pubs.acs.org/est Article
Low-Cost Sensor Performance Intercomparison, Correction Factor
Development, and 2+ Years of Ambient PM2.5 Monitoring in Accra,
Ghana
Garima Raheja,* James Nimo, Emmanuel K.-E. Appoh, Benjamin Essien, Maxwell Sunu, John Nyante,
Mawuli Amegah, Reginald Quansah, Raphael E. Arku, Stefani L. Penn, Michael R. Giordano,
Zhonghua Zheng, Darby Jack, Steven Chillrud, Kofi Amegah, R. Subramanian, Robert Pinder,
Ebenezer Appah-Sampong, Esi Nerquaye Tetteh, Mathias A. Borketey, Allison Felix Hughes,
and Daniel M. Westervelt*
Cite This: Environ. Sci. Technol. 2023, 57, 10708−10720 Read Online
ACCESS Metrics & More Article Recommendations *sı Supporting Information
ABSTRACT: Particulate matter air pollution is a leading cause of
global mortality, particularly in Asia and Africa. Addressing the
high and wide-ranging air pollution levels requires ambient
monitoring, but many low- and middle-income countries
(LMICs) remain scarcely monitored. To address these data gaps,
recent studies have utilized low-cost sensors. These sensors have
varied performance, and little literature exists about sensor
intercomparison in Africa. By colocating 2 QuantAQ Modulair-
PM, 2 PurpleAir PA-II SD, and 16 Clarity Node-S Generation II
monitors with a reference-grade Teledyne monitor in Accra,
Ghana, we present the first intercomparisons of different brands of
low-cost sensors in Africa, demonstrating that each type of low-cost sensor PM2.5 is strongly correlated with reference PM2.5, but
biased high for ambient mixture of sources found in Accra. When compared to a reference monitor, the QuantAQ Modulair-PM has
the lowest mean absolute error at 3.04 μg/m3, followed by PurpleAir PA-II (4.54 μg/m3) and Clarity Node-S (13.68 μg/m3). We
also compare the usage of 4 statistical or machine learning models (Multiple Linear Regression, Random Forest, Gaussian Mixture
Regression, and XGBoost) to correct low-cost sensors data, and find that XGBoost performs the best in testing (R2: 0.97, 0.94, 0.96;
mean absolute error: 0.56, 0.80, and 0.68 μg/m3 for PurpleAir PA-II, Clarity Node-S, and Modulair-PM, respectively), but tree-based
models do not perform well when correcting data outside the range of the colocation training. Therefore, we used Gaussian Mixture
Regression to correct data from the network of 17 Clarity Node-S monitors deployed around Accra, Ghana, from 2018 to 2021. We
find that the network daily average PM2.5 concentration in Accra is 23.4 μg/m3, which is 1.6 times the World Health Organization
Daily PM guideline of 15 μg/m32.5 . While this level is lower than those seen in some larger African cities (such as Kinshasa,
Democratic Republic of the Congo), mitigation strategies should be developed soon to prevent further impairment to air quality as
Accra, and Ghana as a whole, rapidly grow.
KEYWORDS: air quality, low-cost sensors, PurpleAir, clarity, Modulair-PM, statistical methods, machine learning, urban air,
sensor network, Ghana
1. INTRODUCTION improvement in terms of PM2.5 levels, many cities in Africa have
In 2019, over 1 million deaths in Africa were attributable to seen PM2.5 levels worsen over time (though lack of data makes
exposure to ambient particulate matter with an aerodynamic this challenging to ascertain).10−12 The Ghana Environmental
diameter less than 2.5 μm (PM 12.5). Ambient PM2.5 pollution, Protection Agency estimates that about 2800 deaths annually in
from sources including agricultural and waste burning, trans-
portation, and residential cooking, is linked to adverse health
impacts such as stroke, cardiovascular and respiratory diseases Received: December 7, 2022
such as lung cancer, obstructive pulmonary disease, and Revised: June 1, 2023
myocardial infarction.2−6 High levels of PM2.5 exposure during Accepted: June 2, 2023
pregnancy can lead to preterm birth, low birth weight, and Published: July 12, 2023
reduced cognitive function in infants.7−9 While cities in the
United States, Europe, and parts of China have observed much
© 2023 The Authors. Published by
American Chemical Society https://doi.org/10.1021/acs.est.2c09264
10708 Environ. Sci. Technol. 2023, 57, 10708−10720
Downloaded via 197.255.69.76 on September 1, 2023 at 13:28:26 (UTC).
See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.
Environmental Science & Technology pubs.acs.org/est Article
Table 1. Summary of Low-Cost Monitoring Device Attributes
Company Monitor Name (version) Time Resolution (min) Purchase Cost (USD) Cost per year (USD)
Purpleair, Inc. PurpleAir (PA-II SD) 2 $269 (note: product discontinued) $0
Clarity Movement Co. Clarity (Node-S Generation II) 20 $1000 $1400
Quant-AQ, Inc. Modulair-PM (Modulair-PM) 1 $1500 $300
the Greater AccraMetropolitan Area (GAMA) can be attributed monitor of the same cost, supporting continuous long-term
to PM2.5 pollution.
13 Anthropogenic activities such as open solid monitoring of PM2.5 exposures and potentially identifying
(household and agricultural) waste burning and emissions from hotspots. The performance and accuracy of LCS are greatly
vehicles have skyrocketed in recent decades, making low-income influenced by particle physicochemical properties including size
communities, children, pregnant women, and older adults distribution which vary by source, meteorological conditions
increasingly at risk to higher outdoor PM levels.14−212.5 This is such as temperature, and hygroscopic growth affected by
especially concerning during the dry dusty Harmattan season humidity.44−47 Hygroscopic growth occurs when particulates
(December to February), which is marked by dust from the absorb water as a function of relative humidity (which increases
Sahara Desert blowing into Accra due to large-scale circulation as relative humidity increases), thus altering their size and
patterns including the movement of the Intertropical Con- structure affecting the amount of scattered light and therefore
vergence Zone (ITCZ).22−24 For example, Alli et al. observed optically estimated PM2.5 concentrations.
that PM2.5 concentrations during the Harmattan period raised Comparing reference monitors and LCS can help develop
PM2.5 levels by 56 to 71 μg/m3 during a city-wide measurement calibration models to correct for LCS sensitivities like
campaign in the GAMA.25 Additionally, during the Harmattan, hygroscopic growth. While some literature exists on corrections
aerosol composition changes can also be expected in addition to derived from filter-based instruments, very little of the real-time
mass concentration changes.26 sensor comparison or calibration work has been done in Africa,
Ambient PM2.5 measurements enable governmental agencies where environmental conditions and emissions sources are very
to formulate policies and control strategies for PM .27−302.5 different from the United States and Europe. Adong et al.
Current estimates in Africa, however, are limited due to a severe colocated eight devices of a single LCS manufacturer (AirQo)
48
lack of both ground-based air quality observations and and twoMetOne BAM-1020 devices. McFarlane et al. applied
corresponding health outcome data. Globally, the population- Gaussian Mixture Regression (GMR) to calibrate a PurpleAir
weighted mean distance to the nearest air quality monitor is 220 colocated with a MetOne BAM-1020 at the US Embassy in
49
km; monitors are especially sparse in Africa.3,4,31 A recent study Accra, Ghana, for one year. The results revealed that the
has shown that only two out of 15 countries under the economic manufacturer-reported data from PurpleAir was biased slightly
2
community of West African States (ECOWAS) monitor PM high compared to the BAM and only moderately correlated (r =2.5 3
pollution.32 The dearth of PM monitoring sensors across 0.53, MAE = 6.2 μg/m ). However, using a GMR-based2.5
African countries has been attributed to the high capital cost of correction factor reduced the bias (MAE = 2.2 μg/m
3) and
2
the instruments, as well as the high installation and maintenance greatly improved the correlation (r = 0.88).
cost.33−35 PM2.5 monitors such as Teledyne T640 and MetOne
BAM-1020 (β Attenuation Monitor), which have US EPA 2. METHODS
Federal Equivalent Method status, require high installation, Here, we seek to conduct, to our knowledge, the first large-scale
operational and maintenance costs, and are thus not practical to LCS intercomparison study in Africa, by colocating 22 LCS from
install in the large numbers34,36 that would be necessary to three different sensor brands (PurpleAir, Inc., ClarityMovement
capture the strong spatial and temporal variation in cities. This Co., and QuantAQ, Inc.) with a reference monitor (Teledyne
poses challenges for air quality practitioners in Africa to monitor T640). We then compare the performance of four types of
PM2.5 levels and subsequently develop policies to control and statistical and machine learning models (Multiple Linear
regulate PM2.5 pollution. Regression, Gaussian Mixture Regression, XGBoost, and
Therefore, more PM2.5 ambient concentration measurement Random Forest) for calibrating and correcting data from each
has become imperative in Africa to estimate environmental and type of LCS. Finally, we apply the GMR correction factor to the
health impacts to inform policies to protect vulnerable groups network of 17 Clarity devices deployed around various
from the harmful effects of PM2.5 pollution. Sub-Saharan Africa, neighborhoods and major roadways in Accra, Ghana, from
in particular, is going through transitions in energy usage, 2018 to 2021, resulting in, to our knowledge, the longest and
urbanization, and population growth, causing economic most spatially and temporally detailed survey of PM2.5 in Accra
activities such as infrastructure, industrialization, and motoriza- to date.
