Seasonal Field Calibration of Low-Cost PM_2.5 Sensors in Different Locations with Different Sources in Thailand

Abstract

Low-cost sensors (LCS) have been increasingly deployed to monitor PM_2.5 concentrations. More than 1500 LCS have been installed in Thailand to increase public awareness of air quality. However, performance of these sensors has not been systematically investigated. In this study, PM_2.5 LCS were co-located next to a PM_2.5 federal equivalent method (FEM) reference instrument at three Thai locations—in the north, center and northeast. We evaluated the performance of a PM_2.5 LCS (PMS7003, Plantower) to understand the key factors affecting performance, including emission sources, relative humidity, temperature and PM_2.5 concentration. Low PM concentration and high humidity levels had a significant impact on performance. Sensors in a high traffic emission area showed low correlation. The unadjusted PM_2.5 LCS performance varied with locations. Errors were mainly observed at low concentrations. They significantly underestimated concentrations in congested urban environments. After calibration, accuracy was improved with multiple regression models. The performance of sensors only at Chiang Mai (CM) during the dry season and Ubon Ratchathani (URT) during the dry and wet seasons were acceptable with coefficient of variation: 5.8 ± 4.7–6.8 ± 5.0%, slope: 0.829–0.945, intercept: 1.12–5.49 µg/m³, R²: 0.880–0.934 and RMSE: 4.3–5.1 µg/m³. In the congested area in Bangkok (BKK), they underestimated concentrations of small particles.

1. Introduction

Air pollutants in many countries in Southeast Asia are influenced from diverse sources. In Thailand, the major sources are vehicles, industry, long-range transport, secondary aerosol formation as well as open biomass burning [1,2,3,4,5,6,7,8,9,10,11]. Traffic is the most influential factor on air quality in the capital city of Bangkok, though some studies indicated that biomass burning was the major cause of atmospheric aerosols during the dry period [6,7,12]. Air quality in northern Thailand was annually affected by domestic and transboundary sources as a result of forest fires and agriculture waste burning during the dry season [1,3,9]. Agricultural area in Thailand covers 24 M ha or 46.5% of the total land [13]. The northeastern region is the largest agricultural area (~10 M ha), followed by the northern and central regions [14]. Normally, agricultural waste has been burned in the open for clearing and preparing for the next crop [15]: January to April is the biomass intensive burning period, not only in Thailand, but also in other countries in the Mekong Subregion, including Laos, Cambodia, Vietnam and Myanmar [2]. All these events cause Thailand serious and long-term air pollution problems, particularly fine particles or PM_2.5. Though the standard for PM_2.5 in Thailand has been enforced since 2010, it was not until recently that there has been strong public awareness: the advent and widespread use of low-cost and easily acquired PM_2.5 sensors contributed to this.

Air quality monitoring networks in Thailand have been established by the Pollution Control Department (PCD) to monitor the air pollutants, including PM₁₀, PM_2.5 and noxious gases (CO, NO₂, SO₂ and O₃). PCD started to monitor ambient PM_2.5 concentrations in 2011 with only three monitoring stations in three Thai megacities—Rayong, Chiang Mai and Bangkok. National allowable ambient PM_2.5 concentrations in Thailand were set on 28 Jan 2010, with the average 24-h concentration not to exceed 50 µg/m³ [16]. Currently, ambient PM_2.5 concentrations monitoring stations have been expanded to more than 77 stations in 45 cities [16]. However, the FEM instruments [17], such as the tapered element oscillating microbalance (TEOM) and the beta attenuation monitor (BAM), are still relatively expensive and they are only deployed at central monitoring sites in each province. One monitors station cannot represent the correct spatial PM_2.5 concentrations [18,19,20,21].

Low-cost sensors (LCS) for air quality monitoring have been widely used in recent years to fill gaps in the existing air quality monitoring networks and provide the public with locally relevant particle concentrations, especially PM_2.5, in real time, at much higher spatial and temporal resolutions. At present, LCS have been deployed to monitor PM_2.5 concentrations, with more than 1500 monitoring stations installed in Thailand [22]. In general, the performance of the PM_2.5 LCS showed a high correlation relative to FEM instruments (R² > 0.85) in the laboratory under controlled conditions [19,21,23,24,25]. However, typically, they were less accurate in the field, where correlations were found in the range 0.36 ≤ R² ≤ 0.99 [20,21,26,27,28,29], with variations due to sources, particle sizes and shapes, compositions and concentrations [30]. The different meteorological conditions, including temperature and relative humidity (RH), also influenced LCS performance (20). Therefore, a PM_2.5 LCS calibrated under laboratory condition may not be suitable for direct field measurements [21,27,31].

Calibration of LCS is necessary due to both the physical and chemical characteristics of aerosols [32] and meteorological conditions [28,33]. Levy Zamora et al. [32] evaluated three Plantower PMSA003 sensors exposed to eight PM sources, including incense, oleic acid, NaCl, talcum powder, cooking emissions and monodispersed polystyrene latex spheres, under controlled laboratory conditions. The accuracy ranged from 13% to >90% compared with reference instruments; hence, PM sources influenced LCS performance. The LCS were most accurate for sizes < 1 μm. Additionally, meteorological conditions, including temperature and RH, significantly influenced LCS performance [28,33]. Levy Zamora et al. [32] showed that the LCS error increased at higher RH levels (greater than 80%) in the field. Jayarathne et al. [33] also found that the LCS mass concentrations were at least 80% higher than the FEM values at high RH levels (78–89%). Additionally, Levy Zamora et al. [32] showed that an RH level > 50% deceased the accuracy of the LCS. Therefore, they need to be calibrated under realistic ambient conditions before deployed for ambient PM_2.5 monitoring.

