2.1. Sampling sites and strategy

Sampling sites were installed in Mosquera (in Sabana Occidente county) and Soacha (in Soacha county), western neighboring towns of Bogotá’s Capital District, both influenced by industrial activities and the transit of a high volume of trucks, buses, and vehicles. The sampling sites were separated 13.1 km. The Mosquera sampling site (4º42’9.75” N, 74º13’54.94” W) was located at a 7-m high monitoring platform regularly used for manual particulate matter monitoring, in the middle of the Villa Olímpica public sports fields. Mosquera has ca. 83,000 inhabitants and has an important industrial corridor, with food and beverage, metalwork, and plastics sectors contributing significantly. The Soacha sampling site (4º35’4.59” N, 74º13’11.62” W) was located on the roof of the 5-story Mario Gaitán Yanguas Public Hospital building. The hospital is on a narrow but congested street, with a medium volume of vehicles and small buses, at 620 meters from Autopista Sur, the main corridor connecting Bogotá with southern and western Colombia and one of the most congested highways in the country, particularly because of a massive volume of heavy-duty trucks. Soacha also has a large industrial corridor, with mining for construction materials a significant activity, which keeps growing at an alarming rate under insufficient planning provisions.

Simultaneous sampling of 24-h integrated PM_2.5 filters was performed at each site between 12 September 2014 and 11 November 2014 using paired samplers. In Soacha, three 16.7 lpm speciation samplers (Rupprecht & Patashnick, USA), one loaded with a PTFE filter (Zefluor membrane, 2 µm, 47 mm, Pall Corporation, USA) and two loaded with quartz filters (QMA, 47 mm, Whatman, USA), were used. In Mosquera, one 16.7 lpm speciation sampler was loaded with a PTFE filter and one high volume sampler equipped with a brushless monitor (TE-1000, 200 lpm, Tisch Environmental, USA) was loaded with a 100-mm quartz filter, cut from a larger sheet (QM-A, 203.2 x 254 mm, Whatman, USA). PTFE filters were used to determine PM_2.5 mass, ions and trace metals, and quartz filters to determine ECOC and organic tracer speciation. Before and after sampling, filters were stored and transported in Petri dishes, with internal aluminum foil holders and an external Teflon tape seal, refrigerated at -20ºC until chemical analyses.

To ensure an appropriate mass loading for chemical speciation, filters compositing was performed. Composites from stage 1 corresponded to 3 groups of consecutive days of the same week: Monday-Tuesday, Wednesday to Friday, and Saturday-Sunday, which were analyzed for WSOC and ions. Any differences in these components between weekdays and weekends could be detected in this way. Composites from stage 2 corresponded to weekly groups, which were analyzed for metals and organics. As a result, 25 composites from stage 1 and 9 weekly composites of each filter material at each site were formed.

2.2. PM _2.5 mass concentration and chemical analysis

Teflon filters used for gravimetric analysis were conditioned at 20±3ºC and 35±5% for 24-48 h before being weighed with a microbalance (MX5, ±1 µg, Mettler Toledo, USA). Each filter was weighed three times and re-weighed if a difference higher than 15 µg in consecutive weights was found.

Three samples per week from each site were analyzed for organic carbon (OC), elemental carbon (EC), Water-Soluble Organic Carbon (WSOC), and water-soluble inorganic ions (WSII). EC and OC were determined through the NIOSH thermal optical transmission method and flame ionization detection (FID) (^{Schauer et al., 2003}) on a 1.0 cm² quartz filter punch in a Thermal Optical Analyzer (Sunset Laboratories, Forest Grove, OR, USA). WSOC was determined using a TOC-V SCH Shimadzu total organic carbon analyzer on an extract of a quarter of each PTFE filter in 15 mL of Milli-Q water (^{Miyazaki et al., 2011}; ^{Yang et al., 2003}). The difference OC - WSOC produced the Water-insoluble organic carbon (WIOC), with an uncertainty calculated by uncertainty propagation. WSII were analyzed using ion chromatography (IC) (Dionex ICS 2100 and Dionex ICS 100) (^{Wang et al., 2005}). Seven ions were measured: sulfate, nitrate, chloride, sodium, ammonium, potassium, and calcium.

