Air pollution in Kigali, Rwanda: spatial and temporal variability, source contributions, and the impact of car-free Sundays

Introduction Ambient air pollution, especially fine particulate mass (PM2.5) and ozone (O3), has been associated with premature human mortality (Dockery et al., 1993; Jerrett et al., 2009; Laden et al., 2006). The World Health Organization (WHO) estimates that in 2016, ambient air pollution caused about three thousand deaths in Rwanda (Brauer et al., 2012; WHO, 2018). However, such estimates can be uncertain because exposure is inferred from satellite estimates. There has been no long-term groundbased monitoring in major cities like Kigali to validate estimated Abstract Ambient air pollution, particularly fine particulate mass (PM2.5) and ozone (O3), is associated with premature human mortality and other health effects, but monitoring is scarce to non-existent in large parts of Africa. Lower-cost real-time affordable multi-pollutant (RAMP) monitors and a black carbon (BC) monitor were deployed in Kigali, Rwanda to fill the air quality data gap here. PM2.5 data were corrected using data from a coincident, short-term campaign that used standard filter-based gravimetry, while gas data were verified by collocation with reference carbon monoxide (CO) and O3 monitors at the Rwanda Climate Observatory at Mt Mugogo, Rwanda. Over March 2017-July 2018, the ambient average PM2.5 in Kigali was 52 μg/m 3, significantly higher than World Health Organization (WHO) Interim Target 1. Study average BC was 4 μg/m3, comparable to mid-sized urban areas in India and China and significantly higher than BC in cities in developed countries. Spatial variability across various urban background sites in Kigali appears to be limited, while PM2.5 at Mt Mugogo is moderately correlated with PM2.5 in Kigali. A sharp diurnal profile is observed in both PM2.5 and BC, with the Absorption Angstrom Exponent (AAE) indicating that the morning peak is associated with rush-hour traffic-related air pollution (TRAP) while the late evening peak can be attributed to both traffic and domestic biofuel use. PM2.5 in the dry seasons is about two times PM2.5 during the following wet seasons while BC is 40-60% higher. Local sources contribute at least half the ambient PM2.5 during wet seasons and one-fourth during dry seasons. Traffic restrictions on some Sundays appear to reduce PM2.5 and BC by 10-12 μg/m 3 and 1 μg/m3 respectively, but this needs further investigation. Dry season ozone in Kigali can exceed WHO guidelines. These lowercost monitors can play an important role in the continued monitoring essential to track the effectiveness of pollution-control policies recently implemented in Rwanda.

exposures. The situation is similar in many other sub-Saharan African countries (Kalisa et al., 2019;Petkova et al., 2013). This lack of monitoring due to resource limitations also hampers scientific understanding of the sources contributing to air pollution in these countries, which is essential to formulating effective environmental management policies.
In Rwanda, biomass use accounts for 85% of energy consumption, in the form of wood and charcoal (MININFRA, 2018). Wood is used in rural households and charcoal in urban households. Petroleum in the form of transportation fuel, liquified petroleum gas (LPG) for household cooking, and electricity generation accounts for 13% of national energy consumption (MININFRA, 2018). As of 2018, 77% of on-road vehicles (excluding motorcycles) were manufactured before 2005 (Duhuze, 2018). As a result, air pollution in Kigali can be significantly higher at roadside locations than at urban background locations . Rwandan air quality is also influenced by regional forest fires and seasonal weather patterns (DeWitt et al., 2019). Rwanda's electricity generation (218 MW) is composed of 45% hydropower, diesel and heavy fuel oil at 27%, methane (14%), peat (7%), and solar (6%). Diesel is used to fuel peaking power stations and for backup generation during power outages, which can be significant contributors to ambient air pollution (Farquharson et al., 2018;. Some of these sources were identified by Henninger (2013) using scanning electron microscopy of filter samples.