A common approach for field calibration of the LCS is to collocate the LCS with an FEM/federal reference method (FRM) instrument and use mathematical models to adjust the LCS data to reference data [34,35]. Regression model is one of the earliest methods for LCS calibration [28,31,35]. Moreover, the machine learning (ML) methods were recently applied to improve the accuracy of the linear regression [34,36]. However, regression models including linear regression (LR), multi-linear regression (MLR) and nonlinear regression (NLR) are still widely used to calibrate the LCS, because of their simplicity and ease of implementation [35]. Zheng et al. [28] developed calibration models to improve the LCS performance against an FEM by using the LR, MLR and NLR based on PM_2.5 concentration, temperature and RH in India. Zusman et al. [31] also developed an MLR calibration model based on the same parameters. The model showed the moderate accuracy for LCS versus the reference with R² values in the range 0.67–0.84. Hong et al. [21] developed a set of models based on LR, MLR and NLR to correct the PM_2.5 concentration from LCS in Taiwan: they found that NLR methods were superior to LR. Barkjohn et al. [37] proposed an MLR calibration model, based on PM_2.5 concentration and RH, to be used as a US-wide correction model for the Purple Air sensors. Though the LCS performance has been studied in many countries, including India [28], Korea [38], Taiwan [21] and the USA [37], the studies of performance of PM_2.5 LCS for field measurements in Thailand are limited and not well understood, especially related to locations and seasonal conditions.

This study adds new insights to understanding the key factors affecting the performance of LCS, including relative humidity, temperature and PM_2.5 concentration. The tested sensors had been operating for a long-term period, 15–16 months, at different locations in Thailand. In addition, LCS calibration based on reference concentrations, RH and temperature levels by using linear regression (LR) and multiple linear regression (MLR) models were also investigated to improve the PM_2.5 LCS performance so that it became an acceptable suppletory means for public monitoring of PM_2.5 concentration. This will help to increase public awareness on air quality, which directly affects their health.

2. Materials and Methods

2.1. The Low-Cost PM_2.5 Sensors

In the present study, the particulate matter low-cost sensors used for field calibration were made by Plantower (model PMS7003) equipped with custom-designed electric circuit and a housing as shown in Figure 1. The PMS7003 used a light-scattering technique to measure real-time PM mass concentrations [26]. The PM_2.5 measurement range was 0–1000 µg/m³, with a resolution of ±1 µg/m³ and response time 1–10 s, provided by the manufacturer [39]. The dimension of the miniature sensor used was 48 mm × 37 mm × 12 mm. The manufacturer reported that maximum errors were relatively low (±10 µg/m³ for <100 µg/m³ concentrations and ±10% for 100–500 µg/m³ concentrations). The RH and temperature for calibration were measured by thermo-hygrometer (LSI LASTEM, DMA875) at each monitoring site. The primary sensor reported data include PM₁, PM_2.5, and PM₁₀ concentrations with a factory-specified correction factor for ambient measurement concentrations (CF = atm), as recommended by the manufacturer. The description of the selected low-cost PM_2.5 sensors (LCS) in the present study was mentioned in a field evaluation study of PM_2.5 [40].

In the PMS low-cost sensor, the raw PM_2.5 mass concentration was obtained by converting particle number concentrations from different particle size bins. The effect of number-to-mass conversion on the performance of the PMS5003 was studied by Hong et al. [21]. A good correlation between the measured number concentration and the raw PM_2.5 mass concentration (R² = 0.99) was found, but a high conversion ratio was needed. This indicated that using the measured number concentrations from different size bins for calibration can provide similar results to the use of raw PM_2.5 concentrations. In this study, we used the PMS7003, which is expected to perform similarly to the PMS5003 because it has the same particle size bins [39,41]. Additionally, a field test confirmed the similar performance of the two sensors compared to the PM_2.5 FEM monitor [42]. Therefore, the effects of the measured number concentrations from different size bins on the performance of the PMS7003 can be considered negligible.

2.2. Sampling Locations and Periods

The low-cost PM_2.5 sensors were placed at three PCD monitoring stations in Chiang Mai (CM) in the north, Bangkok (BKK) in central and Ubon Ratchathani (URT) in the northeast of Thailand—see

Table 1. Sampling locations, periods, and FEM instruments at three monitoring stations.

Regions	Station Name	Locations	Instruments (U.S. EPA FEM Approved)	Measurement Technique	Duration	Dry Season	Wet Season
Northern Thailand: Chiang Mai (CM site)	Chiang Mai municipal center (PCD station #35)	Lat:18.83737, Lon:98.97132	BAM1020, MetOne, USA	Beta attenuation	1 Dec 2020–31 Mar 2022 (16 months)	Dec 2020–Apr 2021 & Nov 2021–Mar 2022 (n = 7043)	May–Oct 2021 (n = 4005)
Central Thailand: Bangkok (BKK site)	National Housing Authority Public Community Din Daeng (PCD station#54)	Lat:13.76470, Lon:100.55195	MP101M, ENVEA, France	Beta attenuation	1 Jan 2021–31 Mar 2022 (15 months)	Jan 2021–Apr 2021 & Nov 2021–Mar 2022 (n = 6022)	May–Oct 2021 (n = 4042)
Northeast Thailand: Ubon Ratchathani (URT site)	Ubon Ratchathani Provincial Administrative Organization (PCD station#83)	Lat:15.245169, Lon:104.84680	T640, Teledyne, USA	Light Scattering	1 Mar 2020–31 May 2021 (15 months)	Mar–Apr 2020 & Nov 2020–Mar 2021 (n = 5198)	May–Oct 2020 & May 2021 (n = 4671)