Weekly composites for each site with at least 500 μg of OC were prepared to analyze organic compounds by gas chromatography/mass spectrometry (GC-6980, quadrupole MS-5973, Agilent Technology). Isotopically-labeled standard solutions were used to spike the filter composites before the extraction using four alternate and equal volumes of dichloromethane and acetone, each one sonicated for 20 minutes. Solvents were evaporated in a rotavapor and reduced in volume by blowing ultrapure nitrogen. Two aliquots of each extract were analyzed by GC-MS. In the first one, diazomethane was used to derivatize carboxylic acids. In the second one, a silylating reagent derivatized hydroxyl groups (^{Nolte et al., 2002}; ^{Stone et al., 2008}). A sample of SRM 1649a (Urban Dust, NIST) and a standard spike sample were analyzed together with each batch of samples for quality control. All concentrations were blank-corrected. Uncertainties were estimated using the detection limits of the instruments and the standard deviation of field blanks. Details on the analytical methods can be found elsewhere (^{Villalobos, 2015}; ^{Miyazaki et al., 2011}; ^{Stone et al., 2008}; ^{Wang et al., 2005}; ^{Schauer et al., 2003}; ^{Yang et al., 2003}; ^{Nolte et al., 2002}).

2.3. Source apportionment

The EPA Chemical Mass Balance model software CMB v8.2 (^{Watson et al., 1984}) was used to estimate primary source contributions to organic carbon (OC) for the eight weekly composite samples at each site. Organic compounds included in the model were selected for their stability and low volatility during their transport in the atmosphere (^{Schauer et al., 1996}): polycyclic aromatic hydrocarbons (such as benzo(b)fluoranthene, benzo(k)fluoranthene, benzo(e)pyrene, indeno(1.2.3-cd)pyrene, benzo(ghi) perylene, ABB-20R-C27-cholestane, ABB-20R-C29-sitostane, ABB-20S-C29-sitostane) as tracers of vehicle emissions and biomass burning (^{Ravindra et al., 2008}); picene as a tracer of coal combustion (^{Oros and Simoneit, 2000}); hopanes (such as 17α(H)-22. 29. 30-trisnorhopane, 17α(H)-21β(H)-30-norhopane, 17α(H)-21β (H)-hopane) as tracers of gasoline, diesel, and fuel oil combustion (^{Rogge et al., 1997}; ^{Schauer et al., 1996}; ^{Shrivastava et al., 2007}); and levoglucosan as a tracer of biomass burning (^{Bi et al., 2008}; ^{Fine et al., 2004}; ^{Huang et al., 2015}; ^{Pereira et al., 2019}; ^{Rincón-Riveros et al., 2020}; ^{Simoneit et al., 1999}; T. ^{Zhang et al., 2008}; ^{Zhang et al., 2007}). Since no local source profiles were available, we selected published source profiles for the model: natural gas combustion (^{Rogge et al., 1993}), diesel vehicles (^{Lough et al., 2007}; Lough and Schauer, 2007), gasoline vehicles (^{Lough et al., 2007}; ^{Lough and Schauer, 2007}), wood smoke (^{Fine et al., 2004}) and residential coal burning (^{Zhang et al., 2008}) assimilating small facilities. These sources have been used in Latin America by Villalobos et al. (²⁰¹⁷; ²⁰¹⁵) for source apportionment in Santiago and Temuco, Chile. Other source profiles such as meat cooking and vegetative detritus were included in preliminary runs of the model but were not found to produce acceptable results. Source contributions to OC from the CMB model were converted to PM_2.5 contributions using specific OC/PM_2.5 ratios for each source (^{Sheesley et al., 2007}; ^{Turpin and Lim, 2001}; ^{Zhang et al., 2008}). As it will be shown, sources that were not resolved by the CMB model were assumed to be associated with secondary organic compounds and converted using an OC/PM_2.5 ratio of 0.5 (^{Daher et al., 2012}; ^{Turpin and Lim, 2001}; ^{Utembe et al., 2009}; ^{Yin et al., 2010}), even though there may be other unresolved sources such as meat cooking and trash burning.

