Seasonal source variability of carbonaceous aerosols at the Rwanda Climate Observatory

Sub-Saharan Africa (SSA) is a global hot spot for aerosol emissions, which affect the regional climate and air quality. In this paper, we use ground-based observations to address the large uncertainties in the source-resolved emission estimation of carbonaceous aerosols. Ambient fine fraction aerosol was collected on filters at the high-altitude (2590 m a.s.l.) Rwanda Climate Observatory (RCO), a SSA background site, during the dry and wet seasons in 2014 and 2015. The concentrations of both the carbonaceous and inorganic ion components show a strong seasonal cycle, with highly elevated concentrations during the dry season. Source marker ratios, including carbon isotopes, show that the wet and dry seasons have distinct aerosol compositions. The dry season is characterized by elevated amounts of biomass burning products, which approach ∼95 % for carbonaceous aerosols. An isotopic mass-balance estimate shows that the amount of the carbonaceous aerosol stemming from savanna fires may increase from 0.2 µg m⁻³ in the wet season up to 10 µg m⁻³ during the dry season. Based on these results, we quantitatively show that savanna fire is the key modulator of the seasonal aerosol composition variability at the RCO.

1 Introduction

Sub-Saharan Africa (SSA) currently faces major challenges with respect to sustainable development, including industrial development, agriculture, fresh water supply, climate change, energy resources, and air pollution (IPCC, 2014; UNDP, 2018). These challenges are linked, either directly or indirectly, to aerosol emissions. Aerosols offset the ongoing regional climate warming in SSA, shift monsoon and precipitation patterns, and are detrimental to the air quality (IPCC, 2013; WHO, 2016). Ambient air pollution in SSA is estimated to cause 563 000 premature deaths annually, making it one of the main causes of mortality in the region (Bauer et al., 2019). However, the level of scientific understanding of the overall health and climate impact of aerosols is still low, owing to the complex aerosol life cycle during which emissions, transformations, and sinks are associated with large uncertainties, in particular given their vast physical and chemical complexity. A major limiting factor regarding improving our understanding of these effects in SSA is the limited number of in situ observations (Williams et al., 2007; Cais et al., 2011; Kulmala, 2018; López-Ballesteros et al., 2018).

A major source of aerosol emissions in SSA is dry season regional fires, which are clearly visible from space (Fig. 1). These are occasionally ignited naturally by lightning strikes, but they are mainly lit by humans. There is evidence that slash-and-burn agriculture in SSA has been a common practice for thousands of years (Bird and Cali, 1998; Archibald et al., 2012). This long-term anthropogenic perturbation is a significant modulator of current ecosystem structure. A number of studies have specifically focused on characterizing emissions of aerosols and gases from African fires, e.g., the Southern African Regional Science Initiative Project (SAFARI 2000) that was conducted between 1999 and 2001 (Swap et al., 2003). Ground- and airborne chemical characterization from this and other campaigns suggest a rather distinct aerosol chemical composition, including elevated BC, K⁺, and ${\mathrm{NO}}_{\mathrm{3}}^{-}$ concentrations (Table 1).

Figure 1 Fire counts and air mass back-trajectories for the October 2014 to September 2015 campaign at the Rwanda Climate Observatory (RCO, white circle with black outline). The fire counts are from the Fire Information for Resource Management System (FIRMS) derived from the NASA Moderate Resolution Imaging Spectroradiometer (MODIS) satellite product for June–July–August (JJA), 2015. The thin lines represent (04:00 CAT, central African time) 5 d air mass back-trajectories arriving at RCO 100 m above ground level (a.g.l.; 2690 m above sea level, a.s.l.). The blue lines correspond to what we refer to as the “wet” seasons (October–November 2014 and April–May 2015), whereas the green lines represent the dry JJA season.

Table 1 Concentrations of carbonaceous aerosol (µgC m⁻³) and inorganic ions (µg m⁻³) in fine aerosol from ground-based and airborne measurements over sub-Saharan Africa (“bkg” represents background).

^a Present study. ^b Mkoma et al. (2014). ^c Gao et al. (2003). ^d Formenti et al. (2003). ^e Kirchstetter et al. (2003). ^f Sinha et al. (2003). ^g Maenhaut et al. (1996). ^h Puxbaum et al. (2000). ⁱ Brito et al. (2018). ^j Tiitta et al. (2014). ^k Aurela et al. (2016). NA – not available.

Carbonaceous aerosols, which are often quantified as total carbon (TC), are generally divided into two main components: black carbon (BC; here we use elemental carbon, EC, to quantify the amount of BC) and organic carbon (OC). Although they overlap to some extent, these two pools generally have distinct atmospheric life cycles and environmental effects. Formed from incomplete combustion, sunlight-absorbing BC contributes to regional warming and is a particularly detrimental component in air pollution with respect to human health (WHO, 2012; UNEP/WMO, 2012; IPCC, 2013; Bond et al., 2013). BC is chemically inert to atmospheric reactions; thus, its lifetime is mainly determined by deposition. OC is also emitted from incomplete combustion (however, with different emission factors), but it is of non-combustion origin and is formed in the air via secondary processes. OC is thought to have an overall cooling effect on the climate (IPCC, 2013). As it is more chemically reactive, the OC pool has a more complex atmospheric life cycle to some extent, with continuous heterogenous chemistry rendering the lifetime dependent on both precipitation and chemical transformations. Emissions from SSA fires are expected to contribute to a significant part of the global TC atmospheric burden (Liousse et al., 2014).

In general, the actual environmental impact of TC on SSA is poorly constrained. Bottom-up emission projections suggest that the TC emissions from SSA are expected to increase rapidly during the coming decades, perhaps reaching 50 % of the global burden by 2030 (Liousse et al., 2014). To quantify and evaluate such model predictions, as well as to characterize the overall aerosol composition, it is valuable to conduct measurements at regional background sites. Dual carbon isotope characterization (Δ¹⁴C and δ¹³C) of TC at background sites in South and East Asia and the Arctic has been shown to be a valuable tool for quantitatively constraining the emissions from different sources (Gustafsson et al., 2009; Andersson et al., 2015; Sheesley et al., 2012; Kirillova et al., 2014; Winiger et al., 2019).