tion to surge.3,31,32,37 2.1. University of Ghana (UG) Collocation. Low-cost
Within the last 10 years, there has been a dramatic rise in the sensors (LCS) manufactured by three different brands�Clarity
use of real-time, off-the-shelf low-cost sensors (LCS), such as Node-S Generation II, PurpleAir PA-II SD, and QuantAQ
those manufactured by PurpleAir, Inc., Clarity Movement Co., Modulair-PM�were colocated at the University of Ghana
and QuantAQ, Inc. for PM2.5 monitoring. Many LCS are optical (UG), Department of Physics, Legon, Accra, Ghana. The
particle counters using Plantower PMS5003 nephelometers for monitors were mounted on metal poles and fastened with cable
laser light scattering-based estimates of ambient PM 382.5. With ties at a height of five meters. Data were inspected between
substantial validation and careful calibration to correct for individual sensor units for any possible defective units and none
inherent biases, these LCS have been demonstrated to be viable were found.
alternatives to the more expensive reference-grade mon- All of the low-cost air quality monitoring devices in this study
itors.2,39−43 Further, the high spatial density of LCS networks have in-built Plantower PMS5003 nephelometers that estimate
can offer a more detailed view of a city than a single reference the concentration of fine particulate matter dispersed in ambient
10709 https://doi.org/10.1021/acs.est.2c09264
Environ. Sci. Technol. 2023, 57, 10708−10720
Environmental Science & Technology pubs.acs.org/est Article
Figure 1. Locations of Clarity nodes deployed in Accra, Ghana. The inset shows Jamestown. Pie charts show the percentage of measured days that
exceededWHODaily PM 32.5 Guidelines (15 μg/m ). The University of Ghana, where LCS were colocated as mentioned in Section 5.I., is denoted with
a black star.
air based on laser light scattering technology.50−52 The low-cost air quality monitors (Clarity Node-S, PurpleAir PA-II, and
monitoring devices also contain BOSCH BME280 sensors to QuantAQ Modulair-PM) used in this research can be found at
measure internal meteorological parameters including temper- their respective websites, https://www.clarity.io/, https://
ature and relative humidity.50,51 In addition to the Plantower www2.purpleair.com/, and https://www.quant-aq.com/. The
nephelometer, Modulair-PMs include an Alphasense OPC-N3 PM2.5 (μg/m3) columns were extracted from each dataset; for
(optical particle counter), which provides more realistic the PurpleAir, the Sensor A PM2.5 (cf = atm) and Sensor B PM2.5
estimates of supermicron aerosol not captured by Plantower (cf = atm) columns were averaged.
nephelometers, with a measurement range of 0.35−40 The LCS were collocated with a Teledyne T640 (reference-
μm.43,53−56 Modulair-PM-reported PM2.5 includes a manufac- grademonitor) placed on a two-story building rooftop located at
turer-applied correction accounting for particle density, 5.65136°N, 0.18566°W, and elevation of 108 meters above sea
aspiration efficiency, and hygroscopic growth. Summary level for a period of 4 months (11th May to 25th September,
information about the different types of monitors is given in August 2021), recording data with a 1 min resolution. The
Table 1. Detailed information on the different types of low-cost Teledyne T640 contains an aerosol sample conditioner, a
10710 https://doi.org/10.1021/acs.est.2c09264
Environ. Sci. Technol. 2023, 57, 10708−10720
Environmental Science & Technology pubs.acs.org/est Article
sample flow controller, and a 5 lpm vacuum pump with a developed by AlexanderFabisch on Github.61 GMR is beneficial
temporal resolution of 1 min. It uses broadband spectroscopy because it can produce “components” which identify regimes
using 90° white-light scattering with a polychromatic light- under which regression is classified (see McFarlane et al.49 for
emitting diode (LED), measures with a resolution of 256 sizes more information).
over 0.18−20 μm range, combined to 64 channels for mass The final is Extreme Gradient Boosting (XGBoost), which
calculation, and exceeds US EPA PM10 FEM and Class III FEM uses distributed gradient-boosted decision trees for regression.
PM2.5 performance requirements for additive and multiplicative The models are optimized to maximize R2, which is the
bias compared to FRM samplers.50,51 Instrument installation, coefficient of determination, andminimizeMean Absolute Error
calibration, and training were held virtually on August 10−11, (MAE), which is a measure of bias. In this study, XGBoost is
2020 according to the US EPA Teledyne T640 Standard implemented using the open-source xgboost Python library.62
Operating Procedure.57 Further information on the Teledyne 2.3. Accra Deployment. Seventeen Clarity Node-S
model T640 can be found at https://www.teledyne-api.com/.58 Generation I monitors were deployed across the city of Accra
The colocation site is an urban area with low-density housing, beginning in May 2018. Note that the set of monitors deployed
sparse trees, and low traffic flows. The distance to the nearest for this set of the project is different from the set of monitors
road is 500 meters. There are no known major burning or other used for the colocation work in Section 5, Part I−University of
emissions sources near the site. Ghana (UG) Collocation. Four of them were deployed along
2.2. Comparison of Machine Learning Models for major roadsides (Tetteh Quarshie, Amasaman, Malam Junction,
Correction Factors. Hourly averaged low-cost sensor data is and Achimota Interchange) onmetal poles five meters above the
cleaned by only keeping measurements where PM2.5 > 0 μg/m3, ground with a metal basket to securely fasten the monitors.
PM < 1000 μg/m32.5 , relative humidity > 0%, and in the case of Another four of the monitors were mounted at already existing
PurpleAir, where |(Channel A-Channel B)|/(Channel B) < 20%. permanent monitoring sites (East Legon, Odorkor, Dansoman,
The uptime (percent of data where these criteria were met) and North Industrial Area) across the city and one at the Ghana
during the University of Ghana colocation where these EPA Head office. The rest were deployed on electric poles in
conditions are true is 98.7% for Clarity, 99.2% for PurpleAir, Jamestown (Jamestown Bruce/Kofi Oku, Jamestown Clinic,
and 99.9% for Modulair-PM. Jamestown Coast, Ga Mashie Road, and Jamestown Hansen/
Four different models are tested for correcting low-cost sensor Asafoatse) and Chorkor (Chorkor Residence and Chorkor
data: Multiple Linear Regression, Gaussian Mixture Regression, Sackey Ansah) residential areas. The locations for the monitors
Random Forest, and XGBoost. Each model uses three can be seen in Figure 1.
measurements from the LCS as explanatory features: PM2.5, The time series of manufacture-reported and GMR-corrected
temperature (T, °C) and relative humidity (RH, %) (eq 1). In PM2.5 measurements from the Accra deployment are shown in
the case of PurpleAir, the Channel A and Channel B PM2.5 Figure S1. The GMR-corrected values are lower than the
readings are averaged. The reference value is the Teledyne T640 manufacturer-reporter values but maintain the same temporal
measurement. For all models, a 10-fold cross-validation is used trends.
for hyperparameter training, with an 80−20% training−testing Figure S2 shows the timeline of all Clarity nodes deployed
data split via random subsampling without replacement. across Accra. A total of 11,001 valid days of data were retrieved
for analysis from the network before the monitors were
PM2.5[LCS, calibrated] decommissioned.
= f(PM2.5[LCS reported], RH[LCS reported, %], 3. RESULTS
T[LCS reported, °C]) (1) 3.1. Colocation at University of Ghana. 3.1.1. Sensor
Intercomparison.Table 2 summarizes the data collected during
The first is multiple linear regression (MLR), which the University of Ghana colocation.
optimizes the best fit by minimizing the distance between the
“true” y-values (in this case, the reference monitor PM2.5) and Table 2. Summary of University of Ghana LCS
the “predicted” y-values. MLR is beneficial because of its ease of Intercomparison (Averaged by LCS Brand)
understanding, and because the model is easy to convey and
reuse using an equation. Linear (and higher-order polynomial) ReportedLow-cost Reported PM Temperature Reported Relative
fits are the least computationally intensive and simplest to code, 2.5Monitor Range (μg/m3) ange (°C) Humidity ange (%)
and therefore the most common correction methods in low-cost PurpleAir 2.1−57.6 23.1−41.0 29.5−87.6
sensor literature.33 Clarity 5.3−135.3 23.6−37.9 48.0−89.9
The second is Random Forest, a supervised ensemble model Modulair-PM 2.8−55.5 22.7−39.8 44.4−100.0
that uses a combination of decision trees. Random forest (RF) is Teledyne 5.2−60.0 4.6−34.2 58.3−100.0
useful because each decision tree theoretically isolates errors.