2.4. Backward trajectories

Potential pollutant sources were identified based on cluster and 2D kernel density analysis of backward trajectories (BWT) of air masses during the sampling periods (weekly from 15 September 2014 to 9 November 2014). The HYSPLIT Lagrangian back trajectory model (^{Stein et al., 2015}) was used to track the origin and location of air masses arriving at the sampling sites (Mosquera: 4º42’9.75” N, 74º13’54.94” W, and Soacha: 4º35’4.59” N, 74º13’11.62” W). The configuration used by ^{Mendez-Espinosa et al. (2019)} to calculate BWT was followed. BWTs were calculated eight times per day (starting at 1 UTC-5 <local time>, every 3 hours). Each trajectory was computed for 72 h to analyze short, middle, and long-range transport in the mesoscale (less than 600 km). The Global Data Assimilation System of 1º x 1º (GDAS1) was used as input gridded meteorological dataset since it is more suitable over complex topography and diversified land uses (^{Su et al., 2015}), using SplitR v0.4 R-package by (^{Iannone, 2016}). As the inaccuracy of BWT models could be directly proportional to shorter receptor heights given that local winds, turbulence processes, and frictional effects cannot be accurately represented (^{McGowan and Clark, 2008}; ^{Sapkota et al., 2005}; ^{Zielinski et al., 2016}), BWT were calculated for two receptor heights: 500m, and 1000m above ground level (AGL). Receptor heights were set up below and near mixing height (~1500m/year) (Mendez-Espinosa et al., 2019). BWT were linked with pollutant species, hereafter, they were grouped according to their angle. Angular distance clustering was computed using the OpenAir package (^{Carslaw and Ropkins, 2012}). The results are shown as concentration per chemical concentrations associated with clustering trajectories. Then, back-trajectories density maps were created by clusters to estimate the areas with the larger density of points related to the geo-location of air masses.

In the cluster analysis, a normality test (Shapiro-Wilk) was applied to the data. No dataset per cluster and pollutant was found to be normally distributed, and therefore, non-parametric Kruskal-Wallis tests (one-way ANOVA on ranks) were calculated to determine whether or not there was a significant difference (p-value <0.05) among clusters per observed pollutant.

3.1. PM _2.5 mass concentrations, bulk composition, and mass reconstruction

Daily PM_2.5 concentrations, shown in Figure S3, during the sampling period were 31 ± 15 µg m^-3 in Soacha (max. 78 µg m^-3) and 26 ± 12 µg m^-3 in Mosquera (max. 87 µg m^-3). The percentage of samples exceeding the current Colombian Air Quality Standard (37 µg m^-3) and the WHO guideline (25 µg m^-3) was 30%, and 60% respectively in Soacha, 16%, and 45% respectively in Mosquera. Figure 1 shows daily scatter plot comparisons of PM_2.5 and major constituents (3 samples per week were analyzed) between Soacha and Mosquera. For each chemical species, the Pearson correlation coefficient (r) and the Coefficient of Divergence (COD) between both sampling sites were calculated. The Pearson coefficient shows the degree of correspondence, so values close to 1 indicate that concentrations at both sites are proportional. The COD shows the heterogeneity, so values lower than 0.2 indicate that concentrations have a homogeneous spatial distributions between both sites and values close to 1 indicate that concentrations are heterogeneous (^{Xie et al., 2012}; ^{Yatkin et al., 2020}). Nitrate (r = 0.75, COD = 0.29), sulfate (r = 0.66, COD = 0.27) and calcium (r = 0.70, COD = 0.27) concentrations showed the highest r, suggesting that there is some regional influence of secondary inorganic aerosols and dust, although their spatial distribution was heterogeneous. EC (r = -0.02, COD = 0.25), OC (r² = -0.05, COD = 0.23) and PM_2.5 (r = 0.12, COD = 0.25), are poorly correlated and heterogeneous, which suggests that the main sources of PM_2.5 and its carbonaceous fractions are local at both sites.

Fig. 1 Daily scatter plot comparisons of PM_2.5 and major constituents between Soacha and Mosquera. 3 daily samples were analyzed per week. Different symbols represent different weeks. 