To more accurately quantify source contributions (a requirement for effective air quality management), source apportionment is often conducted with chemical mass balance (CMB) or receptor modeling (e.g. positive matrix factorization) using organic molecular markers (Shrivastava et al., 2007;Subramanian et al., 2007) or aerosol mass spectrometry (AMS) (Zhang et al., 2011). However, such studies require extensive sample collection and offline analysis or long-term deployment of expensive equipment, especially if sources can vary between seasons. Pikridas et al. (2013) find that an observation-based method (OBM), which uses the temporal pattern of pollution measured with even a low-cost PM monitor, closely replicates the regional/ urban divide based on PM 2.5 composition measurements inside and upwind of Patras, Greece. Diamantopoulou et al. (2016) use regional air quality modeling to simulate observations and find the OBM analysis of the pseudo-observations comparable to the model-computed regional/local divide.
Traffic restrictions such as car-free days, low-emission zones (LEZ), and "odd/even" policies have been used in cities worldwide to reduce air pollution. The evidence that such policies reduce human exposure is mixed. No effect was observed on air quality in Mestre-Venice (Masiol et al., 2014). London's LEZ saw decreases in nitrogen dioxide (NO 2 ), but no significant changes in PM 2.5 or PM 10 (Mudway et al., 2019). Extensive traffic control measures in Beijing during the 2008 Olympics reduced median black carbon (BC) concentrations by as much as 50% (Wang et al., 2009). The Rwandan government, to promote social welfare, has recently implemented "car-free" Sundays, wherein major roads are blocked off and people take part in group exercises on car-free streets starting at 7 AM and ending either at 10 AM or at noon. In 2017, the first Sunday of each month was designated "car-free", while in 2018 that was expanded to the first and third Sundays of each month. Additionally, to curb the use of older imported cars and harmonize duty structures with other East African countries, in 2017 the Rwandan government increased duties on cars, with the increase depending on the vehicle age (RRA, 2017). The vehicle import rate dropped by 20% in the first half of 2017, likely connected to the higher import duties (Ngabonziza, 2017). Monitoring is required to quantify the impact of these policies on air quality in Kigali.
Here, we show how low-cost sensors can improve scientific understanding of air quality in resource-challenged countries. The high time resolution of the RAMPs and BC monitor enables an examination of diurnal patterns in each season, which is not possible with integrated daily filter samples. We apply OBM to the RAMP and BC datasets to get a preliminary estimate of the regional and local contribution to Kigali air pollution in each season. Measurements at multiple "urban background" sites within Kigali allow us to examine intra-urban variability in air pollution at non-roadside locations across Kigali, unlike the urban background/roadside comparison by Kalisa et al. (2018). Long-term monitoring with RAMP monitors allows examination of the seasonal variability in ambient air pollution across multiple dry and wet seasons. Furthermore, we evaluate the impact of the "car-free Sunday" policy on air pollution. The multiwavelength aerosol light absorption from the BC monitor helps us qualitatively identify contributions from biomass burning and fossil fuel combustion. Comparison of urban and rural air pollution provides insight into the impact of urbanization on air quality. We end with recommendations on ways to implement and improve such studies in Rwanda and other countries in the Global South.  Table 1. Kigali terrain is composed of hills and valleys; the RAMPs in this study were all located at hilltop sites. The Gacuriro and Belle Vue Estate locations are both residential neighborhoods about 0.5 km apart. The Carnegie Mellon University (CMU-Africa) site is about 2.5 km from Gacuriro, located in a commercial building complex near a major road and six stories above ground (this was the campus till late 2019). RAMP #152 and a Met-One 10-wavelength BC monitor (BC-1054) were deployed in July 2017 on the roof of the University of Rwanda's five-story College of Science & Technology building (UR-CST), where Kalisa et al. (2018) had collected filter-based samples for their urban background location over April-June 2017. This manuscript focuses on RAMP data collected at UR-CST between July 2017-July 2018. The BC data ends in April 2018 as the pump malfunctioned, which made the instrument inoperable. The Gacuriro and Belle Vue sites are about 5 km from the UR-CST site.