and Figure 2. Three identical LCS were located next to the PM_2.5 FEM instrument sampling probe on the rooftop of each PCD station (~3 m above ground). The FEM instruments were a BAM-1020 (Met One, Grants Pass, OR, USA) at the CM site, an MP101M (ENVEA, Poissy, France) at the BKK site and a T640 (Teledyne, Thousand Oaks, CA, USA) at the URT site. Key specifications of these instruments and the test periods are in Table 1: brief descriptions of the FEMs are in [21]. The FEM instrument at URT was calibrated to the FRM by the National Institute of Metrology (Thailand) [43]; the correlation between FEM and FRM was acceptable with regression equation: y = 1.084x + 4.093 (R² = 0.94). The data from FEM instruments were collected and provided by PCD. The LCS data were transmitted in real time to our server. The PM_2.5 FEM instruments also monitor meteorological parameters, including ambient temperature and relative humidity, and they were used in our LCS calibration. Weather in Thailand is influenced by two monsoon seasons: the northeast monsoon from December to March and the southwest monsoon from May to October [44] and is usually described as falling into two seasons according to the rainfall: the dry season (November to April) and wet season (May to October) [45]. The location characteristics at each sampling site in all these regions is described below.

In the northern part of Thailand, LCS were placed at Chiang Mai municipal center, Chiang Mai (PCD station#35/CM). Chiang Mai is one of the largest and most famous cities in Upper Southeast Asia (SEA): the geography consists of high mountains along the north–south corridor, covered with forest (69.7%) and agricultural areas (23.7%) [46]. Its air pollution is normally influenced by domestic and transboundary sources, particularly forest fires, as well as agriculture waste burning during the dry season [2,9]. Temperature inversions in the winter enhance accumulation of the pollutants near the ground [9]. In central Thailand, LCS were placed at the National Housing Authority Public Community, Din Daeng District, Bangkok (PCD station#54/BKK): it is a business center, where traffic is among the worst in Bangkok, from normal to slow speed vehicles. In the northeast, LCS were placed in Ubon Ratchathani (PCD station#83/URT), a large province in northeast Thailand, bordering with Cambodia. It is an agricultural-based province, where rice paddy fields cover the major portion of the land [47]. The burned area from agricultural residue in Thailand was mainly found in the northeast [12,47]. Rice and sugarcane residue burning in the northeast contributes to 50% and 40% of total agricultural waste burned in Thailand [15,48]. Open burning of the agricultural waste, as well as transboundary haze from neighboring countries, leads to a high level of air pollution there.

2.3. Sensor Calibration

The parameters used in the calibrations of the low-cost PM_2.5 sensors in the present study include the reference monitors, RH and temperature. First, hourly moving averaged PM_2.5 concentrations from the LCS and the accompanying FEM, along with RH, and temperature were collected in a single dataset for each location; the number of data records are shown in Table 1. The LR and MLR terms were developed to improve the PM_2.5 LCS—listed in

Table 2. Developed calibration models by LR and MLR to improve LCS accuracy.

#Model	Type	Equations
1	LR	${\mathrm{PM}}_{2.5\text{-}\mathrm{FEM}}=a_1+a_2\left({\mathrm{Raw}\mathrm{PM}}_{2.5\text{-}\mathrm{LCS}}\right)$
2	MLR	${\mathrm{PM}}_{2.5\text{-}\mathrm{FEM}}={\mathrm{b}}_1+{\mathrm{b}}_2\left({\mathrm{Raw}\mathrm{PM}}_{2.5\text{-}\mathrm{LCS}}\right)+{\mathrm{b}}_3\left(\mathrm{T}\right)$
3	MLR	${\mathrm{PM}}_{2.5\text{-}\mathrm{FEM}}={\mathrm{c}}_1+{\mathrm{c}}_2\left({\mathrm{Raw}\mathrm{PM}}_{2.5\text{-}\mathrm{LCS}}\right)+{\mathrm{c}}_3\left(\mathrm{R}\mathrm{H}\right)$
4	MLR	${\mathrm{PM}}_{2.5\text{-}\mathrm{FEM}}={\mathrm{d}}_1+{\mathrm{d}}_2\left({\mathrm{Raw}\mathrm{PM}}_{2.5\text{-}\mathrm{LCS}}\right)+{\mathrm{d}}_3\left(\mathrm{T}\right)+{\mathrm{d}}_4\left(\mathrm{R}\mathrm{H}\right)$
5	MLR	${\mathrm{PM}}_{2.5\text{-}\mathrm{FEM}}={\mathrm{e}}_1+{\mathrm{e}}_2\left({\mathrm{Raw}\mathrm{PM}}_{2.5\text{-}\mathrm{LCS}}\right)+{\mathrm{e}}_4\left(\mathrm{R}\mathrm{H}\right)+{\mathrm{e}}_4\left({\mathrm{Raw}\mathrm{PM}}_{2.5\text{-}\mathrm{LCS}}\times \mathrm{R}\mathrm{H}\right)$
6	MLR	${\mathrm{PM}}_{2.5\text{-}\mathrm{FEM}}={\mathrm{f}}_1+{\mathrm{f}}_2\left({\mathrm{Raw}\mathrm{PM}}_{2.5\text{-}\mathrm{LCS}}\right)+{\mathrm{f}}_3\left(\mathrm{T}\right)+{\mathrm{f}}_4\left(\mathrm{R}\mathrm{H}\right)+{\mathrm{f}}_5\left({\mathrm{Raw}\mathrm{PM}}_{2.5\text{-}\mathrm{LCS}}\times \mathrm{R}\mathrm{H}\right)$
7	MLR	${\mathrm{PM}}_{2.5\text{-}\mathrm{FEM}}={\mathrm{g}}_1+{\mathrm{g}}_2\left({\mathrm{Raw}\mathrm{PM}}_{2.5\text{-}\mathrm{LCS}}\right)+{\mathrm{g}}_3\left(\mathrm{T}\right)+{\mathrm{g}}_4\left({\mathrm{Raw}\mathrm{PM}}_{2.5\text{-}\mathrm{LCS}}\times \mathrm{T}\right)$
8	MLR	${\mathrm{PM}}_{2.5\text{-}\mathrm{FEM}}={\mathrm{h}}_1+{\mathrm{h}}_2\left({\mathrm{Raw}\mathrm{PM}}_{2.5\text{-}\mathrm{LCS}}\right)+{\mathrm{h}}_3\left(\mathrm{T}\right)+{{\mathrm{h}}_4\left(\mathrm{R}\mathrm{H}\right)+\mathrm{h}}_5\left({\mathrm{Raw}\mathrm{PM}}_{2.5\text{-}\mathrm{LCS}}\times \mathrm{T}\right)$
9	MLR	${\mathrm{PM}}_{2.5\text{-}\mathrm{FEM}}={\mathrm{i}}_1+{\mathrm{i}}_2\left({\mathrm{Raw}\mathrm{PM}}_{2.5\text{-}\mathrm{LCS}}\right)+{\mathrm{i}}_3\left(\mathrm{T}\right)+{\mathrm{i}}_4\left(\mathrm{R}\mathrm{H}\right)+{\mathrm{i}}_5\left({\mathrm{Raw}\mathrm{PM}}_{2.5\text{-}\mathrm{LCS}}\times \mathrm{T})+{\mathrm{i}}_6({\mathrm{Raw}\mathrm{PM}}_{2.5\text{-}\mathrm{LCS}}\times \mathrm{T}\times \mathrm{R}\mathrm{H}\right)$