PM_2.5 mass reconstruction was performed for the three samples per week analyzed from each site, by adding EC, Organic Material (OM) and WSII. OM concentrations were calculated using an OM/OC factor of 1.6 for both sites, as recommended by ^{Turpin and Lim (2001)} for urban aerosols. Reconstructed mass was 104 ± 25% and 103 ± 26% in Soacha and Mosquera, respectively. OM constituted a large fraction of PM_2.5, with 68 ± 19% (22.3 ± 7.1 µg m^-3) and 63 ± 19% (17.8 ± 7.6 µg m^-3) in Soacha and Mosquera, respectively. EC represented 21 ± 5% and 25 ± 5% of PM_2.5, respectively. The most abundant ion was Sulfate (5.7 ± 2.0% and 5.4 ± 2.8%), followed by ammonium (3.0 ± 1.3% and 2.9 ± 1.4%) and nitrate (1.5 ± 0.7% and 1.8 ± 0.3%). Chloride (1.5 ± 0.6% and 1.1 ± 0.6%) and Calcium (1.3 ± 0.8% and 1.5 ± 0.8%) contributions to PM_2.5 are as high as nitrate. Figure S4 shows the PM_2.5 mass reconstruction for all the samples analyzed at both sites: 3 samples per week for 8 weeks. Figure 2a shows the weekly average PM_2.5 concentrations and bulk composition at Soacha. Figure 2b and Table SII show the incremental excess of PM_2.5 concentrations and bulk composition in Soacha compared to Mosquera. An excess of PM_2.5 was found in Soacha compared to Mosquera for most of the sampling period, excepting weeks 1 and 3. Similarly, an excess of the PM_2.5 carbonaceous fractions (EC, WIOM, and WSOM) was found in Soacha compared to Mosquera, excepting EC in weeks 1, 3, and 5, and WIOM in week 1. A mixed situation was found for ions.

Fig. 2 a) Weekly PM_2.5 mass concentrations and bulk composition in Soacha and b) Incremental excess of major constituents of PM_2.5 in Soacha compared to Mosquera. 

3.2. Molecular markers

Molecular markers used in the CMB model made up a small fraction of OM, but they are useful to identify and quantify the OM and PM_2.5 sources. Table SI shows the weekly average concentrations of organic fractions and molecular markers with their corresponding uncertainties, and variability metrics for each species (median, interquartile range and standard deviation). Even though the sampling period was short, the observed concentration variability was reasonably good for the application of the CMB model.

Figures 2 to 7 show the weekly concentrations of EC + OC and molecular markers in Soacha (part a) and the incremental excess of EC + OC and molecular markers in Soacha compared to Mosquera (part b). EC + OC showed similar temporal trends to PM_2.5 at both sites, with higher-than-average concentrations at Soacha during weeks 2 and, especially, during week 7 (Fig. 3a). There was a consistent incremental excess of EC at Soacha compared to Mosquera, with the only exception of week 1. The incremental excess of OC at Soacha compared to Mosquera (Fig. 3b) was negative in weeks 1, 3, and, to a lesser extent, 5, and positive for the rest of the weeks, with higher excess in weeks 7 and 8. OC/EC ratios were 1.91 ± 0.37 at Soacha and 1.60 ± 0.30 at Mosquera. These ratios can be considered as typical of urban aerosols with an influence of secondary aerosol formation (^{Blanchard et al., 2011}; ^{Khan et al., 2012}). They were lower than the ratios measured in Santiago (^{Villalobos et al., 2015}) and Temuco (^{Villalobos et al., 2017}) - with a high influence of biomass burning for heating in winter. They were similar to the ratio found at Xalostoc industrial site and lower than those at La Merced commercial site and Pedregal residential site in Mexico City (^{Vega et al., 2004}).

Fig. 3 a) Weekly EC and OC concentrations in Soacha and b) Incremental excess of EC and OC in Soacha compared to Mosquera. 

Individual concentrations of PAHs in Soacha (Fig. 4a) were 1.98 ± 0.77 ng m^-3 for benzo[b]fluoranthene, 1.69 ± 0.64 ng m^-3 for benzo[k]fluoranthene and 1.69 ± 0.64 ng m^-3 for benzo[e]pyrene). They were in a similar range to those measured in Temuco and lower than those in Santiago during winter (^{Villalobos et al., 2017}, ²⁰¹⁵), but 20 times higher than near an oil refinery in northeastern Mexico (^{Montaño-Soto et al., 2017}), 10 times higher than in Monterrey, Mexico (^{Longoria-Rodríguez et al., 2020}) and around 70%-80% of the levels found in Cuernavaca, México (^{Murillo-Tovar et al., 2018}; ^{Saldarriaga-Noreña et al., 2015}). Incremental excess of all these compounds in Soacha compared to Mosquera (Fig. 4b) were negative in weeks 1, 3, 5, and, particularly, 8. The negative excess was consistent for benzo(k)fluoranthene for the rest of the period. There was also a slight negative excess of benzo(e)pyrene in week 2 for benzo(e)pyrene. Overall, the incremental excess in Soacha tended to be negative for these PAHs (Fig. 4b).