For this study, the RandomForestRegression from sklearn.en-
semble method in Python59 is used and is optimized using grid Figure 2 shows the hourly average time series of PM2.5
search (maximum features: 1,2,3 and maximum tree depth: measured by the reference monitor, and the averages of
1,2,3,4,5). Optimal parameter results from grid search are in the manufacture-reported PM2.5 (MR) measured by each LCS.
Supporting Information. The gaps in the data are due to monitor malfunction or
The third model is the Gaussian Mixture Regression model, scheduled maintenance. All intercomparison analysis is
which is a multivariate nonlinear regression method that models performed using only the hours (n = 3063) where at least one
the probability density of the output data conditional to the of all four types of monitors was functioning. Each low-cost
input data. It is implemented here using the sklearn.mixtur- monitor is temporally correlated with the reference monitor
e.GaussianMixture60 method in Python and the gmr library (black). Clarity devices (yellow line) tend to overestimate both
10711 https://doi.org/10.1021/acs.est.2c09264
Environ. Sci. Technol. 2023, 57, 10708−10720
Environmental Science & Technology pubs.acs.org/est Article
Figure 2. Time series of hourly averaged PM2.5 (μg/m3) LCS and reference measurements at the University of Ghana betweenMay and September of
2021. Time gaps are due to sensor maintenance and/or failure. Reference monitor in black, PurpleAir in purple, Clarity in yellow, andModulair-PM in
green.
Figure 3. Violin plot of distributions of hourly averaged PM2.5 from each monitor colocated at University of Ghana, May to September 2021. Violins
show data distribution with boxplots inside and outliers labeled with dots. Reference monitor in black, PurpleAir in purple, Clarity in yellow, and
Modulair-PM in green.
the background PM2.5 levels and also the peaks, reporting mean of 15.5 μg/m3. Overall, the Modulair-PM distribution
maximum values of over 100 μg/m3. PurpleAir devices (purple most closely matches that of Teledyne, followed by PurpleAir
line) are also biased high, but to a lesser extent, while Modulair- and Clarity.
PMs (green line) show little bias. There are two distinct rainy Figure 4A shows a scatter plot of hourly LCS PM2.5
seasons in Accra (May/June and September/October) which measurements compared to hourly reference monitor measure-
keep PM2.5 levels reasonably low compared to other ments, shaded by relative humidity. Figure 4B−D isolates each
seasons.41,63−65 Note that the intercomparison time does not LCS from Figure 4A, and the equation for the line of best fit for
include measurements taken during the Harmattan (Decem- each LCS is noted on each panel. (Figure S3 shows the scatter
ber−February), the effects of which are discussed later; further plots of hourly LCS PM2.5 measurements compared to hourly
data collection will hope to address this gap. reference monitor measurements, shaded by temperature.) The
Figure 3 shows the violin plots of hourly averaged MR drawn line indicates the line of equality between LCS and
measurements from each monitor. The mean of the Teledyne reference monitors. As seen in the scatter plots, the relationship
T640measurements is 20.4 μg/m3. The Clarity monitors show a between reference PM2.5 and manufacturer-reported PM2.5 is
broader spread than the other monitors and have means ranging linear with moderate to strong correlation (R2 = 0.82 for
from 33.5 to 39.0 μg/m3. Clarity devices also have recorded PurpleAir, R2 = 0.69 for Clarity, R2 = 0.84 for Modulair-PM).
hourly mean concentrations that exceed 100 μg/m3, which is not The slopes in the relationship, however, vary between the three
seen in any of the other devices. The PurpleAirs show similar low-cost monitoring devices. Clarity has the highest slope (1.8),
measurement ranges between the two channels of each monitor followed by PurpleAir (1.3) and Modulair-PM (0.9). Clarity
and also between monitors, with a total mean of 22.6 μg/m3. data also show the most scatter, withMAE = 13.68 μg/m3, while
The Modulair-PM monitors are within the same range as the PurpleAir MAE = 4.54 μg/m3, and Modulair-PM MAE = 3.04
PurpleAir and Teledyne but have different means due to the μg/m3. Despite Clarity and PurpleAir using the same Plantower
skewed distribution of outlier values above 40 μg/m3, with device to measure PM2.5, Figure 4B,C shows that the reported
MOD_65 having a mean of 20.5 μg/m3 and MOD_74 having a values can be quite different. Clarity and PurpleAir report
10712 https://doi.org/10.1021/acs.est.2c09264
Environ. Sci. Technol. 2023, 57, 10708−10720
Environmental Science & Technology pubs.acs.org/est Article
Figure 4. LCS PM2.5 measurements compared to Teledyne T640 measurements. (A) Hourly average PM2.5 measurements from all three low-cost
monitors, averaged by type, compared to the respective Teledyne hourly average PM2.5 measurements. The one-to-one line is shown for comparison.
The coefficients of the line of best fit are denoted for each LCS type (B−D).
Table 3. Comparison of R2 and MAE of Test Sets of Machine Learning Models to Correct LCS with Reference Measurements
Plantower “Beijing-calibrated” values as is,33 while Modulair- mean normalizedMAE) is calculated as defined by Equation 4 in
PM combines Plantower and Alphasense OPC-N3 with Giordano, et al.33 Selection of optimal model hyperparameters
assumptions about aerosol hygroscopicity to estimate PM2.5. was conducted using parameter grid search (which is a process
Each point is shaded by the relative humidity measurement from that exhaustively tests all combinations of hyperparameters)66
each LCS. PurpleAir, Clarity, andModulair-PMmean measured and 10-fold cross-validation, scoring on minimum testing MAE.
relative humidity and standard deviation are 73.30 ± 10.49,
± Modulair-PM has the best manufacturer-reported (MR)64.44 12.27, and 84.42± 15.32%, respectively. TheModulair-
PM relative humidity readings are significantly higher than the measurements, but the MAE is still improved by applying all2
other LCS. This is a known issue of early Modulair-PM devices four correction factor techniques. Clarity has the lowest R and2
which has since been corrected by switching to a new relative MAE of raw data, but the model-improved R and MAE are
humidity sensor manufactured by Sensirion. comparable to those of other LCS. Root-mean-squared error
3.1.2. ML Model Correction Comparison. Table 3 shows the (RMSE) is also shown, for comparison to US EPA methods.
results of applying the four different models to each LCS. Data Since MLR is the easiest to transfer to more use cases, the
from each monitor is hourly averaged. CvMAE (bias-corrected parameters for MLR correction for each monitor are given in
10713 https://doi.org/10.1021/acs.est.2c09264
Environ. Sci. Technol. 2023, 57, 10708−10720
Environmental Science & Technology pubs.acs.org/est Article
Figure 5. (A) Rawmeasurements (yellow line) andGMR-corrected average Clarity values (red), and Teledyne T640 referencemeasurements (black),
from the UG colocation. (B) Raw measurements (purple line) and GMR-corrected average PurpleAir values (red), and Teledyne T640 reference
measurements (black), from the UG colocation. (C) Raw measurements (green line) and GMR-corrected average Modulair-PM values (red), and
Teledyne T640 reference measurements (black), from the UG colocation.
Table S1. To use the parameters to apply MLR, for PurpleAir, Standard errors and test set predictions forMLR for each LCS
apply eq 2, and likewise for other LCS type are shown in Figures S4−S6. Figures S7−S9 show the range
and combinations of hyperparameters, and training and testing
MLR Corrected PM2.5 = 17.51 + 0.69PA PM2.5 results, for XGBoost for each type of LCS. Figures S10−S12
0.12 PA temperature, C 0.15 PAhumidity,% (2) show the Random Forest training and optimal hyperparameters.
Figures S13−S15 show the Gaussian Mixture Regression
where PAPM2.5 is PurpleAir manufacturer-reported PM2.5, training.