In this paper, we present dual carbon isotope constraints of TC, along with the chemical characterization of inorganic ions and different carbonaceous pools, from a study conducted at the Rwanda Climate Observatory (RCO) from October 2014 to September 2015. A key objective of the study was to estimate the relative contributions from major TC source categories at this regionally representative site in the SSA. In particular, we investigated the source variability associated with the seasonal variations between the prevailing wet and dry monsoon seasons in the region as well as the contributions from savanna fires.

2 Methods and materials

2.1 Field site and regional meteorology

The sampling site, the Rwanda Climate Observatory (RCO), is located on the top of Mt. Mugogo, in mountainous western Rwanda. (1.586^∘ S, 29.566^∘ E; 2590 m a.s.l. and 5 m a.g.l.). The station was established as a collaboration between the Massachusetts Institute of Technology (MIT; USA) and the Rwandan Government in 2013. The station is described in more detail by DeWitt et al. (2019). RCO is an Advanced Global Atmospheric Gases Experiment (AGAGE) network site (for full list of instruments see http://agage.mit.edu, last access: 17 April 2020).

The meteorology of Rwanda is governed by the East African monsoon, with peak rainfalls in April and November. Thus, there are two dry seasons: December–January–February (DJF) and the main dry season in June–July–August (JJA). The dry seasons in SSA are characterized by extensive biomass burning. During JJA the fires mainly occur to the south of Rwanda (Fig. 1). Savannas are the main biomes in SSA, covering ∼65 % of the landmass, and are the main source of fire emissions (Cahoon et al., 1992). Located in a highly elevated region, Rwanda is, broadly speaking, surrounded by savanna regions, except to the west, where the tropical rainforests of Africa are located.

2.2 Filter sampling

Quartz filter samples (Millipore, 150 mm diameter) were collected with a high-volume sampler operating at 30 m³ h⁻¹ using a PM_2.5 inlet (DH-77, DIGITEL Elektronik AG, Switzerland). Nighttime only (01:00 to 06:00 CAT) sampling was conducted to minimize the effects of local emissions and daytime local atmospheric chemistry and to increase the likelihood of capturing the regional, free troposphere, signals. This strategy is supported by high temporal resolution investigations of the diurnal cycle of, e.g., BC (DeWitt et al., 2019). Each sample was collected over a period of 7 d. The samples were pre-combusted in aluminum foil envelopes (400 ^∘C for 5 h) and were treated with special attention to minimize contamination. The samples were subsequently shipped to Stockholm University for chemical analysis and isolation for carbon isotope analysis. The samples were stored in freezers both on site and at Stockholm University. Field blanks were collected on a monthly basis. The present campaign covers the period from October 2014 to September 2015. However, the period from December 2014 to April 2015 is missing due to a lightning strike that damaged the instrument. Thus, this study presents results from the analysis of filter samples (in total 25) collected for the periods that cover the beginning of the 2014 fall rainy season (October–November), the end of the spring 2015 rainy season (April–May), and the dry 2015 summer season (June–September). We jointly refer the October–November 2014 and the April–May 2015 periods as the wet seasons.

2.3 Concentration analysis

The concentrations of elemental carbon (EC – mass-based tracer for black carbon) and organic carbon (OC) were determined with a thermal–optical instrument (Sunset Laboratory Inc., Tigard, Oregon, USA) using the NIOSH 4050 protocol (Birch and Cary, 1996; Table S1 in the Supplement). Pretreatment using acid fumigation with 1M HCl ensured the efficient removal of carbonates. A glucose solution was used to calibrate the flame ionization detector (FID) response of the instrument, and the long-term performance of the instrument was checked by running National Institute of Standards and Technology (NIST) Standard Reference Materials (SRM) standards. All of the concentrations were blank corrected, and the field blank input was 2 % for OC and 0 % for EC on average. The average relative standard deviation of the triplicate analysis was 5 % for OC and 7 % for EC.

Water-soluble organic carbon (WSOC) was extracted from filter subsamples in ultrapure Milli-Q water by shaking for 1.5 h. The extracts were filtered using 0.45 µm cutoff polytetrafluoroethylene syringe filters (Minisart-SRP 10, Sartorius Stedim Biotech, Germany). The concentration of WSOC was quantified in the filtered solutions as the difference between total water-soluble carbon and water-soluble inorganic carbon using a high-temperature catalytic oxidation instrument TOC-5000A (Shimadzu, Japan). The samples were neither acidified nor purged, in order to avoid the loss of volatile organic compounds. The accuracy of the measurement ranges from 7 % (70 µg L⁻¹) for 1 mg L⁻¹ of carbon solution to 3 % for concentrations higher than 2 mg L⁻¹ of carbon (corresponding to 60 µg L⁻¹). All of the measurements were blank corrected. WSOC field blanks corresponded to an average of 0.5 %. The average relative standard deviation of the triplicate analysis was 10 %.

The concentrations of water-soluble inorganic anions were determined by ion chromatography using a Dionex ICS-2000 system. Anions were separated using an IonPac AG11 2 mm × 50 mm Dionex guard column, an IonPac AS11 2 mm × 250 mm Dionex separation column, and an ASRS 300 self-regenerating suppressor. A solution of KOH was used as eluent. Cations were separated using an IonPac CG16 3 mm × 50 mm Dionex guard column, an IonPac CS11 3 mm × 250 mm Dionex separation column, and a CSRS 300 self-regenerating suppressor. The analysis of cations was performed using 30 mM solution of methanesulfonic acid (MSA) as eluent. Field blanks constituted 3 % of NO³⁻, 2 % of ${\mathrm{SO}}_{\mathrm{4}}^{\mathrm{2}-}$, and 1 % of ${\mathrm{NH}}_{\mathrm{4}}^{+}$ and K⁺ ion concentrations on average. The triplicate analysis showed the average relative standard deviation of 2 % for NO³⁻ and K⁺, 5 % for ${\mathrm{SO}}_{\mathrm{4}}^{\mathrm{2}-}$, and 6 % for ${\mathrm{NH}}_{\mathrm{4}}^{+}$.

