Introduction
Exposure to particulate matter (PM) is associated with cardiopulmonary
morbidity and mortality. Multiple complex pathophysiological or mechanistic
pathways have been identified as the underlying causes of this association
(Pope and Dockery, 2006). Fine particles (PM2.5, with a diameter of 2.5 µm and smaller) pose a greater threat to human health than their larger
and coarser counterparts due to their higher levels of toxicity, stronger
tendency towards deposition deep in the lungs, and longer lifetime in the
lungs (Pope and Dockery, 2006). From an environmental perspective,
PM2.5 contributes to decreased visibility, environmental damages such
as depletion of soil nutrients and acid rain effects, and material damages such
as discoloration of the Taj Mahal (U.S.
EPA, 2016a; Bergin et al., 2015).
In the US, PM2.5 is regulated and monitored under the National
Ambient Air Quality Standards (NAAQS) (U.S. EPA, 2016b). The NAAQS compliance
monitoring approves the use of both the Federal Reference Methods (FRMs) and
the Federal Equivalent Methods (FEMs) to accurately and reliably measure
PM2.5 in outdoor air (U.S. EPA, 2017). While these kinds of instruments
provide measurements of decision-making quality, they require skilled staff,
close oversight, regular maintenance, and stringent environmental operating
conditions (Chow, 1995). The personnel, infrastructure, and financial
demands of running a regulatory PM2.5 monitor make it impractical to
deploy them in a dense monitoring network and make it consequently hard to
gather high temporally and spatially resolved air quality information. The
lack of finely grained PM2.5 monitoring data hinders the characterization
of urban PM2.5 gradients and distributions (Kelly et al., 2017), and
prohibits exposure scientists from adequately quantifying the relationship
between air pollution exposures and health effects (Holstius et al., 2014).
The lack of finely resolved ambient PM2.5 data also restricts prompt
empirical verifications of emission-reduction policies and inhibits rapid
screening for urban “hot spots” (Holstius et al., 2014).
These conventional techniques' deficiencies in measuring PM2.5 along
with the technological advancements in multiple areas of electrical
engineering (Snyder et al., 2013) foster a paradigm shift to the use of
small, portable, inexpensive, and real-time sensor packages for air quality
measurement. As these sensors can provide almost instantaneous feedback
about changes in air quality and at a low cost, citizens may be more willing
to be part of “participatory measurement” including determining if they
are in areas with high levels of pollution and exploring how to decrease
their exposure. Air pollution control agencies such as the South Coast Air
Quality Management District (SCAQMD) have already been researching ways of
empowering local communities to answer questions about their specific air
quality issues with sensors and potentially engaging them in future projects
(U.S. EPA, 2016c).
Previous evaluations of numerous low-cost PM sensor models have demonstrated
promising results in comparison with FEMs or research-grade instruments in
some field studies. These models include Shinyei PPD20V (Johnson et al.,
2018), Shinyei PPD42NS (Holstius et al., 2014; Gao et al., 2015), Shinyei
PPD60PV (SCAQMD, 2015a; Jiao et al., 2016; Mukherjee et al., 2017; Johnson
et al., 2018), Alphasense OPC-N2 (SCAQMD, 2015b; Mukherjee et al., 2017;
Crilley et al., 2018), Plantower PMS1003 (Kelly et al., 2017; SCAQMD,
2017b), Plantower PMS3003 (SCAQMD, 2017a), and Plantower PMS5003 (SCAQMD,
2017c). Currently, all Plantower particulate matter sensor (PMS) models have only been tested at low to
moderately high ambient PM2.5 concentrations in the US. Kelly et al. (2017)
assessed the performance of Plantower PMS1003 against an FRM, two FEMs, and
a research-grade instrument in a 41-day field campaign in the southeast
region of Salt Lake City during winter. They reported both high 1 h
PMS–FEM PM2.5 correlations (R2=0.83–0.92) and high 24 h
PMS–FRM PM2.5 correlations (R2>0.88). The SCAQMD's
Air Quality Sensor Performance Evaluation Center (AQ-SPEC) has field-tested
Laser Egg Sensor (Plantower PMS3003 sensors), PurpleAir (Plantower PMS1003
sensors), and PurpleAir PA-II (Plantower PMS5003 sensors) with triplicates
per model located next to FEMs at ambient monitoring sites in southern
California for a roughly 2-month period (SCAQMD, 2017a, b, c). Even
though the evaluation results are still preliminary, they filled in gaps in
the documentation of the performance of the new generation of low-cost PM
sensors. The SCAQMD found that both PMS1003 and PMS5003 raw PM2.5
measurements correlated very well with the corresponding FEM GRIMM model 180
(R2>0.90 and R2>0.93, respectively) and
FEM BAM-1020 (R2>0.78 and R2>0.86,
respectively). The SCAQMD, however, reported a moderate correlation between
1 h raw PMS3003 PM2.5 measurements and the corresponding FEM BAM-1020
(R2∼0.58).