PATemperature°C is PurpleAir manufacturer-reported temperature In Table 3, XGBoost shows the best performance, with
in degree Celsius, and PA 2Humidity% is PurpleAir manufacturer- highest R and lowest MAE, in all three LCS cases. While it
reported relative humidity. would be intuitive to use the best-performing model, it is not
10714 https://doi.org/10.1021/acs.est.2c09264
Environ. Sci. Technol. 2023, 57, 10708−10720
Environmental Science & Technology pubs.acs.org/est Article
Figure 6.GMR-corrected monthly average time series at each site in the Clarity network deployed around Accra. The dotted red line shows the mean
of all sites. Purple boxes denote Harmattan times (December to February).
suited to our application. XGBoost and RF are both tree-based 3.2. Accra Deployment�17 Clarity Nodes. Figure 1
algorithms, which are limited in usability when applied to data shows the locations of Clarity nodes deployed across Accra, with
that is outside the ranges of the testing set. Inside the University the inset showing Jamestown, a neighborhood in Accra that was
of Ghana colocation setting, the LCS PM2.5 measurements targeted for a higher network density due to interest in
remain largely within the 0−60 μg/m3 range. However, PM2.5 understanding the impacts of a local pollution reduction
measurements from LCS deployed around Accra show many education campaign by EPA Ghana and Environment360 (a
raw values >60 μg/m3 (Figure S16), and previous studies of Ghanaian nonprofit) that began after the LCS were installed.70
Accra have demonstrated measured values greater than 100 μg/ Data from these monitors is corrected as discussed in Section
m3.67,68 Figure S17 shows that XGBoost performs well in the 0− 4.a.ii. Each pie chart indicates the percentage of measured days
60 μg/m3 but never extrapolates PM2.5 estimates outside of that exceeding the daily WHO PM2.5 guideline (15 μg/m3), with
range. Tree-based algorithms are very applicable to our study sites city-wide reporting between 37.6 and 95.7% of days above
since they result in excellent increases in R2 and decreases in WHO guidelines. In Jamestown, sites report 87.2 to 100% of
MAE and therefore create a high-performance correctionmodel, days above WHO guidelines. The Dansoman site reports the
and thus present an interesting advance in the field of LCS least percentage of days above WHO guidelines.
correction and could definitely be used in other studies where Figure 6 displays the time series of GMR-corrected monthly
conditions are at least somewhat similar to this site. We note averages from each site in the Accra network, with the dotted red
these caveats as a word of caution to those who might consider line showing the mean of all sites. The January 2019 and January
extrapolating this study using data from a different site. 2020 Harmattan periods show a greater rise in ambient PM2.5
However, the GMR and MLR corrections generally scale well concentrations than the January 2021 Harmattan period. As
outside the bounds of the training data. GMR is used to correct shown in Figure 6, the December 2018 monthly average is 1.9×
the Clarity measurements from the deployed Accra network the April 2019 monthly average; the January 2020 monthly
discussed in Figures 1 and 6−8. Note that for minimal coding average is 2.5× the April 2020 monthly average. However, the
applications, MLR would be suitable to use since the R2 and January 2021 monthly average is only 1.1× the April 2021
MAE are in a comparable range to the GMR. However, since monthly average. Other studies have confirmed the “weaker”
GMR is overall better, and can be implemented in just a few lines effects of the January 2021 Harmattan. (The coronavirus
of code as well, it is used here. lockdowns were not found to be associated with significant
Figure 5A−C shows raw LCS measurements, reference reductions in PM2.5 in this network; only 6 sites out of the 17 in
monitor measurements, and GMR-corrected PM2.5 values for the network had data and showed PM2.5 reductions in March−
each LCS at the University of Ghana colocation. The GMR May 2019 compared to March−May 2020, with a monthly
correction factor is trained on an 80% split of the colocation average reduction of 16.2%.) The small peaks during the
dataset, separately developed for each monitor, and is summertime can be explained by precipitation seasonality in
subsequently applied to all of the data. The correction factor Accra, in which the center of the northward-advancing
generally decreases the raw concentrations, which is expected intertropical convergence zone (ITCZ) passes Accra in June
given the well-known high bias in optical sensors due to relative and July and results in slightly reduced precipitation, resulting in
humidity impacts, but it maintains the trends observed by the less wet deposition and higher PM2.5 concentrations.
71 A similar
LCS. An exception to this is theModulair-PMwhich comes with pattern has been observed previously in Accra49 and also in
an existing relative humidity correction applied by the neighboring Lome,́ Togo.39
manufacturer. All of the sites tend to follow similar regional trends, but some
We note that the time period of the UG colocation does not sites are consistently measuring higher PM2.5 (Malam Junction,
include anymeasurements during theHarmattan; it is confirmed Tetteh Quarshie) and some consistently lower PM2.5 (Danso-
from other data sources25,41,69 that particle concentrations man, Ghana EPA). For the time they were online, the
during this season can be much higher in the Harmattan than Jamestown and Jamestown Coast sites measured significantly
during the wet season. Further data collection during the higher PM2.5 than the rest of the network. The Jamestown Coast
Harmattan is necessary to address this gap. site is predominantly a commercial area characterized by thick
10715 https://doi.org/10.1021/acs.est.2c09264
Environ. Sci. Technol. 2023, 57, 10708−10720
Environmental Science & Technology pubs.acs.org/est Article
Figure 7.Violin plots of daily-averaged GMR-corrected PM2.5 across the Accra, Ghana Clarity network. The dashed gray line indicates theWHODaily
PM2.5 Guideline (15 μg/m3).
Figure 8. GMR-corrected diurnal averages, by hour of day, during the (A) Harmattan (December to February) and (B) outside (March to
November). Note that the Jamestown Coast site has no data collected during the Harmattan.
smoke from singeing of slaughtered animals and fish smoking. 125 ± 86 μg/m3)74 or Isfahan, Iran (2014−2019 mean PM2.5:
The animal hides are usually burned using used car tires. 29.9−50.9 μg/m3).75 However, rapid development in the
Slaughterhouses, also known as abattoirs, have been linked to region, combined with evidence of Harmattan-linked decreasing
toxic air pollution emissions.72,73 visibility over the past 30 years,76 creates a possibility that the
Figure 7 shows the daily-averaged, GMR-corrected PM2.5 ambient PM2.5 will rise in Accra in the coming years.
values from each Clarity monitor in the Accra network. Most Monitor sites around Accra show some consistent trends. For
locations show consistent daily averages surpassing the WHO example, at all sites outside of Jamestown, the upper quartile of
Daily Guidelines, which is shown with a dashed gray line. For daily PM2.5 measurements is below 30 μg/m3, but outliers are as
comparison, the distributions of manufacturer-reported PM2.5 high as 334.5 μg/m3. Some Jamestown sites show a larger spread
from the University of Ghana colocation are shown in Figure than the rest of the network, despite having fewer measured days
S18. There is a drastic difference in the variability between of valid data. The Jamestown Coast site is a clear outlier, and the
colocated and deployed LCS, indicating that the variation in relatively higher mean measurement could be due to the
distributions of PM2.5 measured across the city are not purely proximity of the site to a slaughterhouse. The Jamestown sites
factors of instrument noise. are within 1 km of each other but show quite different trends,
The full network daily mean PM2.5 is 23.4 μg/m3, which is 1.6 further affirming the need for high-density LCS networks for
times the WHO Daily PM 32.5 guideline of 15 μg/m . This is assessing heterogeneity in ambient PM2.5 monitoring.
slightly lower but comparable to other studies which have found Figure S19 shows the annual averages across 4 years in the
the mean daily PM 3.25,67,682.5 in Accra to be 26−37 μg/m The Accra network. Note that 2018 and 2021 are incomplete years,
values in Figure 7 indicate unhealthy levels of ambient PM2.5; as shown in Figure S2. Across the 17 deployed Clarity nodes, 7
they are higher than mean daily concentrations in neighboring show decreasing annual PM2.5 averages, and 9 show increasing
Lome,́ Togo (23.5 μg/m3)39 but not as high as other cities in averages; 1 node only measured data in a single calendar year.
Africa such as Kinshasa (2019 average: 43.5 μg/m3),41 and other Figure 8 shows the difference in diurnal averages at each site
cities around the world such as Delhi (2007−2021 mean PM2.5: during and outside the Harmattan. During the Harmattan, the
10716 https://doi.org/10.1021/acs.est.2c09264
Environ. Sci. Technol. 2023, 57, 10708−10720
Environmental Science & Technology pubs.acs.org/est Article
baseline PM2.5 concentrations are elevated; the mean morning West Africa, and other cities around the world with similar
peak reaches 35.1 μg/m3 at 7 AM and the mean evening peak is meteorologies, allowing researchers to deploy low-cost sensors
28.6 μg/m3 at 6 PM. Outside the Harmattan, the mean morning and retrieve actionable data without performing their own
peak is 25.4 μg/m3 at 6 AM, and themean evening peak 22.8 μg/ unique collocations, which can be cumbersome to carry out.
m3 at 6 PM. The diurnal cycle showing peaks at hours of peak With the growing use of low-cost sensors on the African
human activity indicates qualitatively that the main sources of continent, it is vital that data from these be appropriately
pollution are anthropogenic. During the Harmattan season, validated and calibrated.
there is a large background influence, possibly from dust aerosol,
though local sources of pollution still drive the daily trends. Note ASSOCIATED CONTENT
that the Plantower PMS5003 sensors are limited in their ability ■
to measure dust and other coarse particles larger than 1−2 *sı Supporting Information
μm.77−80 The Supporting Information is available free of charge at
https://pubs.acs.org/doi/10.1021/acs.est.2c09264.