The RAMPs and gas sensor calibration
The RAMPs ( Figure S1) were manufactured by Sensevere (now owned by Sensit Technologies, Valparaiso, IN, USA) and cost about US$ 3,000 each at the time of purchase (base unit without an external PM sensor). The RAMP monitors and calibration methodologies are described in previous work (Malings et al., 2019(Malings et al., , 2020Subramanian et al., 2018;Zimmerman et al., 2018). Briefly, the RAMP uses passive Alphasense (UK) electrochemical sensors to measure CO, nitrogen dioxide (NO 2 ), O 3 , and other gases. The raw signals of the RAMP electrochemical gas sensors (collected at 4 times per minute) are processed and averaged to provide hourly ambient concentrations using generalized RAMP (gRAMP) calibration models (Malings et al., 2019) developed in Pittsburgh, Pennsylvania, USA. The gRAMP calibration models are based on data from several RAMP monitors collocated with reference gas monitors at the CMU campus in Pittsburgh in 2017 and were shown to transfer better to other locations in Pittsburgh than calibration models developed for individual RAMPs. For CO, a quadratic regression (QR) gRAMP model is used and for O 3 , a hybrid random forest/linear regression ("hybrid-RF") gRAMP model is used.
For local verification of CO and O 3 , RAMP #140 was collocated with reference monitors for CO and CO 2 (Picarro G2401) and O 3 (Teledyne T400) at the Rwanda Climate Observatory (RCO) on the summit of Mt Mugogo (about 70 km from Kigali and 2590 m above sea level, DeWitt et al. (2019)) over July-December 2017. Due to instrument malfunctions, collocated measurements are available for only 60 days of this six-month period. "Mugogo" linear regression, QR, and hybrid-RF models are developed using a subset (four weeks) of the collocation data; as shown in Zimmerman et al. (2018), a four week period is sufficient to develop calibration models for these electrochemical sensors. The remaining collocation data (32 days) are set aside to provide an independent or unseen data set for testing model performance, a practice previously established by our group in Zimmerman et al. (2018) and Malings et al. (2019). The performance of the Mugogo and the Pittsburgh gRAMP models is shown in Table 2. For CO, the gRAMP models showed slightly worse correlation (Pearson r) and normalized mean absolute error (CvMAE) than the Mugogo-based calibration models. For O 3 , the gRAMP model is comparable to the better-performing Mugogo QR and hybrid-RF models for r and CvMAE. RAMPspecific Mugogo models show lower bias than the gRAMP models. During the dry season, most of the data from the gRAMP ozone model are within ±30% of the reference monitor data at RCO ( Figure S2).
RAMP-specific models trained on one RAMP may not transfer as well to other RAMPs as the gRAMP models (Malings et al., 2019). As most of the Kigali O 3 data is from RAMPs not collocated at Mugogo, the results presented here for ambient CO and O 3 are based on the gRAMP models. As we do not have local verification for NO, NO 2 , or SO 2 , these data are not presented here. PM 2.5 measurements PM 2.5 is measured using a Met-One neighborhood PM monitor (NPM) paired with each RAMP. The NPM is a nephelometer with a PM 2.5 cyclone, an inlet heater to reduce humidity effects, and a pump (flow rate 2 lpm). The PM 2.5 data are processed using methods developed based on collocations of over two dozen NPMs with US Environmental Protection Agency (EPA) federal equivalent method (FEM) beta attenuation monitors (BAM) at an urban background location and a source-impacted location in Pittsburgh (Malings et al., 2020). Either a physical (assumed composition and hygroscopic growth with scaling to BAM values) or an empirical approach (a quadratic regression with raw data, temperature, and RH as the variables) is used to convert as-reported NPM readings to "BAM-equivalent" PM 2.5 mass concentration (reported at 35% RH and 22 °C); both methods performed similarly in Pittsburgh. For Rwanda, we use the Pittsburgh-based physical approach, then apply a further localized correction to account for differences in aerosol size distribution and composition between Kigali and Pittsburgh. This localized correction of the "BAM-equivalent" PM 2.5 data is based on a comparison of the Gacuriro RAMP with 24-hour integrated filter-based measurements in April 2017 collected by Kalisa et al. (2018) at UR-CST about 5 km away (described in the SI). Briefly, the "BAM-equivalent" PM 2.5 data (corrected for hygroscopic growth but not for aerosol differences between Kigali and Pittsburgh) are strongly correlated (correlation coefficient, r 2 = 0.77) with the filter-based PM 2.5 for non-working days (weekends, holidays, and car-free Sundays). However, the comparison is more scattered for working days (weekdays that are not holidays), which might indicate some intraurban variability related to local activities such as traffic and industrial emissions that may be more prominent on working days. Overall, the working day "BAM-equivalent" RAMP PM 2.5 values are scaled up by 1.69 and the non-workday "BAM-equivalent" RAMP PM 2.5 data are scaled up by 1.39 ( Figure S3).