2.4 Isotope analysis

Approximately every second sample (n=12) was selected for carbon isotope (Δ¹⁴C and δ¹³C) analysis of total carbon (TC = OC + EC; Table S1). The filter samples were combusted using the Sunset analyzer (total carbon protocol), and the evolved CO₂ was collected in glass vials using a liquid nitrogen cryo-trap (Andersson et al., 2015). The vials were subsequently shipped to the National Ocean Sciences Accelerator Mass Spectrometry (NOSAMS) facility at the Woods Hole Oceanographic Institute (Falmouth, Massachusetts, USA) for analysis of the dual carbon isotope signatures. The Δ¹⁴C signature was measured using accelerator mass spectrometry (AMS), while the δ¹³C signature was measured using an isotope ratio mass spectrometer (IRMS).

2.5 Source Apportionment

The Δ¹⁴C signature allows for differentiation between the relative contributions of biogenic ∕ biomass burning and fossil sources. The biogenic ∕ biomass burning fraction (f_bio) may be calculated using isotopic mass-balance ($f_{\mathrm{fossil}}=\mathrm{1}-f_{\mathrm{bio}}$):

$$\begin{array}{}\text{(1)}& f_{\mathrm{bio}}={\displaystyle \frac{{\mathrm{\Delta }}^{\mathrm{14}}{\mathrm{C}}_{\mathrm{sample}}-{\mathrm{\Delta }}^{\mathrm{14}}{\mathrm{C}}_{\mathrm{fossil}}}{{\mathrm{\Delta }}^{\mathrm{14}}{\mathrm{C}}_{\mathrm{bio}}-{\mathrm{\Delta }}^{\mathrm{14}}{\mathrm{C}}_{\mathrm{fossil}}}}\end{array}$$

The fossil endmember is −1000 ‰, as it is completely depleted in ¹⁴C. The biomass endmember is more complex. For annual plants it is fairly straightforward: the biomass Δ¹⁴C signature equals the Δ¹⁴C value of CO₂ for that year (∼ +20 ‰ for 2014/2015, Graven, 2015; Turnbull et al., 2017). For more long-lived species (e.g., trees), the Δ¹⁴C signature is the average of the atmospheric CO₂ values (weighted by yearly carbon accumulation) over the plants' lifetime. Therefore, bottom-up estimation of Δ¹⁴C_bio requires information regarding the plant distribution in the area of interest and the annual bioaccumulation of carbon for the different plants. As an alternative, we use the combined Δ¹⁴C signature of dissolved organic carbon (DOC) in three of the regions' major rivers, the Congo, the Zambezi, and the Tana, to obtain a regional Δ¹⁴C_bio of $+\mathrm{57}\pm \mathrm{52}$ ‰, which is well within the expected range of a mixture of annual and multiyear plants (Marwick et al., 2015; Wild et al., 2019; Winiger et al., 2019).

The vegetation in SSA may be divided into two main photosynthetic classes: C₃ plants and C₄ plants (see discussion in Sect. 3.5). These two groups have distinct δ¹³C signatures, which allow for isotope-based separation. We may then resolve three source classes by combining Δ¹⁴C and δ¹³C – C₃ plants, C₄ plants, and fossil – via isotopic mass-balance (Andersson et al., 2015):

$$\begin{array}{}\text{(2)}& \begin{array}{rl}\left(\begin{array}{c}{\mathrm{\Delta }}^{\mathrm{14}}\mathrm{C}\left(i\right)\\ {\mathit{\delta }}^{\mathrm{13}}\mathrm{C}\left(i\right)\\ \mathrm{1}\end{array}\right)& =\left(\begin{array}{ccc}{\mathrm{\Delta }}^{\mathrm{14}}{\mathrm{C}}_{{\mathrm{C}}_{\mathrm{3}}}& {\mathrm{\Delta }}^{\mathrm{14}}{\mathrm{C}}_{\mathrm{fossil}}& {\mathrm{\Delta }}^{\mathrm{14}}{\mathrm{C}}_{{\mathrm{C}}_{\mathrm{4}}}\\ {\mathit{\delta }}^{\mathrm{13}}{\mathrm{C}}_{{\mathrm{C}}_{\mathrm{3}}}& {\mathit{\delta }}^{\mathrm{13}}{\mathrm{C}}_{\mathrm{fossil}}& {\mathit{\delta }}^{\mathrm{13}}{\mathrm{C}}_{{\mathrm{C}}_{\mathrm{4}}}\\ \mathrm{1}& \mathrm{1}& \mathrm{1}\end{array}\right)\\ & \cdot \left(\begin{array}{c}{f}_{{\mathrm{C}}_{\mathrm{3}}}\left(i\right)\\ f_{\mathrm{fossil}}\left(i\right)\\ f_{{\mathrm{C}}_{\mathrm{4}}}\left(i\right)\end{array}\right)\end{array}\end{array}$$

Endmember variability may significantly influence the calculated source fractional contributions (Andersson, 2011). For a discussion on the specific endmember ranges used here, see Sect. 3.5.

In Eq. (2), the isotopic data are treated as independent. However, here we find that there is a dependence between the isotope ratios and the TC concentrations; thus, Δ¹⁴C(i) ∼ $A/\mathrm{TC}\left(i\right)+B$, where A and B are constants, and i is the sample index (Fig. 5). This is known as a Keeling relation, and it is discussed in more detail in Sect. 3.4. The relation holds for both Δ¹⁴C (R²=0.85, p<0.01) and δ¹³C, whereas the correlation is weaker for δ¹³C (R²=0.55, p<0.1). A method for using correlations within the framework of Bayesian source apportionment has recently been developed (Martens et al., 2019). The rationale is based on both statistical concepts and the averaging expected from atmospheric mixing. The endmember ranges used in the calculations are from isolated sources, but during long-range transport the variability within a given source, e.g., savanna fires, will be reduced. Using correlations between data points, a means for accounting for the mixing is obtained, and more realistic source fraction estimates are derived. When using the estimated source fractions to back-calculate the isotope signatures, the agreement is good compared with direct fits (Figs. 5, S2). A sensitivity analysis is discussed in Sect. 3.5 (Fig. S3)