${\mathrm{PM}}_{2.5\text{-}\mathrm{FEM}}$ = PM_2.5 concentration from FEM; ${\mathrm{Raw}\mathrm{PM}}_{2.5\text{-}\mathrm{LCS}}$ = PM_2.5 concentration from LCS; T = Temperature and RH = Relative humidity.

. In those equations, ${\mathrm{PM}}_{2.5\text{-}\mathrm{FEM}}$ represents is the hourly PM_2.5 concentration of the FEM measurement, ${\mathrm{Raw}\mathrm{PM}}_{2.5\text{-}\mathrm{LCS}}$ represents the PM_2.5 concentration provided by the LCS and $a_j$ to $i_j$ are the fitted model coefficients and intercepts. Model#1 is the simple LR model based on only the PM_2.5 concentration. Model#2–Model#4 are the MLR models with addition of T, RH and both T and RH. MLR models with additional product terms are Model#5 (add RH and RH and PM_2.5 product), Model#6 (add T and RH and product of RH and PM_2.5), Model#7 (add T and product of T and PM_2.5), Model#8 (add T and RH and product of T and PM_2.5) and Model#9 (addition of T and RH and multiplication of T, RH and PM_2.5). We used the linear RH model because it had better performance than the nonlinear RH model [37]. However, we also included the nonlinear RH term, RH²/(1−RH), in the model for comparison. Details of the calculation were described by [28].

2.4. Performance Metrics

The performance of the LCS and FEM at each site were examined based on the U.S. EPA guidelines [43], including precision: coefficient of variation, CV ≤ 30%, bias: slope (1.0 ± 0.35) and intercept of 0 ± 5 µg/m³, linearity: coefficient of determination, R² ≥ 0.7 and error: root mean square error, RMSE ≤ 7 µg/m³. A linear regression for LCS measurements compared to the reference instrument calculated the slope, intercept and R² to represent the LCS accuracy.

Precision between identical sensors was characterized by coefficient of variation, calculated every hour by:

$$\mathrm{C}\mathrm{V}=\frac{\mathrm{S}\mathrm{D}}{\overline{x}}\times 100$$

The root means square error, RMSE, was used to help understand the error associated with sensors for PM_2.5 concentration measurements:

$$\mathrm{RMSE}=\sqrt{\frac{1}{\mathrm{N}\times \mathrm{M}}\sum _{\mathrm{j}=1}^{\mathrm{M}}\left[\sum _{\mathrm{t}=1}^{\mathrm{N}}{({\mathrm{x}}_{\mathrm{t},\mathrm{j}}-{\overline{\mathrm{y}}}_{\mathrm{t}})}^2\right]}$$

where SD is the standard deviation of averaged PM_2.5 LCS concentration, $\overline{\mathrm{x}}$ is the mean PM_2.5 LCS concentration (µg/m³), $\mathrm{N}$ is the number of samples, M is number of identical LCS, ${\mathrm{x}}_{\mathrm{t},\mathrm{j}}$ is the PM_2.5 LCS concentration of sample #$\mathrm{j}$ at time $\mathrm{t}$ (µg/m³) and ${\overline{\mathrm{y}}}_{\mathrm{t}}$ is the PM_2.5 FEM concentration at the same time (µg/m³).

Akaike’s information criterion (AIC) was used to determine the significance of the RH and temperature term in the MRL models [28]. The AIC values were used to select the best model and avoid parameter overfitting. A lower value between any comparative pair of models was preferred. The AIC was calculated from:

$$\mathrm{AIC}=\left[\mathrm{N}\times \ln (\frac{\mathrm{Residual}\mathrm{Sum}\mathrm{of}\mathrm{Squares}}{\mathrm{N}})\right]+(2\times \mathrm{k})$$

where $k$ is the number of parameters.

3. Results and Discussion

3.1. Source Characteristics of PM_2.5 in Different Areas

The overall-averaged PM_2.5 concentrations from FEM instruments, at all three sites, obtained from PCD during the year 2020–2022 are listed in

Table 3. Summaries of average PM_2.5 FEM concentrations, ratios of LCS to FEM, RH and temperature (±standard deviation).