BC mass and AAE measurements
The BC-1054 monitor (Met One Instruments, Inc.) deployed at the UR-CST site measures light attenuation by a filter sample at ten wavelengths between 370-950 nm. The as-reported BC mass concentrations from this monitor were processed using manufacturer-provided software (BC Load Correction 1.3.1), which corrects for known filter-loading artifacts (Kirchstetter and Novakov, 2007) using the algorithm developed by Virkkula et al. (2007). However, no correction is made for potential light absorption enhancement due to BC mixing state (Bond et al., 2006), which means the BC mass concentrations reported here may be overestimates. The Absorption Angstrom Exponent (AAE) was calculated based on all ten wavelengths as the negative slope of the relationship between the wavelength-dependent attenuation and the wavelength in a log-log space (Moosmüller et al., 2011). For fresh combustion aerosol mainly composed of black carbon and for particles small relative to the wavelength, AAE is expected to be near unity. A compilation of studies (Lack and Langridge, 2013) on fossil fuel emissions and urban pollution with BC as the dominant absorber yielded an average AAE of 1.1±0.3 (± one standard deviation). The AAE for lightabsorbing organic compounds (brown carbon or BrC) associated with biomass burning or humic-like substances is higher, with estimates ranging from 2 to 6 (Kirchstetter et al., 2004;Sun et al., 2007). Hence, higher values of AAE can be used to qualitatively identify periods when non-BC aerosol components such as light-absorbing dust or biomass burning containing BrC are significant contributors to aerosol light absorption. We do not attempt to quantify the contributions of fossil fuel and biomass burning emissions here as there can be significant uncertainty associated with such methods, as described in DeWitt et al.

Intra-urban variability across Kigali
The UR-CST and Belle Vue Estate sites are about 5 km apart and have the largest paired data set across both dry (July-September) and wet (October-November) seasons. An orthogonal distance regression (ODR) fit of the paired hourly average PM 2.5 concentrations yields a slope 0.996±0.012 and effectively zerointercept. An ordinary least-squares fit (not forced through zero) has a correlation (r 2 ) of 0.61. The UR-CST and Belle Vue Estate sites in Kigali are typical of urban background hilltop locations (Figure 1), unlike the urban background/roadside comparison where Kalisa et al. (2018) found significant differences. The much smaller dataset of 337 paired hourly average PM 2.5 measurements at the Gacuriro and CMU-Africa sites yields a slope of 0.967±0.03 (and an effectively zero intercept) with the ODR fit. For the sake of simplicity (as the succeeding analysis focuses on longer-term comparisons), the measurements at all four sites are averaged into a single time series providing hourly average PM 2.5 values from March 2017 to July 2018.  Figure S4). An ODR fit of the 1-hour O 3 values measured at Gacuriro and CMU-Africa yields a slope of 0.97±0.01 with negligible intercept. A comparison of the UR-CST and Belle Vue estate O 3 measurements is complicated by the significantly higher O 3 apparently measured at the UR-CST site during the mid-July to mid-September period. The reason for these differences is unclear, as temperature and relative humidity (two variables likely to affect sensor performance) measured at the two urban background locations are identical. This issue of potentially high O 3 at UR-CST needs to be investigated further; here, we take a conservative approach and present the data set as a unified O 3 time-series (as we do for PM 2.5 ), except the UR-CST measurements are excluded from the current analysis. This restricts our analysis to the March-November 2017 period (as there is only a week of data for December 2017 from the Belle Vue estate site), which are the two 2017 wet seasons (March-April-May or MAM and ON) and the long dry season (June-July-August-September, JJAS).