Despite the favorable correlation of these sensors in comparison with
reference monitors during these field evaluations, considerable challenges
have also been acknowledged. To date, there is only limited understanding of
the performance specifications of these emerging low-cost PM sensor models
(Lewis and Edwards, 2016). This situation is further confounded by the fact
that a model's agreement with reference instruments, and the corresponding
calibration curves established, vary with the operating conditions (relative humidity (RH),
temperature, and PM2.5 mass concentrations), the aerosol properties
(aerosol composition, size distribution, and the resulting light-scattering
efficiency), and the choice of reference instruments (Holstius et al., 2014;
Gao et al., 2015; Kelly et al., 2017). Artifacts such as varying RH and
temperature significantly interfere with accurate reporting of PM2.5
results from low-cost PM sensors. To the best of our knowledge, only Crilley
et al. (2018) have adequately compensated for the RH bias in low-cost PM
sensor measurements based on the κ-Köhler theory and they found a
roughly 1 order of magnitude improvement in the accuracy of sensor
measurements after correcting for RH bias. Also, U.S. EPA FEMs are required to
provide results comparable to the FRMs for a 24 h but not a 1 h
sampling period. An inappropriate selection of reference monitors in field
tests (especially in low PM2.5 concentration environments) might
prejudice the overall performance of low-cost sensors' short-term
measurements.
These limitations in the previous scientific work warrant more testing under
diverse ambient environmental conditions alongside various reference
monitors and more rigorous methods (statistical and calibration) to
characterize a particular low-cost sensor model's performance. It is of
paramount importance to quantify the accuracy and precision of these
sensors, as the value of the rest of the related work such as data analyses,
sensor network establishment, and citizen engagement is conditional on this.
This paper focuses on (1) comparing a new low-cost PM sensor model (Plantower
PMS3003) to different reference monitors (including a newly designated U.S.
EPA PM2.5 FEM, i.e., Teledyne API T640 PM mass monitor) in both high
(Kanpur, Uttar Pradesh, India, 1 h PM2.5 average ≥36 µg m-3) and relatively low (Durham and Research Triangle Park (RTP), NC, US, 1 h
PM2.5 average ≤10 µg m-3) ambient PM2.5
concentration environments; (2) calculating metrics including mean of ratios
and error in addition to correlation coefficient (R2) to more
rigorously interpret low-cost sensors' performance capabilities as a
function of averaging timescales; and (3) conducting appropriate RH and
temperature adjustments when possible to sensor PM2.5 responses in
order to account for systematic meteorology-induced influences and
consequently to present PM2.5 measurements with relatively high
accuracy and precision at a low cost. To our knowledge, this is the first
study to evaluate such a low-cost PM sensor model under high ambient
conditions during two typical and distinct seasons (i.e., monsoon and
post-monsoon) in India and the first to use the T640 PM mass monitor
(Teledyne API) as a reference monitor to examine sensor performance.
Materials and methods
Sensor configuration
The low-cost sensors evaluated in the present study are Plantower particulate
matter sensors (model PMS3003). The Plantower PMS3003 sensors were chosen
because (1) they are priced at a small fraction of the cost of reference
monitors (approximately USD 30) and (2) their manufacturer-reported maximum
errors are relatively low (±10 µg m-3 in the
0–100 µg m-3 range, and ±10 % in the
100–500 µg m-3 range). Unlike their PMS1003 and PMS5003
counterparts, the PMS3003 sensors are not designed as single-particle counters. The
sensors employ a light-scattering approach to measure PM1, PM2.5,
and PM10 mass concentrations in real time and are believed to apportion
light scattering to PM1, PM2.5, and PM10 based on a
confidential proprietary algorithm (Kelly et al., 2017). Ambient air laden
with different-sized particles is drawn into the sensor measurement volume
where the particles are illuminated with a laser beam, and the resulting
scattered light is measured perpendicularly by a recipient photodiode
detector. These raw light signals are filtered and amplified via electronic
filters and circuitry before being converted to mass concentrations. The
manufacturer data sheet indicates that the measurement range of this specific
sensor model spans from 0.3 to 10 µm. The configuration of the
PMS3003 sensors suggests that their detection approach is volume scattering
of the particle population rather than light scattering at the single-particle level. This volume-scattering detection approach results in PM
measurements that are independent of flow rate. PM mass concentration
measurements either with or without a manufacturer “atmospheric”
calibration are available from the Plantower sensor outputs. Nevertheless,
the manufacturer did not provide any documentation to elaborate on how the
calibration algorithm was derived. The influence of meteorological factors
(e.g., RH, temperature) was likely not accounted for in the manufacturer
calibrations. Therefore, we used the sensor-reported PM concentration
estimates without an atmospheric calibration in the current study. Prior
to field deployment, no attempt was made to calibrate these sensors under
laboratory conditions due to a potentially marked discrepancy in particle
size, composition, and optical properties of field and laboratory conditions.