4. DISCUSSION Comparison of GMR-corrected PM2.5 (Figure S1);
By comparing measurements from the colocation of 16 Clarity timeline of clarity nodes deployed in Accra (Figure S2);
Node-S, 2 PurpleAir PA-II SD, and 2 QUANT-AQ Modulair- scatter plots of LCS vs Reference PM2.5 from UG
PMmonitors over 4 months at the University of Ghana, we find colocation shaded by temperature (Figure S3); coef-
that manufacturer-reported data from Modulair-PM monitors ficients and standard errors for Multiple linear regression
most closely correlates with data from a colocated Teledyne corrections of LCS (Table S1); multiple linear regression
T640 reference monitor, followed by PurpleAir then Clarity statistics for LCS (Figures S4−S6); XGBoost statistics for
nodes (Modulair-PM R2 = 0.84, MAE = 3.04 μg/m3; PurpleAir: LCS (Figures S7−S9); random forest statistics for LCS
R2 = 0.82, MAE = 4.54 μg/m3; Clarity: R2 = 0.69, MAE = 13.68 (Figures S10−12); GMR statistics (Figures S13−S15);
μg/m3). We then assess four machine learning models for distribution of raw PM2.5 during Accra deployment
correcting low-cost sensor data and find that while XGBoost and (Figure S16); comparison of machine learning-corrected
Random Forest raise the R2 and lower the mean absolute error LCS PM2.5 (Figure S17); distributions of raw PM2.5
the most, tree-based models like these fail at predicting PM2.5 during UG colocation (Figure S18); GMR-corrected
values outside the range of the training dataset, which in this case annual PM2.5 averages across Accra deployment (Figure
is the University of Ghana colocation. Though suchmodels offer S19); and Pearson correlations between Clarity nodes
exciting possibilities for accurate correction of low-cost sensor deployed around Accra (Figure S20) (PDF)
data, they should only be used when the training dataset covers a
broader range than the assumed range of the deployment
measurements, since these models cannot extrapolate well ■ AUTHOR INFORMATION
beyond training ranges. Multiple linear regression can also be Corresponding Authors
used for correcting low-cost sensor data; the trade-off for using Garima Raheja − Department of Earth and Environmental
this low-coding technique is slightly reduced performance. We Sciences, Columbia University, New York, New York 10027,
recommend the usage of the Gaussian mixture regression United States; Lamont-Doherty Earth Observatory of
technique, which is adept at extrapolating correction trends Columbia University, Palisades, New York 10964, United
outside of the bounds of colocation training data and also States; orcid.org/0000-0002-5037-7979;
provides large reductions in mean absolute error from the Email: garima.raheja@columbia.edu
manufacturer-reported data. (Modulair-PM R2 = 0.87, MAE = Daniel M. Westervelt− Lamont-Doherty Earth Observatory of
2.04 μg/m3; PurpleAir: R2 = 0.86, MAE = 1.93 μg/m3; Clarity: Columbia University, Palisades, New York 10964, United
R2 = 0.79, MAE = 2.27 μg/m3). We use the Gaussian Mixture States; NASA Goddard Institute for Space Science, New York,
Regression technique to correct data from 17 Clarity monitors New York 10025, United States; orcid.org/0000-0003-
deployed around Accra, Ghana, to find that the mean daily 0806-9961; Email: danielmw@ldeo.columbia.edu
average PM2.5 in the city is 23.4 μg/m3, which is 1.6 times the
WHO Daily PM2.5 guideline of 15 μg/m3. We also find that Authors
during the Harmattan, mean morning peak concentrations can James Nimo − Department of Physics, University of Ghana,
be elevated 1.2 times the non-Harmattan levels. Anthropogenic Ghana, Ghana; African Institute of Mathematical Sciences,
sources likely drive the PM2.5 concentrations especially outside Kigali, Rwanda
of the months of November throughMarch. Stark heterogeneity Emmanuel K.-E. Appoh − Ghana Environmental Protection
between data from these monitors, some of which are within 1 Agency, Accra, Ghana
km of each other but show vastly different measured Benjamin Essien − Ghana Environmental Protection Agency,
distributions, demonstrates that, as Accra continues to grow, Accra, Ghana
there will be a growing need for high-density ambient PM2.5 Maxwell Sunu − Ghana Environmental Protection Agency,
monitoring, as well as further research about the performance Accra, Ghana
(including degradation over time) of these monitors in this John Nyante−Ghana Environmental Protection Agency, Accra,
region. Potential caveats lie with the fact that the Harmattan Ghana
season was not included in the colocation time period and that Mawuli Amegah − Ghana Environmental Protection Agency,
the entire 18-node Clarity network could not be colocated Accra, Ghana
before deployment; however, these issues have been addressed Reginald Quansah − School of Public Health, University of
in other work such as in McFarlane et al. (2021a).49 The Ghana, Accra, Ghana
correction factors reported in this paper will be useful in future Raphael E. Arku − Department of Environmental Health
low-cost monitoring air quality studies in Accra, other parts of Sciences, School of Public Health and Health Sciences,
10717 https://doi.org/10.1021/acs.est.2c09264
Environ. Sci. Technol. 2023, 57, 10708−10720
Environmental Science & Technology pubs.acs.org/est Article
University of Massachusetts, Amherst, Massachusetts 01003, Funding
United States G.R. and D.M.W. acknowledge funding from the National
Stefani L. Penn − Industrial Economics, Inc, Cambridge, Science Foundation Office of International Science and
Massachusetts 02140, United States Engineering Grant Number 2020677. R.S. andM.R.G. acknowl-
Michael R. Giordano − Univ Paris Est Creteil, CNRS UMS edge State funding managed by the Agence Nationale de la
3563, Ecole Nationale des Ponts et Chaussés, Université de Recherche (ANR) under the “Programme d’Investissements
Paris, OSU-EFLUVE�Observatoire Sciences de L’Univers- d’Avenir” integrated into France 2030, under the reference
Envelopes Fluides de La Ville a ̀L’Exobiologie, F-94010 Créteil, ANR-18-MPGA-0011. The authors declare no conflict of
France interest. Corrected data from the Accra Clarity network can
Zhonghua Zheng − Department of Earth and Environmental be found here.
Sciences, The University of Manchester, Manchester M13 9PL,
U.K.; orcid.org/0000-0002-0642-650X Notes
Darby Jack − Department of Environmental Health Sciences, The authors declare no competing financial interest.
Mailman School of Public Health, Columbia University, New
York, New York 10032, United States; orcid.org/0000- ■ REFERENCES
0002-9932-0201 (1) Fisher, S.; Bellinger, D. C.; Cropper, M. L.; Kumar, P.; Binagwaho,
Steven Chillrud − Department of Environmental Health A.; Koudenoukpo, J. B.; Park, Y.; Taghian, G.; Landrigan, P. J. Air
Sciences, Mailman School of Public Health, Columbia Pollution andDevelopment in Africa: Impacts onHealth, the Economy,
University, New York, New York 10032, United States and Human Capital. Lancet Planet. Health 2021, 5, e681−e688.
Kofi Amegah − University of Cape Coast, Cape Coast, Ghana (2) Abera, A.; Mattisson, K.; Eriksson, A.; Ahlberg, E.; Sahilu, G.;
R. Subramanian − Univ Paris Est Creteil, CNRS UMS 3563, Mengistie, B.; Bayih, A. G.; Aseffaa, A.; Malmqvist, E.; Isaxon, C. Air
Ecole Nationale des Ponts et Chaussés, Université de Paris, Pollution Measurements and Land-Use Regression in Urban Sub-
OSU-EFLUVE�Observatoire Sciences de L’Univers- Saharan Africa Using Low-Cost Sensors�Possibilities and Pitfalls.
Envelopes Fluides de La Ville a ̀L’Exobiologie, F-94010 Créteil, Atmosphere 2020, 11, No. 1357.