Sites		Dry Season					Wet Season
	PM2.5 LCS (µg/m3)	PM2.5 FEM (µg/m3)	Ratio LCS/FEM	Temperature (°C)	RH (%)	PM2.5 LCS (µg/m3)	PM2.5 FEM (µg/m3)	Ratio LCS/FEM	Temperature (°C)	RH (%)
CM	41.3 ± 29.3	38.3±23.0	1.1 ± 0.3	24.4 ± 5.4	61.1 ± 20.3	10.5 ± 7.5	16.1 ± 5.5	0.6 ± 0.3	27.5 ± 3.3	74.5 ± 16.0
BKK	26.5 ± 17.5	40.5 ± 18.4	0.7 ± 0.3	29.0 ± 2.7	66.4 ± 11.9	14.0 ± 10.8	23.6 ± 9.4	0.6 ± 0.4	29.9 ± 2.9	72.0 ± 12.3
URT	39.2 ± 27.3	43.4 ± 37.3	0.9 ± 0.3	26.9 ± 5.4	59.9 ± 14.9	18.1 ± 15.5	15.4 ± 10.6	1.2 ± 0.5	28.1 ± 3.8	74.6 ± 15.1

. In general, the ambient conditions at all stations across Thailand were quite similar. The PM_2.5 concentrations and characteristics could be different from season to season and site to site.

Hourly PM_2.5 mass concentration from FEM measurement at CM during the dry season was as high as 246 µg/m³_. The average concentration during the dry season was 38.3 ± 23.0, about two times as high as that during the wet season (16.0 ± 5.5 µg/m³). Northern Thailand is typically influenced by transboundary and domestic aerosols. Domestic sources of ambient PM in northern Thailand were mostly forest fires (>90% of total burned area in northern Thailand), especially during the dry season. They occurred mostly in deciduous dipterocarp forest national parks [9]. Aerosol transport from open biomass burning area in some parts of Myanmar and Laos also affected air quality in northern Thailand [1,3,9]. Moreover, temperature inversion plays an important role, leading to high accumulation of air pollutants during the winter months (November to March) in northern Thailand [2].

In BKK, the average PM_2.5 concentration during the dry season was also approximately two times as high as that during the wet season. Average PM_2.5 concentrations during the wet season were about 1.5 times higher than those at CM and URT. This is because the local emission sources, especially vehicle emissions, were uniformly high due to traffic congestion [8]. As well as vehicle and industrial emissions, local biomass burning and long-range transport from other countries also contributed to the PM_2.5 concentrations during the dry season which were 2–4 times as high as those in background [6,7]. Hence, fine particles in Bangkok came from diverse sources, including vehicle emissions, biomass burning, industrial, construction activities and secondary aerosols [6].

PM levels at URT were significantly enhanced by agricultural waste burning, especially rice and sugarcane, during the dry season. The northeastern region is the largest agricultural area (~10 M ha or 42% of the total). Most agricultural waste was burned in the open, for land clearing and preparing for the next crop, during January to April [15]. PM_2.5 concentrations during the dry season (as high as 322 µg/m³) were higher than those in CM and BKK.

3.2. Time Series of PM_2.5 Concentrations, Ratio of LCS and FEM, RH and Temperature

Figure 3 shows the seasonal variation of 1-h PM_2.5 concentration from LCS and FEM, ratio of LCS to FEM, RH and temperature at all three sites. Overall, the uncalibrated LCS measurements mostly followed the trend in PM_2.5 FEM measurements and responded well to most sudden spikes of mass concentrations. LCS to FEM ratios and meteorological factors (RH and temperature) are also summarized in Table 3. The LCS/FEM ratio during the dry season at CM and URT were close to 1, whereas it was lower at BKK. This indicated that the 1-h averaged ambient PM_2.5 levels of uncalibrated LCS at CM and URT matched well, in general, to those obtained from the FEM instruments. However, PM_2.5 LCS concentrations at URT during the dry season were lower than those from the T640. This agrees with Hagler et al. [49], who reported that the PM_2.5 concentrations from a T640, an optically based FEM, closely matched those from a BAM-1020 instrument at lower concentrations, but overestimated concentrations above ~80 µg/m³. Average RH and temperature during the dry season ranged from 60 ± 15 to 66 ± 12% and 24.4 ± 5.4 to 29.0 ± 2.7 °C, respectively. Average RH during the wet season was significantly higher than those during the dry season (RH increased 9–24%), while temperature was also slightly higher. Average temperatures at CM and URT in the dry season were lower, because it included the winter months (Nov–Feb).

However, the LCS measurements also followed the FEM instruments during the wet season, except at BKK. The ratio of LCS/FEM using the beta ray technique, during the wet season at CM and BKK, were strongly lower than 1. Zheng et al. [28] showed that the LCS performance decreased with low PM_2.5 concentration and high RH. This indicated that the LCS were highly sensitive to RH but in varying directions [50]. At BKK, the ratios of LCS/FEM were lower than 1 during both periods. This could be because the LCS at this site were located close to major roads in a highly congested area. Most of the particles from fresh traffic emissions were smaller than the LCS minimum detectable size (see detail in Section 3.5). This is consistent with Castell et al. [30], who reported that the LCS measurements, where traffic emission was low, were better correlated with the reference measurements. In contrast, the mean ratio of LCS/FEM during the wet season at URT was higher than 1 (1.2 ± 0.5). The ratio of LCS to FEM higher than 1.2 during the wet season, accounting for 53%, was found at high RH level (79 ± 12%). Moreover, we found that increased LCS errors (>1.7 ± 0.5) mostly occurred at the extreme RH level (84 ± 10%). This agrees with Zamora et al. (2019), who reported that the accuracy of LCS (Plantower PMSA003) decreased for RH > 50% [51,52]. Similarly, Jayarathne et al. [33] noted that the error of LCS increased for RH > 80%, in field monitoring. For the temperature effect, the temperature differential between dry and wet seasons at CM was as high as ~3 °C, while at BKK and URT it was closer to 1 °C—an insignificant change. Some reports have claimed that neither temperature or RH affected LCS errors [53,54]. However, others have noted that calibration to correct a temperature effect improved LCS, and our results (see later) show that temperature correction did improve accuracy [54].