The Plantower PMS3003 sensor (dimensions: 5.0 cm long × 4.3 cm wide × 2.1 cm high; weight: 40 g) along with a Sparkfun SHT15 RH
and temperature sensor, a Teensy 3.2 USB-based microcontroller, a ChronoDot
v2.1 high-precision real-time clock, a microSD card adapter, a Pololu 5 V
S7V7F5 voltage regulator, a DC barrel jack connector, and a basic 5 mm LED
were connected to a custom designed printed circuit board (PCB), shown in
Fig. 1a. We programmed the Teensy 3.2 microcontroller to measure PM mass
concentrations (µg m-3) every second and to store the time-stamped
1 min averaged measurements to text files on a microSD card. To protect
sensors from rain and direct sunlight, all components were housed in a 20.50 cm long × 9.95 cm wide × 6.70 cm high, 363 g lightweight NEMA
(National Electrical Manufacturers Association) electrical box (Bud
Industries NBF32306) as shown in Fig. 1b. The inlet of the Plantower sensor
was aligned with a hole drilled in the electrical box to ensure unrestricted
airflow into the sensor. Each Duke PM air quality monitoring package is
estimated to weigh ∼430 g in total and was continuously
powered up by a 5 V 1 A USB wall charger. The total material costs for one PM
monitoring package including the Plantower PMS3003 sensor (∼ USD 30), the supporting circuitry (∼ USD 140 including PCB
with almost all components), the enclosure (∼ USD 20), and
additional power cords (∼ USD 20) are approximately
USD 210. More detailed instructions on how to assemble the sensor packages
and information on how to use their data can be found on our web page
(http://dukearc.com, last access: 30 March 2018).
(a) The custom-designed printed circuit board (PCB) and its
components for the Plantower PMS3003 sensor packages. (b) Electrical
box housing all components for outdoor sampling.
Summary statistics for 1 h averaged measurements (mean ± SD
(range)) at the three sampling locations. Reference monitors at the sampling
locations are indicated with bold font. The dates are formatted month/day/year.
Location
Date
Instruments
PM2.5
RH
Temperature
Data
(µg m-3)
(%)
(∘C)
completeness
Duke rooftop
2/1/2017–
PMS3003-1
9±9 (0–49)
86 %
(36.003350∘ N,
3/31/2017∗
PMS3003-2
10±10 (0–51)
100 %
78.940259∘ W)
PMS3003-3
11±10 (0–52)
100 %
PMS3003-4
9±9 (0–46)
100 %
PMS3003-5
11±11 (0–55)
100 %
E-BAM
9±9(0–62)
100 %
Average Sparkfun SHT15
45±19 (9–87)
15±8 (0–36)
100 %
U.S. EPA RTP
6/30/2017–
PMS3003-1
15±7 (0–35)
100 %
(35.882816∘ N,
7/31/17
PMS3003-2
15±7 (0–36)
100 %
78.874471∘ W)
PMS3003-3
16±8 (0–39)
100 %
SHARP
7±4(0–19)
99 %
SHARP nephelometer
9±5 (0–22)
99 %
T640_Roof
10±3(3–20)
100 %
T640_Shelter
9±3(2–18)
100 %
Average Sparkfun SHT15
64±22 (27–93)
30±7 (14–45)
100 %
IIT Kanpur rooftop
6/8/2017–
PMS3003-6
55±31 (7–173)
100 %
(26.515818∘ N,
6/29/17
PMS3003-7
49±29 (7–170)
100 %
80.234337∘ E)
(monsoon)
E-BAM
36±17(0–127)
85 %
Weather station
62±15 (30–88)
33±5 (24–43)
93 %
10/23/2017–
PMS3003-6
237±88 (57–523)
98 %
11/16/17
PMS3003-7
219±91 (47–574)
98 %
(post-monsoon)
E-BAM
116±57(19–347)
93 %
Weather station
63±16 (19–88)
22±4 (14–35)
99 %
∗ All the PMS3003 sensor packages and the E-BAM were
shut down between 3 and 12 March for maintenance.
Field deployment
Three field campaigns were conducted to evaluate the performance
characteristics of Plantower PMS3003 sensors and to explore the potential
impacts from artifacts such as RH and temperature on sensors' PM2.5
measurements (Table 1). Two sites were in Durham County, NC, representing
suburban environments with low ambient PM2.5 concentrations. The other
study site was in Kanpur, Uttar Pradesh, India, representing an
urban-influenced environment. The data from Kanpur were subset into the
monsoon season with moderately high PM2.5 concentrations and the
post-monsoon season with high PM2.5 concentrations.