Thus, to account for the correlations in the dataset, we add a second constraint in the source apportionment calculations that is based on the relation to the TC concentrations:

$$\begin{array}{}\text{(3)}& \begin{array}{rl}\left(\begin{array}{c}{f}_{{\mathrm{C}}_{\mathrm{3}}}\left(i\right)\\ f_{\mathrm{fossil}}\left(i\right)\\ f_{{\mathrm{C}}_{\mathrm{4}}}\left(i\right)\end{array}\right)& ={\displaystyle \frac{\mathrm{1}}{\left[\mathrm{TC}\left(i\right)\right]}}\cdot \left(\begin{array}{c}{f}_{{\mathrm{C}}_{\mathrm{3}},\mathrm{slope}}\\ f_{\mathrm{fossil},\mathrm{slope}}\\ f_{{\mathrm{C}}_{\mathrm{4}},\mathrm{slope}}\end{array}\right)\\ & +\left(\begin{array}{c}{f}_{{\mathrm{C}}_{\mathrm{3}},\mathrm{intercept}}\\ f_{\mathrm{fossil},\mathrm{intercept}}\\ f_{{\mathrm{C}}_{\mathrm{4}},\mathrm{intercept}}\end{array}\right)\end{array}\end{array}$$

Here, instead of fitting a source vector ($f_{{\mathrm{C}}_{\mathrm{3}}}$, f_fossil, $f_{{\mathrm{C}}_{\mathrm{4}}}$) for each individual data pair, we fit two vectors – a slope and an intercept of the line – to all data points. This clearly offers the advantage of have fewer fitting parameters. We emphasize that the strength of the correlation of the isotope signatures relative to 1∕TC is naturally incorporated into this relation; hence, the lower correlation of δ¹³C with respect to 1∕TC imposes weaker constraints on the calculated source fractions compared with Δ¹⁴C.

The source fractions were computed using numerical Markov chain Monte Carlo simulations, implemented in MATLAB, version 2015b, using 1 000 000 iterations with a burn-in (initial search phase) of 10 000 and a data thinning of 10 (removing stepwise correlations). The stochastic perturbation parameter was adjusted in order to obtain an acceptance ratio of 0.23, which has been suggested to be optimal for Metropolis–Hastings algorithms (Roberts et al., 1997). For this setup, the variability in the numerically estimated parameters, e.g., the standard deviation of the relative source fraction, is lower than 1 % of the mean value, suggesting good convergence (Winiger et al., 2017).

2.6 Air mass back-trajectories and remote sensing

Using the NOAA Hybrid Single Particle Lagrangian Integrated Trajectory (HYSPLIT) model, 10 d air mass back-trajectories at an arrival height of 2690 m a.s.l. (100 m a.g.l.) and 3090 m a.s.l. (500 m a.g.l.) were calculated (Fig. 1, Fig. S1 in the Supplement). Remote sensing fire-spot detections were retrieved from the NASA Fire Information for Resource Management Services (FIRMS) database, based on retrievals from the Moderate Resolution Imaging Spectroradiometer (MODIS) satellite product.

3 Results and discussion

3.1 Back-trajectory analysis

Air mass back-trajectory analysis shows that the air masses arriving at the Rwanda Climate Observatory (RCO) during the filter collection periods are generally from the east/southeast (Figs. 1, S1). There is some overlap between the wet and boreal summer dry seasons, but there is a seasonal switch overall: the wet seasons air masses are from a more easterly origin (e.g., Uganda, Kenya, and Tanzania), whereas the dry season air masses are from more southeasterly directions (e.g., Burundi, Tanzania, and the Democratic Republic of Congo). During the dry season there are extensive fires to the south of RCO, mainly to the southwest (Democratic Republic of Congo and Angola). However, the air masses also pass over regions with comparably high fire activities in the southeast, mainly in Tanzania. Nevertheless, we emphasize that back-trajectory analysis is challenging in mountainous regions (e.g., Winiger et al., 2019), and the actual geographical footprints are expected to be broader, e.g., due to the propagating effects of turbulence. Here, we interpret the back-trajectories qualitatively to visualize overall air mass transport patterns.

3.2 Concentrations of fine aerosol components

During the present campaign, the PM_2.5 carbonaceous and inorganic ion components show a strong seasonal variability, with elevated levels during the dry JJA season (Fig. 2, Table S1). The dry ∕ wet season ratios for TC, EC, WSOC, ${\mathrm{NO}}_{\mathrm{3}}^{-}$, ${\mathrm{SO}}_{\mathrm{4}}^{\mathrm{2}-}$, ${\mathrm{NH}}_{\mathrm{4}}^{+}$, and K⁺ were 4.2, 7.0, 4.1, 12.6, 3.0, 3.2, and 8.8, respectively (Table 1). The sea-salt contributions to the ions are generally estimated to be less than 1 %, using corrections with sodium ions (Blanchard and Woodcock, 1980). Here, we report the actual concentrations to facilitate direct comparisons with previous studies (Table 1). Overall, these differences reflect differences in the aerosol atmospheric lifetime, air mass transport pathways, and emission seasonality (e.g., fires). Elevated ratios of EC and K⁺ suggests an increased influence from biomass burning during the dry season. ${\mathrm{NO}}_{\mathrm{3}}^{-}$, which displays the largest seasonal shift, is often associated with oxidized NO_x from traffic emissions or lightning strikes; however, it is also typically elevated in emissions from savanna fires (Table 1; e.g., Gao et al., 2003; Formenti et al., 2003).