3.3. Influence of Concentration Range, RH and Temperature Levels on Sensor Performance

The distribution of the relative error of unadjusted LCS are shown in Figure 4. The relative error (%) values were calculated with 1-h averaged data of unadjusted PM_2.5 LCS and FEM concentrations. For the concentration range, the highest relative errors for PM_2.5 concentrations were observed below 20 µg/m³, while the maximum error values were up to 200%. For the CM and URT sites, between 20–50 μg/m³, the errors were mostly below 100% and they were mostly below 50% when the PM_2.5 level went above 50 μg/m³. Values from LCS at BKK were mostly characterized by rather constant maximum relative errors at ~90% independent of the concentration range, RH or T (average: 40.2% for dry and 51.3% for wet). The highest relative errors were found during the wet season at all sites at low concentration (<20 µg/m³), with average PM_2.5 concentrations of 16.1 ± 5.5 µg/m³ (CM), 23.6 ± 9.4 µg/m³ (BKK) and 15.4 ± 10.6 µg/m³ (URT). At CM, the relative error during the dry season was clearly increased at a higher RH level. This is consistent with the previous studies by Liu et al. [51] and Jayaratne et al. [33]. The relative error of LCS during the wet season doubled those in dry season at CM and URT. The error during the wet season at CM and BKK were nearly identical, with constant maximum ~90% at RH greater than 42%. The relative error of LSC at URT (average: 23.8% in dry season and 40.0% in wet season) was consistent with most previous studies, which showed that low PM concentration and RH level significantly impacted low-cost PM sensor performance [20,55,56,57]. However, it should be noted that Kosmopoulos et al. [29] reported that the effect of RH on the response of Plantower PMS5003 sensors for fine PM in southeast Europe was negligible, so the measurement location is important to consider. The relative errors of PM_2.5 in the present study show no pattern for the influence of temperature levels on sensor performance. It is noted that beta attenuation monitor consists of the heater system with a controller to reduce the RH level of samples. It was normally set below a particular point to remove the hygroscopic and condensation effect of particles on the filter influencing the mass concentration reading. Hence, a range of RH set point is the critical factor to assess the performance of beta attenuation measurement [58,59]. Takahashi et al. [59] reported the PM concentration from BAM was higher than those using the gravimetric method at high RH (>80%). Kiss et al. [60] suggested that 1-h PM₁₀ measurements must be used carefully, as water vapor adsorption/desorption from the beta attenuation monitor’s filter material may cause significant bias at low ambient concentration. Zheng et al. [28] also supported that the beta-attenuation-based monitors have drawbacks at low concentrations.

3.4. Responses of the LCS to Emission Sources

The response of the PM_2.5 LCS to the different emission sources was interpreted in the present study (see Figure 5). The different R² of linear regression results between PM_2.5 concentrations from the LCS and the FEM instruments in locations where major aerosol sources were different indicated that the LCS had variable responses to types of aerosols. At BKK site, the LCS were located within a short distance from major roads, and the traffic emissions were mainly constituted of fresh vehicle exhaust as discussed in Section 3.1. The results showed that the R² values of the linear regressions between LCS and FEM at BKK during the dry season (R² = 0.561) were quite low, and very poor during the wet season (R² = 0.291), in response to fresh vehicle exhaust emissions. However, the PM_2.5 characteristics at BKK during the dry season were annually affected by domestic and transboundary aerosol transport from open biomass burning [6]. The R² value of the linear regression between the LCS and the FEM was higher because of a large contribution from biomass burning. For CM and URT, R² values of the regressions were high during the dry season in response to predominant open biomass burning. Not only does the emission source play a significant role, but ambient RH levels also had effects on the LCS performance. The LCS were more responsive to open biomass burning emissions than to fresh traffic emissions. This is because the particle size from diesel and gasoline engines were dominant in the range of 60–120 nm and 40–80 nm, respectively [61]. Hence, most of the particles from fresh traffic emissions were ultrafine particles (<0.1 µm), which are smaller than the minimum detectable size (0.3 µm) of the LCS. Moreover, the particle size from biomass burning were predominantly in the range of 0.5–2.5 µm [62]. Hata et al. [63] also reported that the dominant size of PM from biomass burning was in accumulation mode (PM_0.5–1.0). This is why the linear regression between the PM_2.5 LCS and FEM concentration at CM and URT were better than that at BKK, especially in dry season. This was confirmed by Sresawasd et al. [9], who reported the PM mass concentration was distributed mostly in the accumulation mode PM_0.5–1 (38–48%) of TSP in Chiang Mai, Thailand during the haze period (dry season). Additionally, Dejchanchaiwong et al. [6] found that the PM_0.5–1.0 concentrations in the haze period (Dec–Feb) were 2 to 12 times as high as those in the nonhaze period. However, underestimation of PM_2.5 concentration may partly be a result from the limitation of the Plantower sensor. Kuula et al. [64] found that the large particle size fraction 1.0–2.5 µm significantly reduce the concentration of PM_2.5 when using Plantower PMS5003. The finding of the LCS responses to different aerosol emission sources provided guidance on future application of LCS monitors, such as the distance of the deployment sites to major roads.

3.5. Performance Metrics for Unadjusted PM_2.5 Concentration

The correlation between 1-h PM_2.5 concentration from unadjusted (original) LCS and FEM measurements during the dry and wet seasons are shown in Figure 5. The values of performance metric including precision (CV), bias (slope and intercept), linearity (R²) and error (RMSE) for unadjusted LCS at each site are shown in

Table 4. Summary of performance metrics for hourly PM_2.5 concentrations for unadjusted LCS and adjusted LCS by LR and MLR calibration models at three sites.