Low-concentration region: Durham and Research Triangle Park (RTP),
NC
The first measurement campaign in the low-concentration region was on the
rooftop of the Fitzpatrick Center, a three-story building located on the Duke
University West Campus in Durham, NC (latitude
36.003350, longitude -78.940259). The sampling
location lies in close proximity to the 28.5 km2 Duke Forest and
approximately 3.5 km from the Durham downtown and 4.5 km from the Durham
National Guard Armory monitoring station (latitude 36.0330, longitude
-78.9043). This study location is also about 950 m southwest of the Durham
Freeway, which had an annual average daily traffic of 43 000 vehicles as of
2015 (North Carolina Department of Transportation, 2015). No known principle
point source emissions are located in the surrounding area. The 3-year
average (2013–2015) for PM2.5 concentrations reported by the Durham
National Guard Armory monitoring station was 12 µg m-3, and
the reported 98th percentile daily average from 2013 to 2015 was
18 µg m-3 (North Carolina Department of Environmental
Quality, 2017). At the Duke site, five Plantower PMS3003 sensors (labeled
PMS3003-1 through PMS3003-5) were compared to a collocated environmental β attenuation monitor E-BAM-9800 (Met One Instruments). Unlike its more
advanced counterpart BAM-1020 (Met One Instruments), the E-BAM-9800 is not
currently a U.S. EPA-designated FEM for PM2.5 mass concentration
continuous monitoring, although it is ideal for rapid deployment because of
its portability and its ability to accurately track FRM or FEM results with
proper operation and regular maintenance (Met One Instruments, 2008). The
hourly values reported by the E-BAM (mg m-3) were used in the
analyses. The E-BAM's sporadic negative values caused by low actual ambient
concentrations (such as below 3 µg m-3) were replaced with
0 µg m-3 in this study. The sensor packages were strapped to
the E-BAM tripod and operated in a collocated manner for a period of 50 days
from 1 February to 31 March 2017 (all the sensor packages and the E-BAM were
shut down between 3 and 12 March for maintenance). Over the course of the
deployment, PMS3003-1 was disconnected between 14 and 21 February because of
power supply issues, and this situation rendered PMS3003-1 data 86 %
complete.
The second ambient test in the low-concentration region was performed at the
U.S. EPA's Ambient Air Innovation Research Site (AIRS) on its RTP campus, NC
(latitude 35.882816, longitude -78.874471) about 16 km southeast of the
Duke site. The ambient PM2.5 mass concentrations in the RTP region
are normally well under 12 µg m-3 (Williams et al., 2003). A
Thermo Scientific 5030 SHARP (synchronized hybrid ambient real-time
particulate monitor) monitor (U.S. EPA PM2.5 FEM) was operated by the
U.S.
EPA Office of Research and Development (ORD) and two Teledyne API T640 PM
mass monitors (U.S. EPA PM2.5 FEM) were operated by the U.S. EPA Office of
Air Quality Planning and Standards (OAQPS). The SHARP monitor is a hybrid of
a high-sensitivity nephelometer using 880 nm infrared light-emitting diodes
(IREDs) and a BAM. The SHARP continuously calculates the ratios of
dynamically time-averaged beta concentrations to dynamically time-averaged
nephelometer concentrations and continuously employs these ratios as
correction factors to adjust the raw 1 min averaged nephelometer readings.
The corrected nephelometric concentrations are reported as 1 min SHARP
measurements in micrograms per cubic meter (Thermo Fisher Scientific, 2007). The T640
monitor, first introduced in 2016, is one of the latest additions to the
list of approved U.S. EPA PM2.5 FEM monitors. The T640 is essentially an
optical aerosol spectrometer that uses light scattering to measure particle
diameters in 256 particle size classes over the 0.18–20 µm range at the
single-particle level. The 256 size classes are subsequently combined into
64 channels for mass calculation with proprietary algorithms. The light
source used by the T640 monitor is polychromatic (broadband) light. Compared
to traditional monochromatic laser scattering approaches, the polychromatic
light approach provides more robust and accurate measurements with
significantly less noise, especially over the particle size range of 1 to 10 µm (Teledyne Advanced Pollution Instrumentation, 2016). The
T640 reports 1 min resolution results in micrograms per cubic meter. The SHARP and one
of the T640 units (T640_Shelter) were installed inside an ORD
mobile laboratory and an OAQPS shelter, respectively, with roof penetration
while the other T640 unit (T640_Roof) was installed inside an
outdoor enclosure with heating, ventilation, and air conditioning (HVAC)
control on the rooftop of the OAQPS shelter. Three PMS3003 sensor packages
from the Duke site (labeled PMS3003-1 through PMS3003-3) were attached to the rail
on top of the ORD mobile laboratory approximately 3 to 4 m above the ground. The
SHARP inlet and the sensor packages' inlets were only about a meter apart. The
two T640 inlets were situated on the rooftop of the OAQPS shelter, within
about 30 m of the sensor packages' inlets. The inlets of these instruments
were positioned roughly at the same height above ground. Over the course of
the 32-day field project (30 June to 31 July 2017), all the
instruments' data completeness was 100 % except that of SHARP (99 %). The
slightly incomplete SHARP data stemmed from the removal of midnight
concentration spikes (at approximately 01:00 to 01:10 EDT, eastern daylight time) due to the daily
filter tape advancement.
High-concentration location: Indian Institute of Technology Kanpur
(IIT Kanpur) study site
Identical to the setup at the Duke site, the third field evaluation
involving two PMS3003 sensors (labeled PMS3003-6 and PMS3003-7) alongside an E-BAM
was carried out on the rooftop of the Center for Environmental Science and
Engineering inside the campus of IIT Kanpur (latitude 26.515818, longitude
80.234337). The center is a two-story building (roughly 12 m above the
ground level) that lies approximately 15 km northwest of downtown Kanpur. The institute is located upwind of Kanpur and away from major
roadways, industrial sites, and dense residential communities; therefore it
has comparatively low PM2.5 concentrations (Villalobos et al., 2015).