Figure 2 The time dependence of the concentrations of carbonaceous aerosols (TC denotes total carbon; EC denotes elemental carbon; OC denotes organic carbon; WSOC denotes water-soluble organic carbon) and inorganic ions in PM_2.5 from October 2014 to September 2015 at the Rwanda Climate Observatory. Panel (a) shows TC and EC, panel (b) shows OC and WSOC, panel (c) shows ${\mathrm{NO}}_{\mathrm{3}}^{-}$ and ${\mathrm{NH}}_{\mathrm{4}}^{+}$, and panel (d) shows K⁺ and ${\mathrm{SO}}_{\mathrm{4}}^{\mathrm{2}-}$. Instruments were hit by lightning, resulting in a data gap from November 2014 to April 2015. For visual clarity, the concentrations of EC were multiplied by 5, and the concentrations of K⁺ were multiplied by 2.

The dry season concentrations of carbonaceous aerosol components and inorganic ions reported here show good overall agreement with the concentrations observed in dry season rural and aged savanna fire air masses (Table 1). The BC values are in the same range as those observed at Mt. Kenya (0.72±0.06 µgC m⁻³, Gatari and Boman, 2003). During the atmospheric aging of a biomass plume, the values of OC, EC, and K⁺ decrease by a factor of 2–3, whereas other components are relatively unaffected (Table 1). However, the effects appear variable compared with savanna fires in South Africa (Gao et al., 2003).

The RCO is situated not far from the Nyiragongo and Nyamuragira volcanoes in the eastern Democratic Republic of Congo. High spatial resolution (13×24 km²) satellite monitoring of the SO₂ levels shows near-constant emissions from these volcanoes over the time period covering the present campaign, which likely affected the observed sulfate levels (Barrière et al., 2017). Here we observe a spike in sulfate levels (∼5 µg m⁻³) during the week starting on the 13 June 2015 (Fig. 2) but with no clear linkage to an increase in volcanic SO₂ emissions.

3.3 Source marker ratios and correlations

Overall, the ratios of different aerosol components provide insights into sources or atmospheric processes. Here, the EC∕TC ratio shows a distinct seasonality (Fig. 3, Table S1). More commonly analyzed, however, is the OC∕EC ratio (calculated as $(\mathrm{TC}-\mathrm{EC})/\mathrm{EC}$), which shows elevated levels during the wet season (11±3) compared with the dry season (7±3; Table S1). The OC∕EC ratio is sometimes used as a marker for biomass burning, but it is highly influenced by burning conditions such as flaming or smoldering fires. In addition, it is highly impacted by atmospheric processes such as secondary organic aerosol (SOA) formation or photochemical aging (e.g., Dasari et al., 2019). The dry season values observed here are similar to those that have been observed in background air at other dry season sub-Saharan African sites (Table 1).

Figure 3 The time dependence of ratios of carbonaceous aerosols (EC denotes elemental carbon; OC denotes organic carbon; WSOC denotes water-soluble organic carbon) and inorganic ions relative to total carbon (TC) in PM_2.5 from October 2014 to September 2015 at the Rwanda Climate Observatory. Panel (a) shows the EC∕TC ratio, panel (b) shows the WSOC∕OC ratio, panel (c) shows the ${\mathrm{NO}}_{\mathrm{3}}^{-}/\mathrm{TC}$ and ${\mathrm{K}}^{+}/\mathrm{TC}$ ratios, and panel (d) shows the ${\mathrm{NH}}_{\mathrm{4}}^{+}/\mathrm{TC}$ and ${\mathrm{SO}}_{\mathrm{4}}^{\mathrm{2}-}/\mathrm{TC}$ ratios. Instruments were hit by lightning, resulting in a data gap from November 2014 to April 2015. The concentrations of the ${\mathrm{K}}^{+}/\mathrm{TC}$ and ${\mathrm{NH}}_{\mathrm{4}}^{+}/\mathrm{TC}$ ratios were multiplied by 2 for visual clarity.

Similarly, the ${\mathrm{NH}}_{\mathrm{4}}^{+}/\mathrm{TC}$ and ${\mathrm{SO}}_{\mathrm{4}}^{\mathrm{2}-}/\mathrm{TC}$ ratios are also elevated during the wet seasons (Fig. 3), whereas they decrease during the dry seasons, suggesting a different source profile compared with EC, K⁺, and ${\mathrm{NO}}_{\mathrm{3}}^{-}$, including potential volcanic input from SO₂. In contrast, the WSOC∕OC ratio shows no clear seasonality, indicating small differences in sources and atmospheric processing of water-soluble and water-insoluble organic components during the year. TC correlates with K⁺ (an R² of 0.95 and p<0.01) and ${\mathrm{NO}}_{\mathrm{3}}^{-}$ (an R² of 0.95 and p<0.01), suggesting that the incomplete combustion regime during the present campaign is governed by biomass emissions, e.g., savanna burning. Based on these results, these ratios qualitatively suggest that the aerosol regime at RCO is strongly influenced by the occasional input of biomass burning products during the boreal dry season.

3.4 Carbon isotopes

Radiocarbon (Δ¹⁴C) and stable carbon (δ¹³C) provide detailed information regarding the sources and atmospheric processing of carbonaceous aerosols. Here, we investigated the signatures of TC for roughly every second sample during the campaign. The Δ¹⁴C marker is not influenced by atmospheric processing and may be used to directly compute the relative contributions of fossil and biomass ∕ biogenic sources with high precision, Eq. (1). The Δ¹⁴C signature shows an oscillation over the seasons, ranging between −84 ‰ in November 2014 and +30 ‰ in July 2015 (Fig. 4, Table S1). Thus, during the JJA season, the Δ¹⁴C signature exceeds the signature for atmospheric CO₂ (+20 ‰, Graven, 2015; Turnbull et al., 2017).

Figure 4 Dual carbon isotope data for TC vs. time. Panel (a) shows Δ¹⁴C, and panel (b) shows δ¹³C. The uncertainties for Δ¹⁴C are below 50 ‰, and the uncertainties for δ¹³C are below ∼0.2 ‰. Instruments were hit by lightning, resulting in a data gap from November 2014 to April 2015.