Sites	Periods	Models	CV (%)	Slope (-)	Intercept (µg/m3)	R2 (-)	RMSE (µg/m3)
CM	Dry	Unadjusted	14.47 ± 6.80	1.215	−5.184	0.907	11.942
		LR	8.46 ± 5.99	0.907	3.553	0.907	7.131
		MLR	6.84 ± 4.97	0.945	2.414	0.934	5.102
	Wet	Unadjusted	21.53 ± 9.58	1.111	−7.73	0.635	7.753
		LR	13.76 ± 6.79	0.639	5.549	0.635	5.393
		MLR	11.69 ± 5.38	0.681	4.813	0.672	3.304
BKK	Dry	Unadjusted	9.78 ± 5.77	0.714	−2.398	0.561	17.646
		LR	7.32 ± 4.29	0.561	17.772	0.561	13.971
		MLR	5.49 ± 3.24	0.701	4.657	0.671	11.156
	Wet	Unadjusted	5.85 ± 3.86	0.619	−0.543	0.291	13.707
		LR	4.32 ± 2.46	0.291	16.709	0.291	10.918
		MLR	2.77 ± 1.43	0.322	15.961	0.322	7.773
URT	Dry	Unadjusted	8.86 ± 7.30	0.679	9.733	0.856	16.486
		LR	6.74 ± 5.47	0.855	6.273	0.856	8.383
		MLR	5.76 ± 4.67	0.899	5.492	0.880	4.285
	Wet	Unadjusted	8.77 ± 8.14	0.842	5.087	0.815	7.692
		LR	7.41 ± 6.95	0.815	2.864	0.815	6.711
		MLR	6.33 ± 5.94	0.829	1.123	0.893	5.378

Note: Performance metrics of MLR model show only the best fitting model with highest R² and lowest AIC values (See Section 3.7).

. Reproducibility of individual LCS was assessed on the basis of CV. The mean CV for unadjusted LCS at each site during the dry (8.86 ± 7.30% to 14.47 ± 6.80%) and wet season (5.85 ± 3.86 to 12.53 ± 9.58) were at an acceptable level (CV ≤ 30%) of sensor variations. The R² values from correlation between the unadjusted-LCS and the reference PM_2.5 were acceptable at CM during the dry season with R² = 0.907; however, during the wet season, it was poorer, with R² = 0.635, due to larger scattering in the PM_2.5 concentration at high RH level. The R² values of LCS vs. FEM at URT during the dry (R² = 0.856) and wet seasons (R² = 0.815) were also acceptable with R² > 0.7, as suggested by U.S. EPA [43]. However, at BKK, the R² was only 0.561 during the dry season and 0.291 during the wet season. The RMSE values of unadjusted LCS vs. FEM at all sites were also higher than the values suggested by U.S. EPA [43]. Hence, the performance of PM_2.5 LCS in Thailand was poor to high (R² = 0.291–0.907) as compared with FEM instruments. This is consistent with the study of Johnson et al. [27], who reported the performance of LCS with BAM was only moderate to high (R² = 0.35–0.81). The low performance of LCS could be attributed to the different PM emission sources, low concentration and high RH level. Before calibration, the performance of the LCS only at URT during the wet season was acceptable.

3.6. Calibration by LR Model and Performance Metrics for Adjusted LCS Based on Reference Concentration

The LR (Model 1 in Table 2) was used to calibrate 1-h PM_2.5 concentrations from LCS against the reference during the dry and wet seasons. Figure 6 shows the correlation between the adjusted PM_2.5 LCS by the LR model and FEM measurements at all sites for each season. The performances of the LR models are also in Table 4. After calibration by LR model, the results were improved as compared to the unadjusted PM_2.5 LCS with the reduction of CV values of 42% (dry) and 46% (wet) at CM, 25% (dry) and 26% (wet) at BKK and 24% (dry) and 16% (wet) at URT. The low CV value (below 30%) was acceptable for LCS testing studies, indicating high reproducibility [43]. All CV values were well within this limit. This is consistent with Kelly et al. [26], who reported the good repeatability between Plantower sensors. Additionally, the results were significantly improved as compared to unadjusted PM_2.5 from LCS with the reduced RSME values of 40.29% (dry) and 30.44% (wet) at CM, 20.83% (dry) and 20.35% (wet) at BKK and 49.15% (dry) and 12.75% (wet) at URT. However, the R² values were not improved at all sites. This agreed with Hong et al. [21], who interpreted that the datasets were shifted to the 1:1 line without data correlation improvement.

3.7. Calibration Using MLR Models to Adjust LCS Based on Reference Concentration, RH and Temperature Levels

Performance of the linear RH model (model #3) in the present study was compared with the nonlinear RH model terms [28]. The R² for model #3 was slightly higher than the nonlinear model, as shown in

Table 5. Correlation coefficients for linear and nonlinear RH models.

Seasons	CM		BKK		URT
	Nonlinear RH	Linear RH (Model #3)	Nonlinear RH	Linear RH (Model #3)	Nonlinear RH	Linear RH (Model #3)
Dry	0.90 ± 0.01	0.926	0.56 ± 0.02	0.568	0.88 ± 0.01	0.873
Wet	0.63 ± 0.02	0.664	0.29 ± 0.01	0.292	0.83 ± 0.01	0.846

. Based on this analysis, we used the linear RH model to adjust the models further.