Unlike O 3 and PM 2.5 , differences in CO concentrations were sometimes observed at the residential locations and the campus sites in Kigali ( Figure S5 of the SI). The highest concentrations were observed usually at the campus sites, which might reflect the greater traffic seen by these sites compared to residential neighborhoods. However, the Gacuriro neighborhood did experience two hours when CO exceeded 3 ppm. Overall, the study average CO concentrations were similar at UR-CST and Belle Vue Estate at 0.446±0.322 ppm and 0.447±0.283 ppm (average and standard deviation) respectively. CO was 0.504±0.499 ppm at CMU-Africa and 0.404±0.320 ppm at Gacuriro. Concentrations at rural Mugogo were even lower, at 0.225±0.097 ppm. However, the measured CO concentrations at all locations were low and far below US EPA standards (35 ppm for 1 hour). The WHO has not set guideline values for CO. The higher standard deviations at the urban sites compared to the rural site suggests that the urban areas are more likely to see large spikes of CO.
Average monthly patterns of PM 2.5 and changes from 2017 to 2018 The unified Kigali time series for ambient PM 2.5 is summarized as monthly average PM 2.5 values, covering the period between March 2017 and July 2018 ( Figure 2). The least data were collected in June-July 2018 (100 and 64 hours) due to maintenance issues. For all other months, between 298-744 hours of data are available, averaging 634 hourly values per month (88% of a 30day period). The monthly average PM 2.5 values range from 25 µg/ m 3 in May 2018 to 102 µg/m 3 in July 2017, all higher than the WHO annual guideline of 10 µg/m 3 ; most months also exceed the WHO's first interim target (WHO IT 1) of 35 µg/m 3 . The study average hourly PM 2.5 in Kigali is 52.4±33.7 µg/m 3 ; the large standard deviation suggests significant temporal variability that will be explored further in later sections.
A comparison of the overlapping months shows a reduction of almost 40% from 2017 to 2018 for March, April, and May, and 51% for July, while the June monthly average is practically unchanged. However, June and July 2018 had the least data coverage (as noted in the preceding paragraph), and fewer than 300 hours of data are available for  ) Note that August and September are both considered "dry" here. As Figure 2 showed, the monthly average PM 2.5 value for September is closer to that of August and June, and noticeably higher than the average PM 2.5 for October and November. Additionally, September 2017 saw just 17 mm of rainfall, compared to 104 mm of rainfall in October-November 2017 ( Figure S6). This suggests that our classification is appropriate for 2017. Figure 3 shows that ambient PM 2.5 levels are higher in the dry seasons than in the wet seasons. The hourly data collected with the RAMPs on non-working days (nWD)

BC and AAE in Kigali
A total of 6,850 hours of BC data were collected at UR-CST between July 6, 2017-April 24, 2018. The equivalent BC mass concentration (as measured at 880 nm) hourly averages ranged from 0.14 µg/m 3 to 49.6 µg/m 3 , with a study average of 4.04±2.86 µg/m 3 . This value is significantly higher than the values observed in urban areas of developed countries; e.g. in the early 2000s, urban ambient BC was 2 and 1 µg/m 3 respectively in the states of New Jersey and California in the USA after decades of reductions (Kirchstetter et al., 2017). The Kigali BC values are comparable to mid-sized urban areas in China and India. In Hefei (central China), annual average BC in 2012-2013 was 3.5 µg/m 3 (Zhang et al., 2015). In Pune (western India), the average BC over 2015-2016 was 3.9 µg/m 3 (Kolhe et al., 2018). BC working day seasonal averages are summarized in Table 3. Non-working day seasonal averages were 0.2-0.9 µg/m 3 lower. The difference between the wet and dry seasons (dry season BC 40-60% higher) in Kigali is lower than the factor-of-four difference observed in similarly-tropical Pune (Kolhe et al., 2018)  The study average AAE was 1.53, with most values between 0.9-2.1 ( Figure S7). As shown in Table 3, AAE was slightly lower in  (Figure S7), while higher BC concentrations showed AAE values around 1.4 or less, which suggests that fossil fuel combustion is associated with higher BC levels, but the background BC is dominated by biomass burning.
