-based sites worldwide but could not include East African stations owing to the lack of sustained observations at the time. While highlighting the African region as a major observational gap, they concluded that, globally, free-tropospheric ozone has increased since the mid-1990s, in contrast to Europe and North America, where surface ozone trends have largely flattened or declined since around 2000.

The long-term CO trend is the most difficult to interpret because it combines several measurement techniques, including in-situ NDIR and CRDS observations, different calibration strategies, and a mixture of in-situ, flask, and later again in-situ data. In addition, CO observations in the tropics are strongly influenced by biomass-burning emissions and by the varying strength of El Niño and La Niña events. These factors produce temporary CO enhancements, resulting in large interannual variability that often dominates any long-term trend signal. For the period 2002–2012, the CO trend derived from flask measurements is negative (−0.48 ± 1.28 ppb yr⁻¹), consistent with overall patterns observed at remote background sites . However, also highlight that trends in the tropics are less consistent than at higher latitudes because of the strong influence of biomass burning and climate variability. In contrast, the most recent period (2020–2024) shows a slightly positive trend, similar to that obtained for the full dataset. Depending on whether the early continuous measurements or the flask data are used, the overall CO trend ranges between 0.11 ± 0.33 and 0.63 ± 0.29 ppb yr⁻¹. These differences are likely related to the varying sensitivity of continuous versus intermittent flask sampling to episodic fire emissions. The long data gap from 2012 to 2019 also compromises robust trend determination.

Earlier conclusions from Mount Kenya CO and O₃ data determined a slightly positive trend for MKN CO data from 2002 to 2022. Furthermore, we observe that the seasonality does not differ substantially between previous and recent measurement periods (Fig. e–h). However, the February-maxima in CO₂ (Fig. e) and CO measurements (Fig. h) were larger in the 2002–2006 period (red and yellow lines) than in the recent period (blue line). Indeed, while , based on MKN measurements from June 2002 to June 2006, and , using data from Mt. Mugogo (Rwanda) from 2015 to 2017, reported peak CO values in February, our more recent observations (2020 to 2024) show a shift, with maximum CO levels in July (Fig. h).

Larger and slightly earlier CO peaks in recent dry seasons (June/July 2020 to 2022) were also observed by , who investigated MKN CO measurements from 2002 to 2006 and 2020 to 2022. This shift in maximum CO values from February to July might be linked to seasonal changes in wildfire intensity, but it may also result from dynamical atmospheric changes. Further investigations would be needed to confirm those possibilities.

The temporal evolution from 2020 to 2024 and the seasonal cycles for the five analysed species at Mt. Kenya are shown in Fig. , both for in-situ measurements and CAMS model data. Generally, we observe a good agreement between measurements and CAMS data, suggesting that the measurements are highly valuable for model validation in this region. The general larger variability in the measurements is partly due to their hourly resolution, whereas CAMS data shows 3-hourly values (or 6-hourly for CH₄), but mainly because the model misses some local variability.

The CO₂ CAMS data (Fig. a) have an even smaller variability than the other species. We observed that CAMS invGG generally only shows clear diurnal variations (> 1 ppm) up to approximately 850 hPa. At higher altitudes (lower pressure), the diurnal variability in the invGG product decreases and becomes negligible. At the selected CAMS level, the diurnal amplitude is on average only 0.04 ppm, which is a factor of 100 smaller than expected from measurements (see Sect. ). These results suggest that despite the overall good agreement of the CAMS invGG CO₂ product with MKN measurements, the CAMS model is unable to resolve the diurnal cycle at the given altitude. We suspect the reason for this discrepancy is due to poor representation of the boundary layer in the CAMS model. Also for CH₄, the CAMS model at the selected level underestimates the diurnal cycle by a factor of 4 to 10 (depending on the season), but the underestimation is less pronounced than for CO₂.

CAMS O₃ data agree well with measurements (Fig. c). In addition to in-situ and CAMS O₃ data, we show ozonesonde measurements from the nearby launches in Nairobi. For each sounding profile, we chose measurements at a fixed pressure level (700 hPa) to compare with the in-situ measurements at MKN. Generally, the ozonesonde measurements agree well with the in-situ measurements. However, ozonesondes seem to report slightly higher values in 2024, which requires further investigation.

In 2023, the ozone in-situ measurements show smaller values than in other years, with a decrease starting approximately in September 2023 (Fig. d). This ozone decrease, which is also observed in ozonesonde measurements and in CAMS data, might be related to the El Niño event in 2023/2024 . Earlier El Niño events occurred in 2002/2003, 2009/2010, 2014/2015, 2018/2019 (https://ds.data.jma.go.jp/tcc/tcc/products/elnino/ensoevents.html, last access: 7 July 2025). In line with this hypothesis, previous ozone measurements (Fig. d) show decreased values towards the end of the years 2003, 2015 and 2019 (though less pronounced in 2019), which corresponds to the autumn/early winter following those El Niño events. Unfortunately, no measurements were available during the other El Niño events and further analyses would be required to better understand the underlying dynamical processes.

In the CO and BC time series, we observe that a nearby wildfire event in March 2022 led to a strong peak of up to 1205 ppb and 5 µg m⁻³, respectively. To avoid distortion of the figure, we cut the scale of the axis in Fig. d, e. Figure e shows BC measurements from two different instruments (AE31 and AE33). CAMS CO data agree well with CO measurements, whereas the agreement is slightly worse for BC, especially in 2024. In June/July 2024, CO and BC show increased values that are not reproduced by CAMS. We could not identify a single dominant local wildfire event in that period, but wildfire activity from July to September that was not detected by CAMS could possibly explain the differences.

