Estimating ground-level PM2.5 over Bangkok Metropolitan Region in Thailand using aerosol optical depth retrieved by MODIS.

A number of previous studies have shown that statistical model with a combination of satellite-derived aerosol optical depth (AOD) and PM2.5 measured by the monitoring stations could be applied to predict spatial ground-level PM2.5 concentration, but few studies have been conducted in Thailand. This study aimed to estimate ground-level PM2.5 over the Bangkok Metropolitan Region in 2020 using linear regression model that incorporates the Moderate Resolution Imaging Spectroradiometer (MODIS) AOD measurements and other air pollutants, as well as various meteorological factors and greenness indicators into the model. The 12-fold cross-validation technique was used to examine the accuracy of model performance. The annual mean (standard deviation) concentration of observed PM2.5 was 22.37 (± 12.55) µg/m3 and the mean (standard deviation) of PM2.5 during summer, winter, and rainy season was 18.36 (± 7.14) µg/m3, 33.60 (± 14.48) µg/m3, and 15.30 (± 4.78) µg/m3, respectively. The cross-validation yielded R 2 of 0.48, 0.55, 0.21, and 0.52 with the average of predicted PM2.5 concentration of 22.25 (± 9.97) µg/m3, 21.68 (± 9.14) µg/m3, 29.43 (± 9.45) µg/m3, and 15.74 (± 5.68) µg/m3 for the year round, summer, winter, and rainy season, respectively. We also observed that integrating NO2 and O3 into the regression model improved the prediction accuracy significantly for a year round, summer, winter, and rainy season over the Bangkok Metropolitan Region. In conclusion, estimating ground-level PM2.5 concentration from the MODIS AOD measurement using linear regression model provided the satisfactory model performance when incorporating many possible predictor variables that would affect the association between MODIS AOD and PM2.5 concentration. Supplementary Information: The online version contains supplementary material available at 10.1007/s11869-022-01238-4.

Introduction

A number of previous epidemiological studies have revealed that exposure to particulate matter with aerodynamic diameter less than or equal to 2.5 µm (PM2.5), both short-term and long-term exposure, was associated with an increased risk of morbidity and mortality for particularly cardiovascular and respiratory diseases (Alexeeff et al. 2021; Atkinson et al. 2014; Fan et al. 2016; Farhadi et al. 2020; Sangkharat et al. 2019; Thongphunchung et al. 2021; Wang et al. 2020). Over the recent years, PM2.5 has been considered one of the most serious environmental and public health problems in Thailand, especially in Bangkok Metropolitan Region (BMR) (i.e., Bangkok and its five surrounding provinces, including Nakhon Pathom, Nonthaburi, Pathum Thani, Samut Prakan, and Samut Sakhon), which is the most densely populated region of Thailand. Moreover, the rapid economic development with many commercial activities in BMR has exacerbated in favors of high PM2.5 emissions, leading to increase ambient PM2.5 concentrations where daily average concentration has sometimes higher than the 24-h ambient standard of Thailand, especially during winter (i.e., November to February) (Thongphunchung et al. 2021). Because monitoring ground-level PM2.5 concentration in Thailand, including in the BMR, has been operated by the Pollution Control Department (PCD) with sparsely distributed numbers of monitoring stations across the BMR, the concentrations of PM2.5 in unmonitored locations have not been captured. Therefore, there is necessary to develop the possibly alternative techniques to estimate the concentration of PM2.5 in the areas with no monitoring station in the BMR that can be further applied for informing PM2.5 situation in the region.

Satellite remote sensing with different algorithms is one of the powerful tools used to estimate ground-level PM2.5. Specifically, the Deep Blue (DB) algorithm has been developed to calculate aerosol optical depth (AOD), which is a parameter indicating the extinction of light due to the suspension of aerosols in the atmosphere. The AOD data is generally captured by the instruments aboard the satellites, such as the Moderate Resolution Imaging Spectroradiometer (MODIS) onboard the Terra and Aqua satellites. This data has been extensively applied along with different predictor variables (i.e., meteorological factors, greenness indicators, and planetary boundary layer) to estimate PM2.5 concentration at different spatial and temporal scales (He et al. 2020; Kong et al. 2016; Lv et al. 2017; Zheng et al. 2016). The linear relationship between AOD and PM2.5 has been considered in earlier previous studies taken into account many spatiotemporal predictors that might be affecting the number of atmospheric aerosols (Liu et al. 2005; Ma et al. 2014). In recent years, more advanced statistical models have been utilized to describe the complex relationship between AOD and ground-level PM2.5 concentration such as linear mixed-effects model (Wang et al. 2019; Xie et al. 2015; Zheng et al. 2016), the Bayesian-based downscaling framework (Lv et al. 2017), and linear regression with geographically and temporally weighted regression (GTWR) model (Chen et al. 2020; He et al. 2020).

The different MODIS AOD data products have been used to estimate ground-level PM2.5 concentration in many previous studies. In particular, a spatial resolution of 10 km MODIS AOD has long been used to derive ambient concentration of PM2.5 in the regional scales (Kong et al. 2016; Ma et al. 2014), but a 3-km MODIS AOD have been applied recently to gain insight into finer details of spatial information in urban levels contributing to population-based PM2.5 exposure and its health effect studies (Lv et al. 2017; Xie et al. 2015). Although satellite-derived AOD products have widely used to estimate surface PM2.5 concentration in many parts of the world, especially in mainland China, with various statistical models (Hu et al. 2014; Liu et al. 2005; Ma et al. 2014; Wang et al. 2019; Xie et al. 2015), very few studies have been conducted in Thailand where there is a limited number of air quality monitoring stations. Therefore, ambient concentration of PM2.5 in an area with no monitoring station could not be captured; meanwhile, using PM2.5 concentration from a fixed site monitoring station representing the same exposure level for all residents throughout the region may bias the results of epidemiological studies. Hence, this study aimed to estimate daily PM2.5 concentration in the BMR, the most polluted region of Thailand, using 3-km resolution MODIS AOD dataset from January to December 2020 for exposure estimation of epidemiological studies. The multiple-linear regression with various temporally meteorological variables and vegetation index was established to examine the relationship between satellite-derived AOD and PM2.5 concentration. The accuracy of estimated PM2.5 concentration was then explored using cross-validation technique.