Diurnal profiles of PM 2.5 and BC Figure 4 shows the average diurnal pattern of PM 2.5 and BC for the five seasons (four for BC) during the study period for which we have a large number of measurements. Only working day data are shown for visual clarity; a similar pattern is observed for the non-working day measurements. As noted earlier, PM 2.5 concentrations are substantially higher in the dry seasons than in the wet seasons. A morning peak (likely related to traffic) is observed between 8 AM-10 AM local time for all seasons except MAM 2017. Concentrations then fall during the day (as the boundary layer height increases) before rising back up in the evening, likely a combination of evening emissions and lower boundary layer heights. The MAM 2017 night-time highs (and to some extent the JJAS night-time highs) remain at that level until the morning traffic peak. For the other three seasons, the morning traffic peak is higher than the night-time high values. While the BC concentrations follow a similar diurnal profile, in all seasons the maxima occur during the morning rush hours, with concentrations lowest in the afternoon. In further contrast to the PM 2.5 dry/wet seasonal differences, the wet and dry season BC concentrations are much closer to each other, though dry season BC is still higher.
The highest PM 2.5 concentrations are observed at night-time during MAM 2017 and JJAS, with average PM 2.5 around 80 µg/ m 3 between 8 PM-12 AM in the wet season (MAM 2017) and around 100 µg/m 3 between 7 PM-11 PM in the dry season (JJAS). In contrast, the highest average PM 2.5 levels in other seasons were observed during the morning rush hour -around 60 µg/m 3 between 7 AM-10 AM in ON, around 90 µg/m 3 between 8 AM-10 AM in DJF, and around 45 µg/m 3 between 8 AM-10 AM in MAM 2018. The BC maxima all occur during the morning rush hour between 7 AM-8 AM except for DJF, when levels are marginally higher between 8 AM-9 AM than during the preceding hour.
The morning and evening peaks at similar times for PM 2.5 and BC indicate that these PM peaks are related to combustion emissions. Figure 5 shows the diurnal variation in AAE during the dry and wet seasons. In both cases, AAE is lower between 6 AM to 9 AM, when BC mass concentrations are higher; the lowest median AAE is 1.4 between 7 AM-8 AM during the dry season. However, the night-time peaks in PM 2.5 and BC are associated with higher AAE values, when hourly medians approach 1.6-1.7. These differences suggest that while the morning BC and PM 2.5 peaks are mostly associated with fossil-fuel vehicular emissions, the night-time BC and PM 2.5 peaks are additionally Local and regional contributions to PM 2.5 and BC in Kigali We use OBM (Diamantopoulou et al., 2016;Pikridas et al., 2013) to estimate the regional and local contributions to ambient PM 2.5 in Kigali. During the dry season, a higher regional background is likely, as transported pollution may not be rained out. The background can include regional biomass burning, dust, and other upwind emissions as well as secondary PM 2.5 . The OBM assumes that the minimum value of a seasonal diurnal profile based on hourly measurements is the seasonal average regional contribution. Then, the seasonal average local contribution is the difference between the seasonal average ambient PM 2.5 and the seasonal average regional contribution. These estimates of regional contribution assume the ambient concentrations at the minima are entirely regional, but there may be local vehicular and domestic biofuel emissions all day. Thus, the results are the upper bound of the regional contribution and the lower bound for the local contribution.   Table 3 summarizes the seasonal statistics required to calculate the average regional contribution to Kigali ambient PM 2.5 for each season. The minimum average PM 2.5 values were observed between 3-4 PM local time in all seasons. About half the ambient PM 2.5 in Kigali appears to be from regional sources in the wet season, and the regional share rises to over 70% in the dry seasons. Put another way, local sources contribute at least half the observed PM 2.5 during the 2017 wet seasons, and one-fourth of ambient PM 2.5 during the two dry seasons. The MAM 2018 wet season appears to be an anomaly compared to the two other wet seasons, but in terms of absolute PM 2.5 concentrations, the regional component is about the same in MAM 2018 as during the 2017 wet seasons.