To assess the influence of natural and anthropogenic African emissions on concentrations measured at Mt. Kenya, we combine FLEXPART footprints with emission inventories as described in Sect. . We then add the resulting time series to the estimated background (see Sect. ) to obtain the total simulated concentrations or amount fractions at MKN. Figure  shows the total simulated time series, including various emission contributions for CO, BC, and CH₄ alongside the measured time series and the simulated background. The simulated time series are shown for both domains, the African domain (in red) and the full domain, including the Indian Ocean region (black line).

We observe good overall agreement between measured and simulated data for all three species, when transport and emissions are only accounted for the African domain (red line). However, when considering the full domain (black line), the simulated time series are mostly larger than observations. We saw that the largest contribution to this overestimation comes from anthropogenic EDGAR emissions, suggesting that the inventory overestimates emissions, not only in Africa but also from the larger domain. Figure  illustrates this overestimation of the simulated data, especially for CO and BC. Nevertheless, the correlation between observed and simulated BC and CO is weak in all months and the ODR model performs poorly for CO and BC, probably related to the large number of outliers (Fig. a, b). Furthermore, BC is treated in FLEXPART as an inert species neglecting scavenging, which likely explains why the model-observation agreement is worse for BC compared to CO, especially during the rainy seasons. Standard deviations of the measurements were considered for CO and CH₄ in the ODR fit, but were not available for BC, which might further explain the larger spread of the ODR fits for BC (Fig. b). Simulated CH₄ correlates better with the measurements (Fig. c) than CO and BC, but still slightly overestimates amount fractions in most months (Fig. c), for the full domain especially from December to March. This bias most likely results from overestimated emissions in the inventories or from an overestimated background. Indeed, the background (grey line in Fig. c) sometimes exceeds the measurements, indicating that simulated background values are too high.

As expected, we observe larger simulated contributions during the day (not shown), when the influence of PBL is largest, which is consistent with the results at Maïdo by . However, comparing only day- or nighttime measurements with simulations hardly affects the overall correlation between modelled and simulated time series, suggesting that there is no major improvement when looking only at free-tropospheric or PBL contributions respectively.

The sectoral contributions from natural and anthropogenic emissions to the simulated time series are shown in Fig. , and their relative contributions for different seasons are shown in Fig. . Both figures concentrate on the African domain only, to investigate the effect of African emissions on MKN measurements. We observe that for CO and BC, the energy sector is the dominant contributor (around 46 %, see Fig. a, b), with contributions from the “energy for buildings”-group only. This represents solid fuel combustion for domestic and commercial energy production. The second largest contributing sector is industry (38 %), followed by uncontrolled fire (13 % for CO, 10 % for BC), with largest fire contributions around the dry season from June to October, and transportation. The fire sector here stands for uncontrolled burning detected by GFAS, in contrast to solid fuel burning for domestic energy production, which is part of the energy sector. As expected, CO and BC show very similar patterns, but the contribution from the transportation sector is slightly larger for BC (6 % for BC compared to 3 % for CO), due to larger shipping BC emissions.

As mentioned in Sect. , we use annual EDGAR inventories for anthropogenic emissions without seasonal variation. Our observed seasonal variability in anthropogenic emissions is thus only due to variations in atmospheric transport. For fire emissions, however, seasonal variations in atmospheric transport and emissions are considered.

For methane (Figs. c and c), the largest contribution comes from agriculture (64 %), which is mainly due to enteric fermentation, followed by industry (13 %), energy (due to fuel exploitation, 8 %), natural wetland emissions (8 %), waste (7 %), and uncontrolled fire (1 %). Wetland contributions, however, show a clear seasonal cycle, with contributions that exceed the ones from industry and energy during the wet seasons. In contrast to the anthropogenic emissions, the wetland seasonality is a combination of seasonal patterns in transport and emissions, because we use monthly resolved wetland inventory data.

Similar results for CH₄ were reported by at Réunion island, with dominating anthropogenic CH₄ contributions compared to wetland and fire. However, showed that emissions from one of the worldwide largest floodplain wetlands, the Sudd wetland in South Sudan, are highly underestimated, especially by the WetCHARTs product, due to the poor representation of wetland extent and dynamics. Even though the Sudd area is no major origin of air for MKN, their study indicates that wetland emissions may also be underestimated in our results.

In an ideal scenario, if the dispersion model and the inventories were perfect, the resulting simulated time series in Fig.  would perfectly match the measured concentrations at MKN. However, as discussed previously, inventories carry large uncertainties. Furthermore, like for any other dispersion model, tracer transport uncertainties in FLEXPART result from imperfect knowledge of the flow field (input fields) and approximations in the model formulation itself. Both uncertainties lead to differences between the observed and simulated time series.

Similar results have been reported by , who modelled emission contributions to measured CO₂, CH₄ and CO at Réunion island. They found that EDGAR likely overestimates anthropogenic emissions.

Our overall agreement between measured and simulated time series suggests that the inventories for the African continent are broadly reliable in capturing underlying temporal patterns, which is an important insight for validation of regional emission inventories. However, in terms of absolute accuracy, the inventories are at the current state not good enough for accurate chemical modelling of the atmosphere (CO) or assessing the radiative forcing (BC). The underlying issues are well known, mainly related to the inaccurate and highly variable BC and CO emission factors for domestic energy production (cooking, lighting). More regional observations and a complete inversion analysis would be required to verify the quality of the inventories quantitatively. Due to the lack of continuous observations across the continent, this is not possible at the moment. Finally, it is important to note that the sectorial and seasonal variations are not representative for regional or African emissions, because they depend on the origin of air that is transported to MKN as considered by the FLEXPART dispersion model.