Using Eq. (1), the percentage biomass ∕ biogenic TC for the Δ¹⁴C = +30 ‰ sample is 97 %. During the wet season, the percentage derived from fossil sources reaches 13 %, which possibly points to a more local derivation. Δ¹⁴C correlates with 1∕TC (an R² of 0.85 and p<0.01), which suggests that the variability in concentrations can be explained as a two-component mixture: a stable background and a temporally fluctuating source (Keeling, 1958; Fig. 5a). This inverse relation gives a Δ¹⁴C of $+\mathrm{37}\pm \mathrm{6}$ ‰ as TC approaches infinity, showing that the non-background component is of biogenic ∕ biomass burning origins. The Δ¹⁴C signatures for TC reported here are generally higher than for monitoring sites in South and East Asia (Sheesley et al., 2012; Kirillova et al., 2014; Bikkina et al., 2016).

Figure 5 Interrelations of carbon isotope signatures and TC (blue circles). Panel (a) shows Δ¹⁴C vs. TC, and panel (b) shows δ¹³C vs. TC. The black line is the mean fit of the equation Δ¹⁴C, ${\mathit{\delta }}^{\mathrm{13}}\mathrm{C}=A/\left[\mathrm{TC}\right]+B$, using Markov chain Monte Carlo simulations, where A and B are fitting parameters. For Δ¹⁴C, A is $-\mathrm{135}\pm \mathrm{16}$ ‰ µg m⁻³ and B is 37±6 ‰. For δ¹³C, A is $-\mathrm{5.8}\pm \mathrm{1.5}$ ‰ and B is $-\mathrm{21.8}\pm \mathrm{0.6}$ ‰ µg m⁻³. The gray shaded areas display the 1σ spread of the fit.

In contrast to Δ¹⁴C, the δ¹³C value is influenced by both atmospheric processes (i.e., kinetic isotope effects, KIEs) and source signatures. Here, the δ¹³C value shows a similar pattern relative to the Δ¹⁴C value: it is depleted in ¹³C (min δ¹³C of−27 ‰) during wet seasons and higher during the dry season (max δ¹³C of −21 ‰) (Fig. 4b). The correlation with respect to 1∕TC (an R² of 0.55 and p<0.1) is weaker than that for Δ¹⁴C (Fig. 5b). The direct fossil vs. biomass source correlation from the Δ¹⁴C Keeling curve is also driving the δ¹³C signatures, but the higher variability is explained by larger endmember variability and the potential influence of KIEs (see Sect. 3.5). An overall enrichment in ¹³C has been found in aged air masses in South Asia, especially for WSOC (Kirillova et al., 2013; Dasari et al., 2019), although less so for TC. In fact, the enrichment of ¹³C in WSOC often appears to be counteracted by a decrease in water-insoluble OC (e.g., Yan et al., 2017; Fang et al., 2017).

The TC δ¹³C values, and their seasonal trend, are similar to those previously observed in fine aerosols at a rural site in Tanzania (May–August 2011; Mkoma et al., 2014). However, the temporal trend appears shifted in the RCO samples from values of around −25 ‰ to around 22 ‰ in mid-May. At the Tanzanian site, a similar shift occurs in mid-June. In addition to the complications of comparing measurements conducted at different sites during different years, there is good agreement in the δ¹³C values, and the temporal offset may be explained by variability in the position of the intertropical convergence zone. Similarly, the δ¹³C value for TC at a savanna woodland site in Zambia, observed during August–September 2000, was $-\mathrm{21.8}\pm \mathrm{0.8}$ ‰ (Billmark et al., 2005), while values between −19.3 ‰ and −23.6 ‰ were observed at sites in Côte d'Ivoire (Cachier et al., 1985).

3.5 Carbon isotope-based source apportionment

By combining the Δ¹⁴C and the δ¹³C values, we can resolve three major sources of TC at the RCO using isotopic mass-balance, Eq. (2). However, there are some important considerations to this approach: first, the δ¹³C value is not an exclusive source marker and is also affected by atmospheric processing (e.g., photochemical oxidation and secondary formation); second, the main source categories must be defined and distinguishable using carbon isotopes; and third, the source values of the isotope signatures, the endmembers, and their natural variability need to be established.

As mentioned, the δ¹³C value of bulk TC appears to be considerably less affected by atmospheric processing compared with subcomponents, such as WSOC. Here, the temporal variation of the δ¹³C value is qualitatively similar to that of the Δ¹⁴C value (Fig. 4). As Δ¹⁴C is not affected by atmospheric reactions, this suggests that source variability is a key driver of the δ¹³C variability. Furthermore, the WSOC∕OC ratio is virtually constant throughout the year (Fig. 3); the WSOC∕OC ratio has been found to be highly affected by atmospheric processing and related to shifting δ¹³C (Kirillova et al., 2013; Yan et al., 2017; Fang et al., 2017; Dasari et al., 2019). Here, we therefore assume that the δ¹³C ratio of TC is not strongly perturbed by atmospheric processing during long-range transport and may, thus, be used as a source marker. Nevertheless, we explore the potential influence of the KIEs, as well as endmember variability, on the source apportionment results in a sensitivity analysis.

Turning to potential sources, there is a multitude of potential source categories for TC in SSA. However, many of these falls into broad categories with similar carbon isotope signatures. Around the world, the applications of dual carbon isotope source apportionment techniques in ambient TC has mainly identified/considered six broad source categories: C₃ plants, C₄ plants, liquid fossil fuels (e.g., traffic), coal combustion (solid fossil), gas flaring (gaseous fossil), and marine emissions (Winiger et al., 2019; Andersson et al., 2015; Kirillova et al., 2013). Overall, the practice of coal combustion in SSA is expected to be much less frequent than in, for example, South and East Asia; therefore, we do not consider this source further. In addition, marine emissions are not expected to have a large influence at RCO, which is supported by the low estimates of marine contributions to the inorganic ions (<1 %). For gas flaring, there are potential distant sources around the Arabian Peninsula and off the west coast of Africa, in the Gulf of Guinea. However, given the distances to the RCO station and the prevailing wind directions, emissions from flaring are not expected to affect the site, while the δ¹³C signatures for gas flaring are strongly depleted in ¹³C (δ¹³C < −38 ‰; Winiger et al., 2017) and even a small contribution would shift the observed values significantly.