Following a similar procedure for BC, the regional contribution on working days is 54-55% in the JJAS and DJF dry seasons, but only 33% and 40% in the wet MAM 2018 and ON seasons. The regional contribution to Kigali BC on non-working days is 7-8% higher during JJAS, ON, and MAM 2018, which likely reflects lower local emissions (including car-free Sundays). Overall, local sources are larger contributors to ambient BC than to PM 2.5 in Kigali, ranging from about half in the dry season to two-thirds in the wet seasons.
The impact of Sunday car-free hours on urban background air pollution An estimate of the impact of traffic-related air pollution (TRAP) at non-roadside, urban background locations can be obtained by comparing the RAMP and BC measurements on "car-free" Sundays with the measurements on all other Sundays. In this study period, PM 2.5 (BC) measurements were made on 16 (11) "car-free" Sundays that can be compared to data from 42 (30) regular Sundays. However, in Rwanda, the car-free policy is only in effect on Sunday mornings. The time-resolved data from the RAMPs and BC monitor allows investigation of the benefits of this policy for the specific periods when such policies are in effect.  Figure 6 shows a comparison of the hourly average PM 2.5 between regular Sundays and car-free Sundays. The average hourly PM 2.5 concentration is lower in the morning period by an average of 7 µg/m 3 . In the afternoon and evening periods, the diurnal patterns converge, indicating no significant difference in pollution. The reduction in ambient PM 2.5 is highest between 7 AM-9 AM, when PM 2.5 on regular Sundays is highest; during these hours the average PM 2.5 is reduced by 10-12 ug/m 3 on car-free Sundays or 20% of PM 2.5 at the same time on regular Sundays. BC concentrations are lower by about 1 µg/m 3 between 7 AM-10 AM on car-free Sundays, which is 18%-28% of the corresponding concentrations on regular Sundays. However, there are also similar differences for pre-7 AM "business-as-usual" times on these Sundays, and so the observed reductions during the morning car-free periods may not be directly attributable to the car-free policy.

Ozone in Kigali and Mugogo
Only 942

Conclusions and recommendations for future work
We have presented the results of a long-term ground-based monitoring campaign, the first of its kind in Kigali, Rwanda. Lower-cost and relatively low maintenance RAMP monitors were used for this study, with local verification and correction of sensor calibrations by collocation with reference monitoring during an overlapping campaign and at RCO.
O 3 pollution in Kigali was usually below WHO guidelines, but the 50-ppb threshold could be exceeded in the dry season. Periods of high ozone can be identified using low-cost sensors and a general calibration, though local calibrations improve sensor performance. The RAMP PM 2.5 (before filter-based correction) correlates strongly with filter-based PM 2.5 on non-working days when TRAP and other working-day contributions are lower, but the RAMP PM 2.5 values were still a significant underestimate. These differences suggest that the size distribution of PM 2.5 in Kigali is quite different from that in Pittsburgh, with substantial contributions from sub-300 nm particles (where low-cost optical sensors are less sensitive). Future studies with lower-cost monitors should include collocated filter-based measurements or short-term intensive studies with aerosol sizing instruments (e.g. a scanning mobility particle sizer, SMPS) to account for such differences.