France; Kigali Collaborative Research Centre, Kigali, Rwanda; (3) Cohen, A. J.; Brauer, M.; Burnett, R.; Anderson, H. R.; Frostad, J.;
Present Address: Qatar Environment and Energy Research Estep, K.; Balakrishnan, K.; Brunekreef, B.; Dandona, L.; Dandona, R.;Feigin, V.; Freedman, G.; Hubbell, B.; Jobling, A.; Kan, H.; Knibbs, L.;
Institute, Hamad Bin Khalifa University, Doha, Qatar Liu, Y.; Martin, R.; Morawska, L.; Pope, C. A.; Shin, H.; Straif, K.;
Robert Pinder − Environmental Protection Agency, Raleigh, Shaddick, G.; Thomas, M.; van Dingenen, R.; van Donkelaar, A.; Vos,
North Carolina 27709, United States T.; Murray, C. J. L.; Forouzanfar, M. H. Estimates and 25-Year Trends
Ebenezer Appah-Sampong−Ghana Environmental Protection of the Global Burden of Disease Attributable to Ambient Air Pollution:
Agency, Accra, Ghana An Analysis of Data from the Global Burden of Diseases Study 2015.
Esi Nerquaye Tetteh − Ghana Environmental Protection Lancet 2017, 389, 1907−1918.
Agency, Accra, Ghana (4) Rao, S.; Pachauri, S.; Dentener, F.; Kinney, P.; Klimont, Z.; Riahi,
Mathias A. Borketey − Ghana Environmental Protection K.; Schoepp, W. Better Air for Better Health: Forging Synergies in
Agency, Accra, Ghana Policies for Energy Access, Climate Change and Air Pollution. Glob.
Allison Felix Hughes − Department of Physics, University of Environ. Change 2013, 23, 1122−1130.
Ghana, Ghana, Ghana (5) Balakrishnan, K.; Dey, S.; Gupta, T.; Dhaliwal, R. S.; Brauer, M.;Cohen, A. J.; Stanaway, J. D.; Beig, G.; Joshi, T. K.; Aggarwal, A. N.;
Complete contact information is available at: Sabde, Y.; Sadhu, H.; Frostad, J.; Causey, K.; Godwin, W.; Shukla, D.
https://pubs.acs.org/10.1021/acs.est.2c09264 K.; Kumar, G. A.; Varghese, C. M.; Muraleedharan, P.; Agrawal, A.;
Anjana, R. M.; Bhansali, A.; Bhardwaj, D.; Burkart, K.; Cercy, K.;
Author Contributions Chakma, J. K.; Chowdhury, S.; Christopher, D. J.; Dutta, E.; Furtado,
M.; Ghosh, S.; Ghoshal, A. G.; Glenn, S. D.; Guleria, R.; Gupta, R.;
G.R.: data curation, formal analysis, methodology, software, Jeemon, P.; Kant, R.; Kant, S.; Kaur, T.; Koul, P. A.; Krish, V.; Krishna,
validation, visualization, writing�original draft, writing�re- B.; Larson, S. L.; Madhipatla, K.; Mahesh, P. A.; Mohan, V.;
view and editing J.N.: writing�original draft, data curation, Mukhopadhyay, S.; Mutreja, P.; Naik, N.; Nair, S.; Nguyen, G.;
methodology E.K.-E.A.: conceptualization, data curation, Odell, C. M.; Pandian, J. D.; Prabhakaran, D.; Prabhakaran, P.; Roy, A.;
funding acquisition, project administration, resources, super- Salvi, S.; Sambandam, S.; Saraf, D.; Sharma, M.; Shrivastava, A.; Singh,V.; Tandon, N.; Thomas, N. J.; Torre, A.; Xavier, D.; Yadav, G.; Singh,
vision B.E.: data curation M.S.: data curation J.N.: data curation S.; Shekhar, C.; Vos, T.; Dandona, R.; Reddy, K. S.; Lim, S. S.; Murray,
M.A.: data curation R.Q.: project administration, conceptualiza- C. J. L.; Venkatesh, S.; Dandona, L. The Impact of Air Pollution on
tion R.E.A.: writing�review and editing, S.L.P.: conceptualiza- Deaths, Disease Burden, and Life Expectancy across the States of India:
tion, project administration, funding acquisition, resources The Global Burden of Disease Study 2017. Lancet Planet. Health 2019,3, e26−e39.
M.R.G.: writing�review and editing, data curation, method- (6) Cohen, A. J.; Anderson, H. R.; Ostro, B.; Pandey, K. D.;
ology Z.Z.: methodology D.J.: conceptualization S.C.: con- Krzyzanowski, M.; Künzli, N.; Gutschmidt, K.; Pope, A.; Romieu, I.;
ceptualization K.A.: data curation, resources, writing�review Samet, J. M.; Smith, K. The Global Burden of Disease Due to Outdoor
and editing R.S.: conceptualization, supervision, data curation Air Pollution. J. Toxicol. Environ. Health A 2005, 68, 1301−1307.(7) Johnson, M.; Shin, H. H.; Roberts, E.; Sun, L.; Fisher, M.; Hystad,
R.P: conceptualization, funding acquisition, resources, E.A.-S.: P.; Van Donkelaar, A.; Martin, R. V.; Fraser, W. D.; Lavigne, E.; Clark,
funding acquisition, supervision, E.N.T.: data curation M.A.B.: N.; Beaulac, V.; Arbuckle, T. E. Critical Time Windows for Air
data curation A.F.H.: supervision, data curation, project Pollution Exposure and Birth Weight in a Multicity Canadian
administration, methodology D.M.W.: conceptualization, meth- Pregnancy Cohort. Epidemiology 2022, 33, 7−16.(8) Šrám, R. J.; Binkov, B.; Dejmek, J.; Bobak, M. Ambient Air
odology, funding acquisition, project administration, super- Pollution and Pregnancy Outcomes: A Review of the Literature.
vision, resources, writing�review and editing. Environ. Health Perspect. 2005, 113, 375−382.
10718 https://doi.org/10.1021/acs.est.2c09264
Environ. Sci. Technol. 2023, 57, 10708−10720
Environmental Science & Technology pubs.acs.org/est Article
(9) Woodruff, T. J.; Parker, J. D.; Kyle, A. D.; Schoendorf, K. C. (30) Development of an integrated policy making tool for assessing air
Disparities in Exposure to Air Pollution during Pregnancy. Environ. quality and human health benefits of air pollution control | SpringerLink
Health Perspect. 2003, 111, 942−946. 2015 https://link.springer.com/article/10.1007/s11783-015-0796-8.
(10) Health Effects Institute. State of Global Air 2020 Special Report, (accessed November 11, 2022).
2020. (31) Martin, R. V.; Brauer, M.; van Donkelaar, A.; Shaddick, G.;
(11) WHO. Report - WHO Air Quality Database (Update) 2022 Narain, U.; Dey, S. No One Knows Which City Has the Highest
https://www.who.int/data/gho/data/themes/air-pollution/who-air- Concentration of Fine Particulate Matter. Atmos. Environ.: X 2019, 3,
quality-database. (accessed March 22, 2023). No. 100040.
(12) Singh, A.; Avis, W. R.; Pope, F. D. Visibility as a Proxy for Air (32) Mir Alvarez, C.; Hourcade, R.; Lefebvre, B.; Pilot, E. A Scoping
Quality in East Africa. Environ. Res. Lett. 2020, 15, No. 084002. Review on Air Quality Monitoring, Policy and Health in West African
(13) Mudu, P. Ambient Air Pollution and Health in Accra, Ghana; Cities. Int. J. Environ. Res. Public Health 2020, 17, No. 9151.
World Health Organization: Geneva, 2021. (33) Giordano, M. R.; Malings, C.; Pandis, S. N.; Presto, A. A.;
(14) Marais, E. A.; Wiedinmyer, C. Air Quality Impact of Diffuse and McNeill, V. F.; Westervelt, D. M.; Beekmann, M.; Subramanian, R.
Inefficient Combustion Emissions in Africa (DICE-Africa). Environ. Sci. From Low-Cost Sensors to High-Quality Data: A Summary of
Technol. 2016, 50, 10739−10745. Challenges and Best Practices for Effectively Calibrating Low-Cost
(15) McDuffie, E. E.; Smith, S. J.; O’Rourke, P.; Tibrewal, K.; Particulate Matter Mass Sensors. J. Aerosol Sci. 2021, 158, No. 105833.
Venkataraman, C.; Marais, E. A.; Zheng, B.; Crippa, M.; Brauer, M.; (34) Hagan, D. H.; Kroll, J. H. Assessing the Accuracy of Low-Cost
Martin, R. V. A Global Anthropogenic Emission Inventory of Optical Particle Sensors Using a Physics-Based Approach. Atmos. Meas.
Atmospheric Pollutants from Sector- and Fuel-Specific Sources Tech. 2020, 13, 6343−6355.
(1970−2017): An Application of the Community Emissions Data (35)Morawska, L.; Thai, P. K.; Liu, X.; Asumadu-Sakyi, A.; Ayoko, G.;
System (CEDS). Earth Syst. Sci. Data 2020, 12, 3413−3442. Bartonova, A.; Bedini, A.; Chai, F.; Christensen, B.; Dunbabin,M.; Gao,
(16) Robust relationship between air quality and infant mortality in J.; Hagler, G. S. W.; Jayaratne, R.; Kumar, P.; Lau, A. K. H.; Louie, P. K.