Fig. 4 a) Weekly PAH concentrations in Soacha and b) Incremental excess of PAH in Soacha compared to Mosquera. 

Picene concentrations were around one half and 75% of those measured in Temuco (^{Villalobos et al., 2017}) at Soacha (Fig. 7a) and Mosquera (0.37 ± 0.09 ng m^-3 and 0.55 ± 0.13 ng m^-3, respectively), showing that coal burning contributes to PM_2.5 at both sites. Picene was not detected during week 3 at Mosquera. The incremental excess (Fig. 7b) was positive only for weeks 3, 4, and 7 (Fig. 7b). Overall, picene concentrations suggest that the impact of coal burning is similar at both sites.

Hopane concentrations were 1.27 ± 0.20 ng m^-3 at Soacha (Fig. 5a) and 2.81 ± 0.24 ng m^-3 at Mosquera, a similar range to that found in Santiago (^{Villalobos et al., 2015}) and Mexico City (^{Stone et al., 2008}), and higher than in Temuco (^{Villalobos et al., 2017}). All the hopanes included in this analysis had a negative incremental excess in Soacha, showing a lower impact of gasoline vehicles compared to Mosquera (Fig. 5b).

Fig. 5 a) Weekly hopanes concentrations in Soacha and b) Incremental excess of hopanes in Soacha compared to Mosquera. 

The sum of concentrations of n-Alkanes from C27 to C33 were 34.2 ± 9.46 ng m^-3 at Soacha (Fig. 6a) and 54.8 ± 5.86 ng m^-3 at Mosquera, 1.8 and 3-fold the concentration measured in Temuco (^{Villalobos et al., 2017}) and similar to those measured in Santiago before and after wintertime (^{Villalobos et al., 2015}). For most of the sampling weeks, the incremental excess of n-alkanes in Soacha was negative, except for octacosane in week 4; heptacosane, octacosane, and nonacosane in week 7 (Fig. 6b). The Carbon Preference Index - CPI applied to the C28 to C33 n-alkanes had values of 0.82 to 0.99 at Soacha and 1.04 to 1.42 at Mosquera, indicating that n-alkanes in PM_2.5 originate from anthropogenic sources.

Fig. 6 a) Weekly alkanes concentrations in Soacha and b) Incremental excess of alkanes in Soacha compared to Mosquera. 

Fig. 7 a) Weekly picene concentrations in Soacha and b) Incremental excess of picene in Soacha compared to Mosquera. 

Levoglucosan concentrations were 599.7 ± 72.35 ng m^-3 at Soacha (Fig. 8a) and 258.3 ± 28.5 ng m^-3 at Mosquera, similar to those measured in Santiago between April and May (^{Villalobos et al., 2015}), and 70% and 30%, respectively, of those measured in Temuco (^{Villalobos et al., 2017}), a city that is known for its high wood-burning contribution to PM_2.5 in winter. The incremental excess in Soacha was positive throughout the sampling period, indicating a much stronger impact of wood burning compared to Mosquera (Fig. 8b). The average Levoglucosan/PM_2.5 ratio was slightly lower than that in Temuco in winter and Santiago from June to September at Soacha (0.018) and one half of that value at Mosquera (0.009). Levoglucosan had a similar trend to PM_2.5 at Soacha, with a higher peak during week 2 and low concentration during week 5. At Mosquera, levoglucosan did not have strong variations and showed a fairly similar trend to PM_2.5.

Fig. 8 a) Weekly levoglucosan concentrations in Soacha and b) Mosquera. 

3.3. Chemical mass balance source apportionment to organic carbon

The sources that are more likely to contribute to organic carbon at the two sites are diesel trucks, gasoline vehicles, natural gas combustion, coal combustion in small facilities, and wood combustion in local bakeries. The ratio-ratio plots for Indeno(1,2,3-cd)pyrene x 1000/EC and benzo(g,h,i)perylene x 1000/EC in Figure S7, suggested by ^{Robinson et al. (2006)} as a useful tool to visualize the potential combination of sources, shows that all the ambient samples fell within the region formed by connecting the ratios of the source profiles selected for applying the CMB model. Ambient samples in Soacha (Fig. S7a) were close to the line connecting the ratios for wood combustion and diesel vehicles, whereas those in Mosquera (Fig. S7b) fell in the center of the polygon, showing that gasoline vehicles and coal combustion may have a higher contribution than in Soacha.