Eight MLR models (Model#2–9 in Table 2) were used to calibrate LCS PM_2.5 concentrations from LCS—see Figure A1, Figure A2 and Figure A3 in Appendix A. The correlations of the adjusted PM_2.5 concentration from LCS were achieved after correction by using the MLR models based on reference concentration, RH and temperature. The best-fitted model was chosen based on lowest AIC and highest R² (see Figure A4). The results showed that Model#6 is best for CM (wet season), BKK (wet season) and URT (dry season), while Model#9 is best for the remaining periods at each site (See Figure 7). The performance metrics of MLR model were examined only for the best fitted model at each site, as also shown in Table 4. After MLR calibration, the LCS performances were improved as compared to adjusted PM_2.5 from LCS by LR model with the R² increase of 2.98% (dry) and 5.83% (wet) at CM, 19.61% (dry) and 10.65% (wet) at BKK and 2.80% (dry) and 9.57% (wet) at URT. Additionally, the results were significantly improved as compared to the adjusted PM_2.5 by LR model with the RSME reduction of 28.45% (dry) and 38.74% (wet) at CM, 20.15% (dry) and 28.81% (wet) at BKK and 48.88% (dry) and 19.86% (wet) at URT, and CV reduction of 19.15% (dry) and 15.04% (wet) at CM, 25.00% (dry) and 35.88% (wet) at BKK and 14.54% (dry) and 14.57% (wet) at URT.

The performance metrics of the LCS and FEM only at CM during the dry season and URT during the dry and wet season were acceptable with the CV: 5.76 ± 4.67–6.84 ± 4.97% (CV ≤ 30%), slope: 0.829–0.945 (slope: 1.0 ± 0.35) and intercept 1.123–5.492 µg/m³ (0 ± 5 µg/m³), R²: 0.880–0.934 (R² ≥ 0.7) and RMSE: 4.285–5.102 µg/m³ (RMSE ≤ 7 µg/m³). The constants parameters of the best fitting calibration model at CM during the dry season and URT during both seasons are shown in

Table 6. MLR calibration: constants for the best fitting models.

Method	Best Fitting Model	Constants Parameters	CM	URT
			Dry	Dry	Wet
MLR	Model 6	f1 (Intercept)	-	5.327	-
		$\mathrm{f}{2}_{}\left({\mathrm{Raw}\mathrm{PM}}_{2.5\text{-}\mathrm{LCS}}\right)$	-	1.853	-
		$\mathrm{f}{3}_{}\left(\mathrm{T}\right)$	-	−0.294	-
		$\mathrm{f}{4}_{}\left(\mathrm{R}\mathrm{H}\right)$	-	−0.076	-
		$\mathrm{f}{5}_{}\left({\mathrm{Raw}\mathrm{PM}}_{2.5\text{-}\mathrm{LCS}}\times \mathrm{R}\mathrm{H}\right)$	-	−0.009	-
	Model 9	i1 (Intercept)	27.192	-	78.664
		$\mathrm{i}{2}_{}\left({\mathrm{Raw}\mathrm{PM}}_{2.5\text{-}\mathrm{LCS}}\right)$	0.352	-	−0.955
		$\mathrm{i}{3}_{}\left(\mathrm{T}\right)$	−0.397	-	−1.847
		$\mathrm{i}{4}_{}\left(\mathrm{R}\mathrm{H}\right)$	−0.15	-	−0.367
		$\mathrm{i}{5}_{}{(\mathrm{Raw}\mathrm{PM}}_{2.5\text{-}\mathrm{LCS}}\times \mathrm{T})$	0.013	-	0.063
		$\mathrm{i}{6}_{}({\mathrm{Raw}\mathrm{PM}}_{2.5\text{-}\mathrm{LCS}}\times \mathrm{T}\times \mathrm{R}\mathrm{H})$	4.34 × 10−5	-	4.264 × 10−5

. However, the correlations of the adjusted PM_2.5 from LCS at BKK during both periods were not entirely consistent with the FEM values and the adjusted PM_2.5 performances were not acceptable. This discrepancy might occur from variations in aerosol compositions and impacts of particle size from emission sources as described in the previous section. Moreover, The MLR model at CM during the wet season was also not suggested to use for calibration. The results showed that the FRM calibration plays more important role to the accuracy of FEM. At URT, the results showed that the LCS at URT can be calibrated during both seasons with the FEM instrument.

4. Conclusions

The effects of key factors—emission source, relative humidity, temperature and PM_2.5 concentration range—on the performance of PM_2.5-LCS for long periods at three different locations in Thailand during both seasons were investigated. Additionally, the calibration for LCS based on reference concentration, RH and temperature levels using LR and MLR models were also investigated to improve the performance of PM_2.5-LCS in Thailand. We concluded:

• At PM concentration < 20 µg/m³ and RH > 85%, PM_2.5-LCS performance was significantly influenced.
• Location of the PM_2.5-LCS was crucial to performance: a high traffic emission area (BKK) showed low correlation with reference monitors due to the effect of small particles.
• Unadjusted PM_2.5-LCS performance varied with location, showing low to high correlations with FEM instruments (0.29 < R² < 0.91).
• Performances of the adjusted PM_2.5-LCS at BKK in both seasons and at CM during the wet season were not acceptable due to very small particle size from emission sources, and effect of low concentrations and RH level.
• After MRL calibration, performances of the PM_2.5-LCS only at CM during the dry season and URT site during both seasons were acceptable with the CV: 5.76 ± 4.67%–6.84 ± 4.97%, slope: 0.829–0.945, intercept: 1.123–5.492 µg/m³, R²: 0.880–0.934 and RMSE: 4.285–5.102 µg/m³.

The present study showed the importance of evaluating the long-term field performance of PM_2.5-LCS at different ambient conditions and locations. Calibration and validation of the PM_2.5 LCS with FEM instruments in the field test is crucial to high-quality data. Moreover, understanding the key factors, including emission sources, relative humidity, temperature and PM_2.5 concentration, on PM_2.5-LCS performance provides crucial information for setting an achievable and reasonable guideline to enhance the quality of PM_2.5-LCS data. Since the calibration of the sensors used the full dataset to develop and test the models at each station, the calibration results may differ slightly for shorter collocation periods when different dataset are used. In the future, the PM_2.5-LCS performance with higher performance in low concentrations and high humidity environment are recommended.