Kanpur is a heavily polluted industrial city on the Indo-Gangetic Plain with
a large urban area of dense population (approximately 2.7 million)
(Villalobos et al., 2015). Various small-scale industries, a coal-fired
power plant (Panki Thermal Power Station), indoor and outdoor biomass
burning, heavy vehicles on the Grand Trunk Road (a major national highway)
running through Kanpur, fertilizer plants, and refineries are the prime
contributors to air pollution (Shamjad et al., 2015;
Villalobos et al., 2015). The local climate is primarily defined as humid
subtropical with extremely hot summers and cold winters (Ghosh et al.,
2014). The monsoon season (June to September) is documented to have lower
PM2.5 concentrations than the post-monsoon season (October and
November) (Bran and Srivastava, 2017). The two sensor packages were first
deployed at the study site on 8 June 2017 for approximately 22 days (early
monsoon), and then on 23 October 2017 for approximately 25 days
(post-monsoon). Since these two sensor units were not embedded with
temperature and RH sensors, the temperature and RH data (available as 15 min
averages) were simultaneously collected from an automatic weather station,
roughly 500 m away from the study site and 2 to 3 m above the ground. Throughout
the sampling periods, error-flagged E-BAM measurements (including delta
temperature set point exceeded, flow failure, abnormal flow rate, beta count
failure) during the operation were excluded from the analyses for quality
assurance purposes, and this caused the E-BAM data to be 85 % and 93 %
complete for monsoon and post-monsoon seasons, respectively. The two sensor
packages had data completeness close to 100 % for both monsoon and
post-monsoon seasons. The temperature and RH data from the automatic weather
station were only occasionally missing due to power supply issues with an
overall 93 % and 99 % completeness for monsoon and post-monsoon
seasons, respectively.
Flow path for sensor calibrations. Note raw sensor PM2.5
measurements are uncalibrated sensor PM2.5 measurements.
Sensor calibrations
Sensor PM2.5 measurement adjustments/corrections were made as described
in the following three subsections. First, we evaluated the dependence of
sensor response on RH (Sect. 2.3.1); if this was significant we adjusted
sensor PM2.5 values for RH. Next, we investigated the sensor response
dependency on temperature (Sect. 2.3.2); if this was significant we
simultaneously adjusted sensor PM2.5 values for temperature and
calibrated sensor values based on reference monitors. If this was not
significant, we simply applied a calibration based on the reference
PM2.5 values and corrected for any nonlinear performance (Sect. 2.3.3). The calibration strategy is shown graphically in Fig. 2.
RH adjustment to sensor PM2.5 measurements
FEMs and research-grade PM analyzers typically control for RH by dynamically
heating the sample air inlet. Our sensor packages, similar to many low-cost
designs, are not equipped with any heaters or conditioners to reduce RH impact.
Therefore, the RH can significantly bias the PM2.5 mass concentrations
reported by our sensor packages. The effect of RH on the mass of atmospheric
aerosol particles has been well documented for decades. Sinclair et
al. (1974) showed that there was a 2- to 6-fold increase in the mass of
particles, depending on the properties of the particles, as the RH reached
100 %. Waggoner et al. (1981) also showed that RH above roughly 70 %
can enhance scattering coefficients of hygroscopic or deliquescent particles
in various locations in the western and midwestern US due to the growth of these
particles associated with water uptake. Zhang et al. (1994) described the
calculated scattering efficiencies of ammonium sulfate in the Grand Canyon as
a function of RH with empirical Eq. (1). This equation was later employed by
Chakrabarti et al. (2004) to predict the effect of RH on the relationship
between the nephelometric personal monitors' PM2.5 mass concentration
measurements and the results of a reference monitor (BAM). They found that
the model agreed quite well with the field data collected from both their
study and a previous study (Day and Malm, 2001). An identical equation was also among a wide
variety of approaches assessed by Soneja et al. (2014) to adjust
nephelometric personal monitor PM2.5 readings for the RH impact. We
believe lessons learned from these previous studies can be directly applied
to RH adjustments for low-cost nephelometric sensors' PM2.5 measurements
in the present study by using Eq. (1):
RHcorrectionfactor=scatteringefficiency(foragivenRH)scatteringefficiency(RH=30%)=rawPMS3003PM2.5conc.(foragivenRH)referencePM2.5conc.(foragivenRH)=a+b×RH21-RH.
Ordinary least-squares (OLS) regressions were conducted to obtain the
empirical regression parameters a and b in Eq. (1), where the dependent
variable was the RH correction factors calculated as the ratio of PMS3003
PM2.5 mass concentrations averaged across all the sensor package units
to the corresponding reference monitor concentrations at each point in time
at a sampling location, and the independent variable was the entire RH2/(1-RH) term. The RH was the measurements averaged across all the
embedded Sparkfun SHT15 RH and temperature sensors at each point in time for
the calibration models of Duke University and EPA RTP study sites, and the
measurements from the automatic weather station for the models of the IIT Kanpur
study site. The empirical equations derived were used to compute the RH
correction factor for a given RH at the sampling sites. The RH interferences
were compensated for by dividing each individual raw PMS3003 PM2.5 mass
concentration for a given RH by the RH correction factor yielded for that RH
(Eq. 2):
RHadjustedPMS3003PM2.5conc.=rawPMS3003PM2.5conc.(foragivenRH)RHcorrectionfactor(foragivenRH).