Africa | Nature. https://www.nature.com/articles/s41586-018-0263-3. K.; Mazaheri, M.; Ning, Z.; Motta, N.; Mullins, B.; Rahman, M. M.;
(accessed November 11, 2022). Ristovski, Z.; Shafiei, M.; Tjondronegoro, D.; Westerdahl, D.; Williams,
(17) Baayoun, A.; Itani, W.; El Helou, J.; Halabi, L.; Medlej, S.; El R. Applications of Low-Cost Sensing Technologies for Air Quality
Malki, M.; Moukhadder, A.; Aboujaoude, L. K.; Kabakian, V.; Monitoring and Exposure Assessment: How Far Have They Gone?
Mounajed, H.; Mokalled, T.; Shihadeh, A.; Lakkis, I.; Saliba, N. A. Environ. Int. 2018, 116, 286−299.
Emission Inventory of Key Sources of Air Pollution in Lebanon. Atmos. (36) Raheja, G.; Harper, L.; Hoffman, A.; Gorby, Y.; Freese, L.;
Environ. 2019, 215, No. 116871. O’Leary, B.; Deron, N.; Smith, S.; Auch, T.; Goodwin, M.; Westervelt,
(18) Magnoudéwa, B. B. Rapport Sectoriel D’ Inventaire Des Gaz A D. M. Community-Based Participatory Research for Low-Cost Air
Effet De Serre Du Togo, 2021.
(19) Anderson, C. M.; Kissel, K. A.; Field, C. B.; Mach, K. J. Climate Pollution Monitoring in the Wake of Unconventional Oil and Gas
Change Mitigation, Air Pollution, and Environmental Justice in Development in the Ohio River Valley: Empowering Impacted
California. Environ. Sci. Technol. 2018, 52, 10829−10838. Residents through Community Science. Environ. Res. Lett. 2022, 17,
(20) UNICEF. Silent Suffocation in Africa Air Pollution Is a Growing No. 065006.
Menace, Affecting the Poorest Children the Most, 2019. (37) Katoto, P. D. M. C.; Byamungu, L.; Brand, A. S.; Mokaya, J.;
(21) Air Pollution From Forest and Vegetation Fires in Southeast Asia Strijdom, H.; Goswami, N.; De Boever, P.; Nawrot, T. S.; Nemery, B.
Disproportionately Impacts the Poor - Reddington - 2021 - GeoHealth Ambient Air Pollution and Health in Sub-Saharan Africa: Current
- Wiley Online Library. https://agupubs.onlinelibrary.wiley.com/doi/ Evidence, Perspectives and a Call to Action. Environ. Res. 2019, 173,
full/10.1029/2021GH000418. (accessed November 11, 2022). 174−188.
(22) Schwanghart, W.; Schütt, B. Meteorological Causes of (38) Ouimette, J. R.; Malm, W. C.; Schichtel, B. A.; Sheridan, P. J.;
Harmattan Dust in West Africa. Geomorphology 2008, 95, 412−428. Andrews, E.; Ogren, J. A.; Arnott, W. P. Evaluating the PurpleAir
(23) He, C.; Breuning-Madsen, H.; Awadzi, T. W. Mineralogy of dust Monitor as an Aerosol Light Scattering Instrument. Atmos. Meas. Tech.
deposited during the Harmattan season in GhanaGeogr. Tidsskr.-Dan. J. Discuss. 2021, 1−35.
Geogr. 107 9 15 DOI: 10.1080/00167223.2007.10801371. (39) Raheja, G.; Sabi, K.; Sonla, H.; Gbedjangni, E. K.; McFarlane, C.
(24) A long-term experimental study of the Saharan dust presence in M.; Hodoli, C. G.; Westervelt, D. M. A Network of Field-Calibrated
West Africa - ScienceDirect 2008 https://www.sciencedirect.com/ Low-Cost SensorMeasurements of PM2.5 in Lome,́ Togo, OverOne to
science/article/pii/S0169809507001196. (accessed November 11, Two Years. ACS Earth Space Chem. 2022, DOI: 10.1021/acsearthspa-
2022). cechem.1c00391.
(25) Alli, A. S.; Clark, S. N.; Hughes, A.; Nimo, J.; Bedford-Moses, J.; (40) Shusterman, A. A.; Teige, V. E.; Turner, A. J.; Newman, C.; Kim,
Baah, S.; Wang, J.; Vallarino, J.; Agyemang, E.; Barratt, B.; Beddows, A.; J.; Cohen, R. C. The BErkeley Atmospheric CO<Sub>2</Sub>
Kelly, F.; Owusu, G.; Baumgartner, J.; Brauer, M.; Ezzati, M.; Agyei- Observation Network: Initial Evaluation. Atmos. Chem. Phys. 2016, 16,
Mensah, S.; Arku, R. E. Spatial-Temporal Patterns of Ambient Fine 13449−13463.
ParticulateMatter (PM2.5) and Black Carbon (BC) Pollution in Accra. (41) McFarlane, C.; Isevulambire, P. K.; Lumbuenamo, R. S.; Ndinga,
Environ. Res. Lett. 2021, 16, No. 074013. A. M. E.; Dhammapala, R.; Jin, X.; McNeill, V. F.; Malings, C.;
(26) Weinstein, J. P.; Hedges, S. R.; Kimbrough, S. Characterization Subramanian, R.; Westervelt, D. M. First Measurements of Ambient
and Aerosol Mass Balance of PM2.5 and PM10 Collected in Conakry, PM2.5 in Kinshasa, Democratic Republic of Congo and Brazzaville,
Guinea during the 2004 Harmattan Period. Chemosphere 2010, 78, Republic of Congo Using Field-Calibrated Low-Cost Sensors. Aerosol
980−988. Air Qual. Res. 2021, 21, No. 200619.
(27) Israel, B. A.; Schulz, A. J.; Parker, E. A.; Becker, A. B. Community- (42) Baron, R.; Saffell, J. Amperometric Gas Sensors as a Low Cost
Campus Partnerships for Health. Community-Based Participatory Emerging Technology Platform for Air Quality Monitoring Applica-
Research: Policy Recommendations for Promoting a Partnership tions: A Review. ACS Sens. 2017, 2, 1553−1566.
Approach in Health Research. Educ. Health 2001, 14, 182−197. (43) Crilley, L. R.; Shaw, M.; Pound, R.; Kramer, L. J.; Price, R.;
(28) Li, S.-T.; Shue, L.-Y. Data Mining to Aid Policy Making in Air Young, S.; Lewis, A. C.; Pope, F. D. Evaluation of a Low-Cost Optical
Pollution Management. Expert Syst. Appl. 2004, 27, 331−340. Particle Counter (Alphasense OPC-N2) for Ambient Air Monitoring.
(29) An integrated assessment of two decades of air pollution policy Atmos. Meas. Tech. 2018, 11, 709−720.
making in Spain: Impactscosts and improvements - ScienceDirect 2015 (44) Kumar, V.; Sahu, M. Evaluation of Nine Machine Learning
h t t p s : / /www . s c i e n c e d i r e c t . c om/ s c i e n c e / a r t i c l e / p i i / Regression Algorithms for Calibration of Low-Cost PM2.5 Sensor. J.
S0048969715300735. (accessed November 11, 2022). Aerosol Sci. 2021, 157, No. 105809.
10719 https://doi.org/10.1021/acs.est.2c09264
Environ. Sci. Technol. 2023, 57, 10708−10720
Environmental Science & Technology pubs.acs.org/est Article
(45) Malings, C.; Tanzer, R.; Hauryliuk, A.; Saha, P. K.; Robinson, A. (67) Arku, R. E.; Vallarino, J.; Dionisio, K. L.; Willis, R.; Choi, H.;
L.; Presto, A. A.; Subramanian, R. Fine Particle Mass Monitoring with Wilson, J. G.; Hemphill, C.; Agyei-Mensah, S.; Spengler, J. D.; Ezzati,
Low-Cost Sensors: Corrections and Long-Term Performance Evalua- M. Characterizing Air Pollution in Two Low-Income Neighborhoods
tion. Aerosol Sci. Technol. 2020, 54, 160−174. in Accra, Ghana. Sci. Total Environ. 2008, 402, 217−231.