The Health Effects Institute (2019) (HEI) using data from the Global Burden of Disease Study 2017 estimates that the population-weighted annual average PM 2.5 in Rwanda was 43 µg/m 3 in 2017 (and the same for 2015-2016). Our study found that the study average PM 2.5 in Kigali is 52±34 µg/m 3 at residential or university locations, which can be considered as "urban background" sites. Kalisa et al. (2018) show that pollution levels can be significantly higher at the roadside in Kigali than at the urban background locations where our measurements were made. Air pollution in the low-lying valleys of Kigali can also be higher (Henninger, 2013). RAMP measurements at Musanze in late 2017 showed PM 2.5 concentrations at this rural site were moderately correlated (r 2 = 0.54) with Kigali PM 2.5 and about 20% lower (not shown). These results suggest that the HEI population-weighted average for Rwanda and other similar satellite-based estimates could be underestimates that need to be updated with ground-based monitoring, like the campaign presented here.
Three recent studies -this paper, DeWitt et al. (2019), and Kalisa et al. (2018) -have found that air pollution is significant in urban and rural Rwanda, with considerable spatial variability due to local conditions and sources. Hence, future studies with lowcost monitors should include a variety of locations, such as low-lying areas and roadside locations. Our results should be considered a lower estimate of the PM 2.5 pollution that Kigali residents are exposed to -which is concerning since the average monthly concentrations in our 16+ month study were above the WHO annual guideline and often over WHO's first interim target.
Long-term monitoring over five seasons shows that differences in PM 2.5 between working days and non-working days are smaller than the differences between the dry and wet seasons. The average dry season PM 2.5 levels are about two times the succeeding wet season PM 2.5 while the comparable seasonal differences in BC are about 40-60%, indicating that transported non-BC (e.g. dust) or low-BC (e.g. forest fires) pollution is important in the dry season. Particulate pollution in Kigali has a distinct diurnal profile, as the morning rush-hour usually results in the maximum BC and PM 2.5 due to traffic pollution, and minima occur in the afternoons. The higher AAE values associated with the overnight BC and PM 2.5 peaks suggest that domestic biofuel use could also be a significant contributor to overnight air pollution.
Overall, local sources could contribute half the PM 2.5 and twothirds of the BC in the wet seasons in Kigali, which means that local pollution control policies can significantly improve Kigali's air quality. In MAM 2018, heavier local rainfall may have helped reduce ambient PM 2.5 levels, though the January-July 2018 average PM 2.5 was 43.7 µg/m 3 , still above the WHO's first interim target for annual average PM 2.5 . New pollution control policies are being implemented in Rwanda, including doubling the number of car-free Sundays and increased duties on older imported used cars. Continued monitoring is essential to evaluate the impact of these policies.
Air pollution in Kigali is also influenced by substantial regional contributions especially in the dry season; for example, the dry season daily minimum PM 2.5 was over 50 µg/m 3 , or over 70% of average PM 2.5 . About half the BC during the dry season also appears to be regional, which suggests a significant biomass burning influence in the regional/background contribution in line with DeWitt et al. (2019). Hence, controls on regional biomass burning are essential to reducing PM 2.5 concentrations in Kigali. Studies with more advanced instrumentation (such as an AMS) would help better quantify the time-resolved contribution of different local and regional sources to air pollution in Rwanda, which in turn would help identify additional air quality interventions.
This study was carried out in collaboration with the University of Rwanda and RCO (co-authors on this paper) and with the support of local residents who hosted the RAMPs. Towards the end of this study, the Rwanda Environmental Management Authority (REMA) acquired a reference monitoring station as well as eight RAMPs to establish their own air quality monitoring network. Low-cost monitors can significantly reduce the costs of air quality monitoring for developing countries, but local support and buy-in are keys to success.  Figure S3: Developing scaling factors for optical RAMP PM 2.5 measurements by comparison with filter-based 24-hour PM 2.5 measurements (reported by Kalisa et al. 2018