The remaining three main source categories are the two biomass sources of C₃ (e.g., trees) and C₄ plants (e.g., sugarcane and certain grasses) and liquid fossil fuels (Fig. 6). Aerosols from liquid fossil fuel sources have a Δ¹⁴C_fossil of −1000 ‰ (completely depleted in ¹⁴C) and a δ¹³C_fossil of $-\mathrm{25.5}\pm \mathrm{1.3}$ ‰ (Widory, 2006; Andersson et al., 2015). The Δ¹⁴C of biomass was established in Sect. 2.4 as ${\mathrm{\Delta }}^{\mathrm{14}}{\mathrm{C}}_{{\mathrm{C}}_{\mathrm{3}}}={\mathrm{\Delta }}^{\mathrm{14}}{\mathrm{C}}_{{\mathrm{C}}_{\mathrm{4}}}=+\mathrm{57}\pm \mathrm{52}$ ‰. The δ¹³C of C₃ plants in general is $-\mathrm{27.1}\pm \mathrm{2}$ ‰ (Bender, 1971; O'Leary, 1988). However, for aerosols generated from C₃ plants, this value may be either enriched (e.g., ∼0.5 ‰ biomass burning) or depleted (e.g., ∼0 ‰ to 4 ‰ during SOA formation) (Turekian, 1998; Das et al., 2010; Mkoma et al., 2014; Aguilera and Whigham, 2018). In any case, the numerical spread in the δ¹³C value of these different sources is largely overlapping with that of the raw materials; therefore, we use this value here. The δ¹³C of C₄-plant materials is $-\mathrm{13.1}\pm \mathrm{1.2}$ ‰ (Bender, 1971; O'Leary, 1988; Turekian et al., 1998). However, during incomplete combustion, the ${\mathit{\delta }}^{\mathrm{13}}{\mathrm{C}}_{{\mathrm{C}}_{\mathrm{4}}}$ may be reduced by a factor ranging from 0 ‰ to 7 ‰, which is largely dependent on burning conditions and species (Martinelli, 2002; Das et al., 2010; Aguilera and Whigham, 2018). Accounting for such effects in source apportionment is a challenge, especially as the reported values are ranges and not mean and variability; thus, they are highly influenced by potential outliers. Here, we use a method discussed in Andersson et al. (2015) to address the issue of the statistical analysis of ranges by assuming that the total range corresponds to the 95 % confidence intervals of a normal distribution. This corresponds to the range of 4 times the standard deviation, yielding $\mathit{\sigma }=\mathrm{7}/\mathrm{4}$ ‰, while the mean is $-\mathrm{7}/\mathrm{2}$ ‰. Combining this with the variability of the of pure C₄ plants we obtain the following: ${\mathit{\delta }}^{\mathrm{13}}{\mathrm{C}}_{{\mathrm{C}}_{\mathrm{4}}}$: $-\mathrm{16.6}\pm \mathrm{2.2}$ ‰, where ${\mathit{\sigma }}^{\mathrm{2}}={\mathrm{1.2}}^{\mathrm{2}}+(\mathrm{7}/\mathrm{4}{)}^{\mathrm{2}}$ ‰². These values are also what is obtained by numerical estimation of the convolution of a normal distribution ($\mathit{\mu }=-\mathrm{16.6}$ ‰, σ=1.2 ‰) with a uniform distribution ([$-\mathrm{7},\mathrm{0}$] ‰), adding to the strength of the statistical representation.

Figure 6 Dual carbon (Δ¹⁴C vs. δ¹³C) isotope plot of TC. Blue circles represent October–November 2014 (wet), yellow circles represent April–May 2015 (wet), and red circles represent June–September 2015 (dry). The boxes represent the endmember ranges (mean ± SD; see Sect. 3.5 for details) of the three main sources: C₃ plants (green), C₄ plants (orange), and fossil (black).

The fractional source contributions of fossil fuel, C₃, and C₄ to TC are computed using Eqs. (2) and (3) (Fig. 7). It is well established that the accurate estimation of the fractional source contributions requires explicit incorporation of the endmember variability, and we use a Bayesian framework driven by Markov chain Monte Carlo simulations for this purpose (Andersson, 2011; Andersson et al., 2015). To estimate the influence of the intra-endmember mixing during atmospheric transport, we use the correlations of the isotopes with TC within the Bayesian framework (see Sect. 2.5) to account for the endmember averaging during atmospheric transport (Martens et al., 2019; Fig. 5). The resulting fractional contributions display a large variability when comparing wet and dry conditions (Fig. 7a, Table S2). The dry season is characterized by relatively higher C₄-plant contributions, whereas the relative contributions of fossil fuels and C₃ plants increase during the wet seasons. Back-calculating the isotope signatures from the computed source fractions from the MCMC simulations essentially reproduces the Keeling relations relative to 1∕TC (Figs. 5, S2). To check the influence of the number of data points used in the Keeling-based MCMC, we computed comparative scenarios where every third data point was used (starting at data point 1, 2, and 3, respectively) (Fig. S3). The standard deviations for the calculated $f_{{\mathrm{C}}_{\mathrm{3}}}$ are doubled on average when only every third point is used (5 % vs. 10 %), showing how correlations between multiple data points aid in constraining the sources.

Figure 7 Carbon isotope source-segregated fractions and concentrations of TC vs. time computed with the “best endmember scenario”. (a) Relative source contributions (%) of C₃ plants (green circles), C₄ plants (orange diamonds), and fossil (black triangles). Estimated savanna contributions are shown using blue squares. (b) Source-segregated concentrations of TC of C₃ plants (green circles), C₄ plants (orange diamonds), and fossil (black triangles). The error bars (standard deviations) were calculated using Markov chain Monte Carlo simulations.