(46)Nowack, P.; Konstantinovskiy, L.; Gardiner, H.; Cant, J. Towards (68) Amegah, A. K.; Dakuu, G.; Mudu, P.; Jaakkola, J. J. K. Particulate
Low-Cost and High-Performance Air Pollution Measurementsusing Matter Pollution at Traffic Hotspots of Accra, Ghana: Levels, Exposure
Machine Learning Calibration Techniques. Atmos. Meas. Tech. 2020, Experiences of Street Traders, and Associated Respiratory and
1−30. Cardiovascular Symptoms. J. Exposure. Sci. Environ. Epidemiol. 2022,
(47) Si, M.; Xiong, Y.; Du, S.; Du, K. Evaluation and Calibration of a 32, 333−342.
Low-Cost Particle Sensor in Ambient Conditions Using Machine- (69) Aboh, I. J. K.; Henriksson, D.; Laursen, J.; Lundin, M.; Ofosu, F.
Learning Methods. Atmos. Meas. Tech. 2020, 13, 1693−1707. G.; Pind, N.; Selin Lindgren, E.; Wahnström, T. Identification of
(48) Adong, P.; Bainomugisha, E.; Okure, D.; Sserunjogi, R. Applying Aerosol Particle Sources in Semi-Rural Area of Kwabenya, near Accra,
Machine Learning for Large Scale Field Calibration of Low-Cost PM2.5 Ghana, by EDXRF Techniques. X-Ray Spectrom. 2009, 38, 348−353.
and PM10 Air Pollution Sensors. Appl. AI Lett. 2022, 3, No. e76. (70) Stefani, P. Jamestown Open Burning Neighborhood Pilot Data
(49) McFarlane, C.; Raheja, G.; Malings, C.; Appoh, E. K. E.; Hughes, Analysis Draft Report 2021.
A. F.; Westervelt, D. M. Application of GaussianMixture Regression for (71) Nicholson, S. E. The ITCZ and the Seasonal Cycle over
the Correction of Low Cost PM . Monitoring Data in Accra, Ghana. Equatorial Africa. Bull. Am. Meteorol. Soc. 2018, 99, 337−348.2 5
ACS Earth Space Chem. 2021, 5, 2268−2279. (72) Singh, A. L.; Jamal, S.; Baba, S. A.; Islam, M. M. Environmental
(50) PurpleAir, Inc. PurpleAir. Our Technology - PurpleAir Laser and Health Impacts from Slaughter Houses Located on the City
Particle Counters & Arduinos. 2021 https://www2.purpleair.com/ Outskirts: A Case Study. J. Environ. Prot. 2014, 05, 566−575.
pages/technology. (accessed October 18, 2021). (73) Jonah, A. E. Determination of Some Air Pollutants
(51) MODULAIR-PM. MODULAIR-PM 2022. https://docs.quant- AndMeteorological Parameters in Abattoir, NTAK INYANG in UYOL.G.A of AKWA IBOMState in Nigeria. Int. J. Sci. Manag. Stud. 2020, 3,
aq.com/modulair-pm. (accessed November 11, 2022).
| 1−8.(52) Air Quality Measurement Resources Clarity Movement Co. (74) Chetna; Dhaka, S. K.; Longiany, G.; Panwar, V.; Kumar, V.;
h t t p s : / / w w w . c l a r i t y . i o / a i r - q u a l i t y - m o n i t o r i n g - Malik, S.; Rao, A. S.; Singh, N.; Dimri, A. P.; Matsumi, Y.; Nakayama,
resources#documentation. (accessed November 16, 2022). T.; Hayashida, S. Trends and Variability of PM2.5 at Different Time
(53) Hagan, D. H.; Cross, E. S. Introduction to the MODULAIR-PM Scales over Delhi: Long-Term Analysis 2007−2021. Aerosol Air Qual.
2022 DOI: 10.5281/zenodo.7062168. Res. 2022, 22, No. 220191.
(54) Bezantakos, S.; Costi, M.; Barmpounis, K.; Antoniou, P.; (75)Hajizadeh, Y.; Jafari, N.;Mohammadi, A.;Momtaz, S.M.; Fanaei,
Vouterakos, P.; Keleshis, C.; Sciare, J.; Biskos, G. Qualification of the F.; Abdolahnejad, A. Concentrations and Mortality Due to Short- and
Alphasense Optical Particle Counter for Inline Air Quality Monitoring. Long-Term Exposure to PM2.5 in a Megacity of Iran (2014−2019).
Aerosol Sci. Technol. 2021, 55, 361−370. Environ. Sci. Pollut. Res. 2020, 27, 38004−38014.
(55) Sousan, S.; Koehler, K.; Hallett, L.; Peters, T. M. Evaluation of (76) Balarabe, M. A. The Thirty Years Trend Analysis of Harmattan
the Alphasense Optical Particle Counter (OPC-N2) and the Grimm Season Visibility and Temperature in Sahel Zone of Nigeria. Phys. Mem.
Portable Aerosol Spectrometer (PAS-1.108).Aerosol Sci. Technol. 2016, - J. Theor. Appl. Phys. 2019, 1, 15−21.
50, 1352−1365. (77) Stavroulas, I.; Grivas, G.; Michalopoulos, P.; Liakakou, E.;
(56) Particulate Matter Sensors | PM Sensor. Alphasense. https:// Bougiatioti, A.; Kalkavouras, P.; Fameli, K. M.; Hatzianastassiou, N.;
www.alphasense.com/products/optical-particle-counter/. (accessed Mihalopoulos, N.; Gerasopoulos, E. Field Evaluation of Low-Cost PM
February 22, 2023). Sensors (Purple Air PA-II) Under Variable Urban Air Quality
(57) Standard Operating Procedure Teledyne Model 640. 2019. Conditions, in Greece. Atmosphere 2020, 11, No. 926.
(58)Model T640 with NumaView Software- T640 PMMass Monitor (78) Robinson, D. L. Accurate, Low Cost PM2.5 Measurements
2022 https://www.teledyne-api.com/products/particulate- Demonstrate the Large Spatial Variation in Wood Smoke Pollution in
instruments/t640. (accessed November 11, 2022). Regional Australia and Improve Modeling and Estimates of Health
(59) sklearnensembleRandomForestRegressor. scikit-learn. https:// Costs. Atmosphere. 2023 https://www.mdpi.com/2073-4433/11/8/
sciki t - learn/stable/modules/generated/sklearn.ensemble . 856. (accessed February 22, 2023).
RandomForestRegressor.html. (accessed November 11, 2022). (79) Kosmopoulos, G.; Salamalikis, V.; Pandis, S. N.; Yannopoulos,
(60) sklearnmixtureGaussianMixture. scikit-learn 2022 https://scikit- P.; Bloutsos, A. A.; Kazantzidis, A. Low-Cost Sensors for Measuring
learn/stable/modules/generated/sklearn.mixture.GaussianMixture. Airborne Particulate Matter: Field Evaluation and Calibration at a
html. (accessed November 11, 2022). South-Eastern European Site. Sci. Total Environ. 2020, 748,
(61) Fabisch, A. Gmr 2022 https://github.com/AlexanderFabisch/ No. 141396.
gmr. (accessed November 11, 2022). (80) Kuula, J.; Friman, M.; Helin, A.; Niemi, J. V.; Aurela, M.;
(62) XGBoost Documentation � xgboost 171 documentation 2022 Timonen, H.; Saarikoski, S. Utilization of Scattering and Absorption-
https://xgboost.readthedocs.io/en/stable/index.html. (accessed No- Based Particulate Matter Sensors in the Environment Impacted by
vember 11, 2022). Residential Wood Combustion. J. Aerosol Sci. 2020, 150, No. 105671.
(63) Kwak, H.-Y.; Ko, J.; Lee, S.; Joh, C.-H. Identifying the
Correlation between Rainfall, Traffic Flow Performance and Air
Pollution Concentration in Seoul Using a Path Analysis. Transp. Res.
Procedia 2017, 25, 3552−3563.
(64) Tian, X.; Cui, K.; Sheu, H.-L.; Hsieh, Y.-K.; Yu, F. Effects of Rain
and Snow on the Air Quality Index, PM2.5 Levels, and Dry Deposition
Flux of PCDD/Fs. Aerosol Air Qual. Res. 2021, 21, No. 210158.
(65) Yoo, J.-M.; Lee, Y.-R.; Kim, D.; Jeong, M.-J.; Stockwell, W. R.;
Kundu, P. K.; Oh, S.-M.; Shin, D.-B.; Lee, S.-J. New Indices for Wet
Scavenging of Air Pollutants (O3, CO, NO2, SO2, and PM10) by
Summertime Rain. Atmos. Environ. 2014, 82, 226−237.
(66) 32Tuning the hyper-parameters of an estimator. scikit-learn.
https://scikit-learn/stable/modules/grid_search.html.(accessed Feb-
ruary 22,2023).
10720 https://doi.org/10.1021/acs.est.2c09264
Environ. Sci. Technol. 2023, 57, 10708−10720