Considering the concentration/uncertainty ratios shown in Figure S6, and results from preliminary runs of the CMB model, the fitting species selected for the definite run were: EC, benzo(b)fluoranthene, benzo(k)fluoranthene, benzo(e)pyrene, benzo(g,h,i)perylene, 17A(H)-22,29,30-Trisnorhopane, 17A(H)-21B(H)-30-Norhopane, 17A(H)-21B(H)-Hopane, ABB-20R-C27-Cholestane, ABB-20R-C29-Sitostane + ABB-20S-C29-Sitostane, Levoglucosane and Picene.

CMB model results are shown in Table I, including the values for the fitting statistics r² and χ², and the percent of OC mass explained by the model. r² values above 0.97 and χ² values below 2.2 indicate a good fit between the selected sources and ambient measurements at both sites. The percent of OC mass explained by the model was within the accepted range (80% - 120%) for just half of the weekly samples at both sites and lower than 80% for the other half. The unexplained mass was considered to be due to “Other” source, which was calculated as the difference between the measured OC concentration and the sum of the calculated concentrations of the significant fitting sources. It is, therefore, a measure of unresolved sources. Wood combustion, diesel vehicles, and gasoline vehicles had significant contributions for all samples at both sites. Natural gas combustion was not significant during weeks 2 and 3 at both sites. Coal combustion at small facilities was not significant during weeks 1 and 3 at Soacha and was significant only during weeks 4, 7, and 8 at Mosquera. Other sources were not significant during weeks 2, 4, and 6 at Soacha, and during weeks 3 and 7 at Mosquera.

The main contributor to OC in Soacha (Fig. 9a) was wood combustion (6.02 ± 1.35 µg m^-3) followed by diesel vehicles (2.46 ± 0.25 µg m^-3), other sources (3.86 ± 1.23 µg m^-3), gasoline vehicles (2.15 ± 0.25 µg m^-3), coal combustion in small facilities (0.17 ± 0.05 µg m^-3) and natural gas combustion (0.06 ± 0.02 µg m^-3). In Mosquera (Fig. 9b), the main contributors were gasoline vehicles (4.92 ± 0.55 µg m^-3), followed by other sources (2.42 ± 0.56 µg m^-3), diesel vehicles (1.87 ± 0.21 µg m^-3), wood combustion (1.71 ± 0.46 µg m^-3), coal combustion in small facilities (0.26 ± 0.09 µg m^-3) and natural gas combustion (0.11 ± 0.03 µg m^-3).

Fig. 9 a) Weekly source contribution to OC in Soacha and b) Mosquera. 

The use of chemical source profiles from the USA in Colombia is indeed a source of uncertainty in determining source contribution by using the CMB model, especially regarding vehicle technologies and fuels. The vehicle emission profiles used here were developed from a wide range of technologies of the Californian fleet (1975-2001), in which the Colombian fleet can be considered to be included. On the other hand, Colombian fuels meet similar standards to those in place in California over that period. Therefore, the true emission profiles can be thought to be inside the range of the profiles documented by ^{Lough and Schauer (2007)}.

3.4. Source apportionment to PM _2.5

Figure 10 shows the results of the source contribution to PM_2.5, after applying specific OC to PM_2.5 factors to the source contributions to OC and adding the concentrations of WSII. The main contributors at Soacha (Fig. 10a) were wood combustion (7.20 ± 2.59 µg m^-3), diesel vehicles (6.05 ± 2.51 µg m^-3), secondary organic aerosols (5.99 ± 3.21 µg m^-3), coal combustion in small facilities (3.49 ± 1.87 µg m^-3) and gasoline vehicles (2.77 ± 0.68 µg m^-3). At Mosquera (Fig. 10b), the main contributors were gasoline vehicles (6.33 ± 1.52 µg m^-3), diesel vehicles (4.58 ± 1.07 µg m^-3), secondary organic aerosols (3.73 ± 2.49 µg m^-3), coal combustion at small facilities (3.65 ± 2.91 µg m^-3) and wood combustion (2.04 ± 0.32 µg m^-3). The average undetermined mass was 8% of PM_2.5 at Soacha (max 27%) and 11% of PM_2.5 at Mosquera (max 30%). This fraction may be associated with dust and other industrial and commercial sources. Previous source apportionment studies in Bogotá had been made for PM₁₀ (^{Ramírez et al., 2018a}; ^{Ramírez et al., 2018b}; ^{Vargas et al., 2012}), not for PM_2.5, and could not resolve the contributions of different types of vehicles.