We only performed the RH adjustments when the fitted models for any of the
sampling locations over any time-averaging interval had at least a moderate
coefficient of determination (R2≥0.40). The slightly high
correlation cutoff value was implemented in this study to ensure that the RH
corrections can effectively lower the error of the low-cost sensor PM2.5
measurements. Despite the similarity of the general shape of correction
factor curves in different studies, the detailed behaviors of aerosols
diverged greatly due to considerable difference in particle chemical
composition and diameter (Waggoner et al., 1981; Zhang et al., 1994; Day and
Malm, 2001; Chakrabarti et al., 2004; Soneja et al., 2014). In a previous
study (Day and Malm, 2001), aerosol mass at some locations increased
continuously above a relatively low RH (such as 20 %), whereas at other
locations it exhibited a distinct deliquescent behavior (i.e., aerosol water
uptake occurred at a relatively high RH). Even for aerosols showing
deliquescent behavior, the observed deliquescence RH (RH threshold) varies
from study to study. Soneja et al. (2014) also found underestimation of PM
concentrations (correction factors less than 1) below 40 % RH. Because of
these uncertainties, we conducted RH adjustments across the entire range of
recorded RH without incorporating an RH threshold. Additionally, the RH
adjustments in this study were always performed separately from and prior to
either temperature adjustments or reference monitor adjustments.
Temperature adjustment to sensor PM2.5 measurements
The Akaike's information criterion (AIC) is a widely used tool for model
selection to address the fact that including additional predictors may
overfit the data (Crawley, 2017). It was used to determine the significance
of the temperature term in the PMS3003 calibration models for all the study
locations at various averaging times. The AIC penalizes more complex models
based on the number of parameters fit in that model. A lower AIC when
comparing two models for the same data set indicates a better fitting model.
In a linear regression model, an AIC difference between two models of less
than or equal to 2 indicates that the more complex model does not improve
predictive performance. Therefore, the simpler model should be adopted. We
specifically compared the AIC value of a multiple linear regression model,
which included both the reference monitor measurement and temperature as
predictor variables and without considering an interaction term (i.e., Eq. 3)
to the value of a univariate linear regression model with only the reference
monitor measurement as a predictor variable (i.e., Eq. 4). We performed the
temperature adjustments using Eq. (5) only when the AIC indicated that the
temperature predictor was significant in the calibration model (i.e.,
AICEq.(4)-AICEq.(3)>2).
[raw(orRHadjusted)PMS3003PM2.5conc.]=β0+β1referencePM2.5conc.+β2temperature[raw(orRHadjusted)PMS3003PM2.5conc.]=β0+β1referencePM2.5conc.temperatureandreferencemonitor(andRH)adjustedPMS3003PM2.5conc.=[raw(orRHadjusted)PMS3003PM2.5conc.]-β0-β2temperatureβ1
The temperature was the measurements averaged across all the embedded
Sparkfun SHT15 RH and temperature sensors at each point in time for the
models of Duke University and EPA RTP study sites and the measurements from
the automatic weather station for the models of the IIT Kanpur study site. Since
the RH adjustments in this study were always performed first, the PMS3003
PM2.5 concentration in Eqs. (3) and (4) were RH-adjusted PMS3003 PM
concentrations when RH adjustments were significant and were otherwise raw
PMS3003 PM2.5 concentrations. Additionally, temperature adjustments and
reference monitor adjustments were always conducted simultaneously when the
temperature predictor was significant because Eq. (3) consists of both the
reference monitor concentration and temperature terms as independent
variables. The AIC values for models with 24 h data are not reported in the
present study as 24 h observations generally have limited statistical power
to determine the significance of temperature in the models.
PM2.5 sensor calibrations based on reference monitor
values
The most basic calibration is a direct comparison with reference monitor
measurements. We derived reference instrument calibration equations (Eq. 4) by fitting a linear least-squares regression model to each pair of
PMS3003 (dependent variable) and collocated reference instrument's
PM2.5 mass concentrations (independent variable). The PMS3003
PM2.5 values were RH-adjusted concentrations when RH adjustments were
significant and were otherwise raw concentrations. Each PMS3003 measurement
was subsequently calibrated using Eq. (6).
When the relationship between PM2.5 mass concentrations of reference
monitors and PMS3003 sensors was nonlinear, PM2.5 sensor calibration
equations based on reference monitor values in a quadratic form (Eq. 7)
were used to describe the nonlinear performance and each PMS3003
measurement was subsequently calibrated using Eq. (8) since calibrated
values should always be on the left side of the axis of symmetry of the
parabola with a2<0. The AIC values (discussed in Sect. 2.3.2), and the root-mean-square errors (RMSEs) (Eq. 9) were used in
combination to assess the goodness of fit and accuracy of the two model
approaches (i.e., univariate linear and quadratic models) as a function of
integration times.