As the δ¹³C endmembers are not well constrained (in particular for C₄ plants), we also employed a sensitivity analysis with respect to endmembers and the potential influence of the KIEs (Tables S2, S3, S4, S5; Figs. S4, S5, S6). In addition to the abovementioned best estimate scenario, we tested two ${\mathit{\delta }}^{\mathrm{13}}{\mathrm{C}}_{{\mathrm{C}}_{\mathrm{4}}}$ scenarios: a “minimum KIE scenario” with zero KIE (${\mathit{\delta }}^{\mathrm{13}}{\mathrm{C}}_{{\mathrm{C}}_{\mathrm{4}}}$ of $-\mathrm{13.1}\pm \mathrm{1.2}$ ‰) and a “maximum KIE scenario” with a depletion of 5.9 ‰ (${\mathit{\delta }}^{\mathrm{13}}{\mathrm{C}}_{{\mathrm{C}}_{\mathrm{4}}}$ of $-\mathrm{19.0}\pm \mathrm{2.2}$ ‰). The maximum KIE scenario was established such that the $f_{{\mathrm{C}}_{\mathrm{4}}}/(f_{{\mathrm{C}}_{\mathrm{4}}}+f_{{\mathrm{C}}_{\mathrm{3}}}$) ratio would be 62 % as TC approach infinity, which amounts to 100 % savanna contributions, see Eq. (4). As expected, these scenarios significantly shift the estimated relative C₄ contributions, resulting in a total range of the sample period averages of 24 % (min KIE; min 6 %, max 32 %) to 42 % (max KIE; min 10 %, max 58 %), thereby providing lower and upper bounds (Figs. S4, S5; Tables S3, S4). The corresponding value for our best estimate is 32 % (max 44 %, min 8 %). In addition, we investigated a scenario with a 3 ‰ depletion of the fossil endmember (δ¹³C_fossil of $-\mathrm{28.5}\pm \mathrm{1.3}$ ‰). As the fossil contribution is low overall, as determined by Δ¹⁴C, and as Δ¹⁴C constrains the fossil contribution independently of the δ¹³C data, this shift has no significant influence on the computed source fractions of 6 % (max 11 %, min 3 %; Fig. S6; Table S5). Overall, we stress that these three sensitivity test scenarios represent extreme limits, and the a priori least biased scenario is the initially outlined best scenario.

By combining the estimated fractional source contributions with the TC concentrations, we can estimate the concentrations from the different sources (Fig. 7b), revealing a more accentuated source variability. The average dry-to-wet ratios of the TC to C₃ plants, C₄ plants, and fossil fuels are 3, 5, and 2, respectively.

Savannas are the main biome supporting C₄ plants in SSA. For East African savannas, δ¹³C data suggest that ∼62 % ($f_{{\mathrm{C}}_{\mathrm{4}},\mathrm{NPP}}$) of the net primary production (NPP) is from C₄ plants (the rest mainly stems from C₃ plants; Lloyd et al., 2008). Thus, one may assume that the source characteristics of TC emitted from savanna burning should represent this plant-signature distribution. However, the aerosol emissions modulate the NPP activity through emission factors (EF). The uncertainties of EFs from different biomass burning activities are generally large and overlapping (Andreae, 2019). As a first approximation, we use $f_{{\mathrm{C}}_{\mathrm{4}},\mathrm{NPP}}$ to estimate the fractional contribution of savanna emissions to TC (f_savanna) as follows:

$$\begin{array}{}\text{(4)}& f_{\mathrm{savanna}}\left(i\right)={\displaystyle \frac{f_{{\mathrm{C}}_{\mathrm{4}}}\left(i\right)}{f_{{\mathrm{C}}_{\mathrm{4}},\mathrm{NPP}}}},\end{array}$$

where i is the sample index.

This analysis shows that the dry season carbonaceous aerosol regime is dominated by savanna fire emissions reaching up to 71 % (Fig. 7, Table S2). These results agree with the elevated levels of EC, K⁺, and ${\mathrm{NO}}_{\mathrm{3}}^{-}$ during the dry season (Table S1).

4 Outlook

In this paper, we find that the aerosol composition of the emissions affecting the Rwanda Climate Observatory (RCO) may be described as a two-state source mixture: a regional/local background signal modulated by savanna fire emissions. Multiple studies have shown that savanna fires strongly influence the aerosol regime in SSA. Here, we estimate the savanna fire contributions for carbonaceous aerosols to range from 13 % (wet season; TC_savanna of 0.2 µg m⁻³) to 71 % (dry season; TC_savanna of 9.7 µg m⁻³) at RCO (Fig. 7). The savanna fires are believed to be mainly lit by humans and, although these activities have possibly been ongoing throughout the Holocene, they strongly perturb the regional ecosystems, climate, and air quality (e.g., Bird and Cali, 1998; Archibald et al., 2012). The annual SSA savanna carbon budget represents a net CO₂ source to the atmosphere (Williams, 2007; Cais et al., 2011; Valentini et al., 2014; Palmer et al., 2019). Therefore, finding more sustainable alternatives to the slash-and-burn practices in SSA could possibly turn the region into a carbon sink. For instance, the implementation of early dry season burning may be a possible strategy (Lipset-Moore et al., 2018). Savanna fire mitigation would also improve the regional air quality and stabilize precipitation patterns (Hodnebrog et al., 2015; Heft-Neal et al., 2018; Bauer et al., 2019).

Nevertheless, the current level of scientific understanding of the impact of savanna burning on the environmental system is poor, as are the couplings/responses to climate change, population growth, urbanization, and other key socioeconomic and environmental challenges for sustainable development in SSA (e.g., IPCC, 2014; Liousse et al., 2014; Brandt et al., 2017; UNDP, 2018). Savanna burning mitigation, or shifts in fire regime due to climate change, may change the present steady state in unpredictable ways (e.g., Abreu et al., 2017). To better constrain the multiple environmental impacts of savanna burning in SSA, the comparably few ongoing ground-based in situ observations should be expanded and solidified (Williams et al., 2007; Cais et al., 2011; Kulmala, 2018; López-Ballesteros et al., 2018). For instance, observations of source-segregated aerosol concentrations provide multiple opportunities for advancing our knowledge of environmental processes relevant to SSA, including providing means for testing chemical-transport models; examining the relative importance of cooling vs. warming (e.g., BC) aerosols; ground-truthing remote sensing products; and detailed monitoring of the expected rapid change over the coming decades, including the effects of climate warming, population growth, and urbanization.