Fig. 10 a) Weekly source contribution to PM_2.5 from molecular markers in Soacha and b) Incremental excess of source contributions to PM_2.5 in Soacha compared to Mosquera. 

The scatter plot comparisons of source contributions between Soacha and Mosquera (Fig. 11) shows relatively high correlations for ammonium (r² = 0.64) and sulfate (r² = 0.72), associated with the regional contribution of secondary inorganic aerosols during the sampling period. The correlation was also high for the contribution of diesel exhaust, which can be attributed to similar behavior of the cargo fleet traffic at both sites. Natural gas contributions were well correlated between both sites most of the time, but a higher contribution in Mosquera in week 8 appears as an outlier. Other sources are not well correlated, indicating that they only have local influence. The high variability in the contribution of wood combustion in Soacha contrasts with the fairly constant contribution in Mosquera. The opposite is shown for the contribution of gasoline, more variable in Mosquera, and less variable in Soacha.

Fig. 11 Weekly scatter plot comparisons of source contributions between Soacha and Mosquera. 

In summary, even though chemical source profiles were not developed in Colombia, these source apportionment results are reasonable when considering the characteristics and activities found around the sites and suggest that future emission inventories exercises should not underestimate the importance of wood and coal burning in small facilities, often neglected.

3.5. Backward trajectories

Back-trajectory kernel densities for both sites for the whole sampling period (Fig. 12) show three clusters. C1 corresponds to air masses arriving from the Orinoco basin, passing over Bogotá, with 64.6 % of the trajectories. C2 (24.7 %) has mixed trajectories from the Orinoco basin and influence of the Amazon rainforest in the South, arriving at the sites from the East, having passed over Bogotá. C3 (10.7 %) arrives from the Northwest, with air masses from the Magdalena Valley, between the Central and Eastern Colombian Andes, and the influence of the Pacific Coast.

Fig. 12 Kernel densities of backward trajectories arriving at Soacha and Mosquera. Percentages correspond to the contribution of C1, C2, and C3 back trajectory clusters at each site. 

For a more detailed view, weekly kernel densities are shown in Figure S8. Three clusters per week are plotted, and the percentage of trajectories represented by each cluster per week is indicated. The great majority of the air mass trajectories arrive at both sites from the eastern savanna, which is part of the Orinoco basin, having passed over Bogotá. In several weeks, two of the three clusters, or even the 3 clusters found do not differ significantly from each other. In week 7, clusters C1 and C2 arriving at Soacha can be associated with C1 arriving at Mosquera, while cluster C3 arriving at Soacha can be associated with clusters C2 and C3 arriving at Mosquera.

Higher wind speeds, and therefore longer trajectories, were observed during week 1 and, to a lesser extent, during week 8, for trajectories arriving from the East. Trajectories in cluster C3 in week 4 (21.4 % and 23.8 %), and week 6 (14.3 %) arrive from the North, at Soacha and Mosquera, respectively. Week 3 showed a slight influence of southeastern trajectories. Week 5 was the only week with a persistent influence of air masses arriving from the South/Southeast (69 % and 71.4 % of trajectories in cluster C2 at Soacha and Mosquera, respectively). It was during this week that a higher contribution of SOA was found at both sites; this may be due to the influence of emissions from rainforest vegetation.

In contrast, air masses from the Northwest (Magdalena Valley) were observed only in week 7. This was the week with the highest PM_2.5 concentrations and the highest contribution of several sources, such as diesel combustion, wood combustion, and secondary inorganic aerosol compounds (sulfate and ammonium). Emission sources located along trajectories from the Northwest are few and weak when compared to those from the East (i.e. the city of Bogotá). This suggests that there was also poorer dispersion of local emissions associated with these Northwest trajectories and a contribution of transport of secondary inorganic aerosols from the Magdalena Valley.

The percentage of clusters per week was very similar for both sites, suggesting that both are influenced by similar air masses. However, the influence of Bogotá’s emissions or long-range transport of pollutants does not seem to have been dominant over the concentrations and compositions of PM_2.5 of the sites during the sampling period. Therefore, the incremental excess in the concentrations of PM_2.5 and its constituents, as well as in the source contribution to organic carbon and PM_2.5, is mostly a consequence of local emissions.