[referencemonitor(andRH)adjustedPMS3003PM2.5conc.]=[raw(orRHadjusted)PMS3003PM2.5conc.]-β0β1[raw(orRHadjusted)PMS3003PM2.5conc.]=a0+a1×referencePM2.5conc.+a2×referencePM2.5conc.2referencemonitor(andRH)adjustedPMS3003PM2.5conc.=-a1+a12-4a2×a0-[raw(orRHadjusted)PMS3003PM2.5conc.]2a2RMSE=1n∑i=1ny^i-yi2
Here n is the number of observations, y^i is the calibrated
PMS3003 PM2.5 mass concentrations, and yi is the reference monitor
PM2.5 mass concentrations.
Comparison of hourly PM2.5 mass concentrations between the
E-BAM and the five uncalibrated PMS3003 sensor packages between 1 February
and 31 March 2017 at Duke University.
Sensor performance metrics
Metrics such as the intercept, slope, and coefficient of determination
(R2) obtained from OLS models of sensor outputs with reference
instrument measurements are widely used to evaluate sensor performance
(Holstius et al., 2014; Gao et al., 2015; Wang et al., 2015; Jiao et al.,
2016; Cross et al., 2017; Kelly et al., 2017; Zimmerman et al., 2018). In
this study, all the R2 values in figures represent regression coefficients of
the (calibration) equations while all the R2 values in tables represent
regression coefficients between the calibrated sensor and reference
measurements. To date, only a few studies have attempted to compute
parameters other than R2 to gauge the overall performance of low-cost
sensor technologies. They typically focus on the RMSE (Holstius et al.,
2014; Cross et al., 2017; Zimmerman et al., 2018), the mean absolute error
(MAE) and the mean bias error (MBE) (Cross et al., 2017; Zimmerman et al.,
2018), and normalized residuals (Sousan et al., 2017; Kelly et al., 2017).
In addition to the intercept, slope, and R2, we also used ratios of the
calibrated PMS3003 PM2.5 mass concentrations to reference monitor
values to examine sensors' post-calibration performance. From this set of
ratios, we calculated an average ratio and 1 standard deviation (SD),
which are defined as mean of ratios and error for each sensor unit,
respectively. The mean of ratios should be close to 1 after calibration, and
we would expect the error of any PM2.5 mass concentration reported by a
particular PMS3003 unit to be within ±1 SD × 100 %
for 68 % of the time. Knowing the performance of calibrated PMS3003
sensors is particularly important for understanding these sensors' potential
for future applications such as investigating the source and transport
patterns of PM in an urban environment or examining the effectiveness of
certain PM abatement strategies.
While longer averaging times (i.e., ≥24 h) typically smooth out
noisy signals and result in enhanced sensor performance, shorter averaging
times (i.e., hours or minutes) are of growing interest, particularly in the
field of exposure assessment (Williams et al., 2017). Similar to Williams et
al. (2017), we also evaluated sensor performance over a wide range of time-averaging intervals, namely 1 min (for the EPA RTP – the only
site where 1 min reference data were available), 1 h, 6 h, 12 h, and 24 h. The purpose of
such an examination is to better understand the tradeoff between errors and
averaging times when using this type of sensor so that data accuracy and
precision can be weighed against the need for highly time-resolved data for
various desirable research or citizen science applications.
Linear regressions between aggregated PM2.5 mass concentrations
(µg m-3) of the E-BAM and the five uncalibrated PMS3003 sensors at 1
and 24 h time intervals from 1 February to 31 March 2017 at Duke University
(6 and 12 h results not shown). Marginal rugs were added to better visualize
the distribution of data on each axis.
Summary of sensor performance characteristics for the five PMS3003
PM2.5 measurements at 1, 6, 12, and 24 h time intervals from 1 February
to 31 March 2017 at Duke University. The fit coefficients for the calibration
models are provided. The R2, mean of ratios, and error are performance
characteristics for the calibrated sensor PM2.5 measurements in
comparison with reference values. The results are displayed in mean (range)
format. Note the mean statistics were obtained by fitting the models to the
PMS3003 PM2.5 measurements averaged across all five sensor package units
at each point in time.
Performance
1 h
6 h
12 h
24 h
characteristics
Adjustment
E-BAM
E-BAM, T
E-BAM
E-BAM
E-BAM
β0
3.7 (3.2–4.1)
4.5 (4.1–5.1)
-1.9 (-2.3 to -1.4)
-2.4 (-2.8 to -1.8)
-4.2 (-4.6 to -3.6)
β1
0.7 (0.6–0.8)
0.7 (0.7–0.8)
1.4 (1.2–1.5)
1.4 (1.3–1.5)
1.6 (1.5–1.8)
β2
–
-0.06 (-0.07 to -0.05)
–
–
–
R2
0.40 (0.36–0.41)
0.41 (0.36–0.42)
0.80 (0.77–0.82)
0.84 (0.81–0.86)
0.93 (0.90–0.94)
Mean of ratios1
0.97 (0.96–0.97)
0.90 (0.90–0.91)
1.05 (1.04–1.06)
1.01 (1.01–1.02)
1 (1–1.01)
Error2
201 % (195 %–223 %)
207 % (201 %–229 %)
53 % (50 %–55 %)
35 % (33 %–39 %)
15 % (13 %–18 %)
β0: intercept. β1: coefficient for
E-BAM. β2: coefficient for temperature (T). 1 Mean of
ratios of calibrated PMS3003 to E-BAM PM2.5 concentration. 2 Defined as 1 SD
of ratiosof calibrated PMS3003 to E-BAM PM2.5 concentration.
Conclusions
This study comprised three distinct field campaigns in both an
urban-influenced setting in Kanpur, India, during both monsoon (1 h averages:
[PM2.5] =36±17 µg m-3; RH =62±15 %;
temperature =33±5 ∘C) and post-monsoon seasons
([PM2.5] =116±57 µg m-3; RH =63±16 %; temperature =22±4 ∘C) and two suburban settings
in Durham ([PM2.5] =9±9 µg m-3; RH =45±19 %; temperature =15±8 ∘C) and RTP, NC, US
([PM2.5] =10±3 µg m-3; RH =64±22 %;
temperature =30±7 ∘C). The goal is to provide the
adequate range of conditions to characterize how variability in ambient
PM2.5 concentrations, meteorological factors (such as temperature and
RH), and reference monitor types (Durham and Kanpur: E-BAM; RTP: T640 and
SHARP) can affect the performance of low-cost Plantower PMS3003 sensors'
PM2.5 measurements compared with reference instruments at 1 min, 1 h, 6 h, 12 h, and 24 h integration times. This information is ultimately important for
identifying suitable research or citizen science applications for these
sensors given their quantified capabilities.
The lower mean errors of PMS3003 sensors at the EPA RTP site (from 27 % for 1 h
to 9 % for 24 h) than those at the remaining sites (Duke: from 201 % to
15 %; Kanpur monsoon: from 46 % to 17 %; Kanpur post-monsoon: from
35 % to 11 %) underscore the critical role the precision of reference
instruments (T640: ±0.5 µg m-3 for 1 h; SHARP: ±2 µg m-3 for 24 h, better than the E-BAM) plays in evaluating sensor
performance and the potential unfavorableness of beta-attenuation-based
monitors for testing sensors at low concentrations. Nonetheless, longer
averaging times (such as 24 h) typically smoothed out noisy signals and
resulted in similar levels of error, indicating the feasibility of
calibrating sensors using suboptimal reference analyzers as long as an
appropriate averaging time is chosen. Even though the RH correction factor
models might be highly location specific, it is striking to see that they
were capable of explaining up to nearly 30 % of the variance in 1 min, 1 h, and 6 h aggregated sensor measurements and reducing mean errors down from
∼22 %–27 % to roughly 10 % even at the finest 1 min time
resolution. Compared to the RH corrections, temperature corrections were
found to be relatively small and can only scale uncertainties down by 7 %
at most; however, in addition to the other corrections, this may help to
achieve the highest possible accuracy level. It is important to note that
the success of both RH and temperature corrections relies on the precision
of reference instruments. Properly accounting for these systematic
meteorology-induced influences is helpful in making high-quality PM2.5
measurements at a low cost. Additionally, we observed that PMS3003 sensors
exhibited nonlinear PM2.5 responses relative to an E-BAM when ambient
PM2.5 levels exceeded ∼125 µg m-3. We found
that the quadratic model is more suitable than the simple linear regression
model for effectively capturing this nonlinearity and can further reduce
mean errors by up to 11 %. Furthermore, we demonstrated that the
quadratic model should be chosen over the simple linear model as the
starting point (default approach) in calibrating PMS3003 PM2.5 responses since the quadratic model can always be of larger benefit to the
accuracy of PMS3003 measurements than the simple linear model even when the
nonlinearity is weak at low ambient PM2.5 concentrations or at longer
time-averaging intervals. The empirical error curves constructed by pooling
the results of all the field tests with E-BAMs as the reference monitor were
indicative of relatively stable and consistent calibrated responses above
∼70 µg m-3 for 1 h aggregated data and above
∼50 µg m-3 for 6 to 24 h aggregated data with
uncertainties roughly confined within 25 %, particularly when the
quadratic calibration models are employed.
Overall, we conclude that appropriate calibration models using ideal
reference monitors and dynamic adjustments for meteorological parameters are
an essential prerequisite for the Plantower PMS3003 sensors to achieve high
accuracy and precision over a wide range in PM2.5 concentration
typically encountered in the ambient monitoring. After proper calibration,
the Plantower PMS3003 low-cost PM sensors are promising monitors for dense,
wireless, real-time PM sensor network development in hazy urban areas such
as Delhi and Mumbai, India; to complement the existing networks by better
approximating the location of major PM2.5 sources (local vs.
regional) and by advancing our understanding of the influence of meteorology,
such as specific wind patterns, on the resulting regional PM2.5 levels
in order to guide local and regional air quality management (Hagler et al.,
2006).