Methods

The Bangkok Metropolitan Region (BMR) is consisting of Bangkok Metropolitan Area and its five surrounding provinces (i.e., Nakhon Pathom, Nonthaburi, Pathum Thani, Samut Prakan, and Samut Sakhon). This region is located in the central part of Thailand with 10.9 million populations in 2020, representing 16.5% of Thailand’s total populations (Thailand National Statistical Office 2020). The BMR has been considered the most economic development area in Thailand that often experiences heavy concentration of PM2.5 generated mainly by vehicle emission, biomass burning, and industrial activities, respectively (Narita et al. 2019). The BMR covers an area of approximately 8000 km2, and 23 air quality monitoring stations have been operated by the Pollution Control Department throughout the BMR as shown in Fig. 1.

Fig. 1

Study location of the BMR, consisting of Bangkok and its five adjacent provinces. The black dots illustrate air quality monitoring stations operated by the Pollution Control Department that are situated in the BMR 

MODIS AOD data product

The MODIS AOD data was obtained from the Atmosphere Archive and Distribution System (LAADS) Distributed Active Archive Center (DAAC) (https://ladsweb.modaps.eosdis.nasa.gov/search/order/1), which is one of the DAACs under the National Aeronautics and Space Administration (NASA)’s Earth Observing System (EOS). In particular, the AOD data in this study was acquired from the Moderate Resolution Imaging Spectroradiometer (MODIS) instrument aboard the Terra and Aqua satellites Collection 6.1, Level 2 Deep Blue aerosol product (Optical_Depth_Land_And_Ocean at the 550 nm wavelength) with a resolution of 3 km × 3 km (MODO4_3K and MYD04_3K) during January through December 2020 and then resampled the data into fixed grid that covers spatial resolution of 1 km × 1 km using bilinear interpolation technique. The daily average AOD value was calculated by taking the average of daily AOD data from the Terra and Aqua satellites, overpassing the study area respectively at approximately 10:30 and 13:30 local time, because a previous study revealed that the MODIS AOD data from Terra and Aqua satellites may provide good estimates of daily AOD values (Zhang et al. 2012). The AOD values were then averaged throughout the BMR by date during the study period, and the missing AOD data was imputed using the expectation–maximization (EM) algorithm under the assumptions of a multivariate normal distribution (Junger and Ponce de Leon 2015).

Ground-level PM2.5 and meteorological data

Hourly concentration of air pollution (i.e., PM2.5 (µg/m3), NO2 (ppb), and O3 (ppb)) and meteorological variables (i.e., ambient temperature (°C), relative humidity (%), and wind speed (m/s)) were obtained from the air quality monitoring stations operated by the Pollution Control Department of Thailand that situated across the BMR as presented in Fig. 1 during January through December 2020. Because MODIS AOD data are monitored at a single point in time when the satellites are overpassing the study area (i.e., 10:30 and 13:30 local times for Terra and Aqua satellites, respectively), the 7-h average concentrations of every single air pollutant and meteorological variable from 9:00 through 15:00 local times each day were computed that covers the time for Terra and Aqua satellites overpassing the study area. In the case that there were missing data more than 2 h out of 7 h, data of that particular day was assigned as missing. To obtain daily average concentrations of individual air pollutants and meteorological variables, the 7-h average concentrations obtained from all monitoring stations located in the BMR were then averaged. The missing air pollutant and meteorological data was imputed using the EM algorithm (Junger and Ponce de Leon 2015).

Greenness indicator data

The 16-day interval greenness data during January through December 2020 was acquired from the Moderate Resolution Imaging Spectroradiometer (MODIS) aboard the NASA’s Terra satellite with a spatial resolution of 0.1° × 0.1° (https://neo.gsfc.nasa.gov/view.php?datasetId=MOD_NDVI_M). The satellite-based greenness data have been measured as the Normalized Difference Vegetation Index (NDVI), which is calculated from the ratio of the difference between reflected near-infrared light and infrared visible light to the addition of near-infrared and infrared visible light reflectance. The value of NDVI has no unit, ranging from − 1.0 to 1.0, in which higher value of NDVI indicates the larger green vegetation space (Denpetkul and Phosri 2021). The daily average NDVI value was then computed by taking the average of all 0.1° × 0.1° grids for every single day over the BMR.

Statistical analysis

The multiple-linear regression model was constructed in this study to examine the relationship between daily variation of AOD and daily concentrations of ground-level PM2.5 taken into account many potential confounders that may affect the relationship of daily varying AOD value with ground-level PM2.5. In particular, the linear regression model was tested whether a number of predictor variables increase the model performance using the forward stepwise selection method, where the multiple-linear regression model in this study was ultimately accounted for surface temperature, relative humidity, wind speed, NO2, O3, and NDVI. Daily surface temperature, relative humidity, and wind speed were included in the model because previous study indicated that increased levels of temperature, relative humidity, and wind speed were associated with reduced concentration of PM2.5 (Sritong-aon et al. 2021). Furthermore, NDVI was incorporated into the model because PM2.5 concentration might be explained by the variation of NDVI. Specifically, higher vegetation coverage may associate with lower concentration of PM2.5 (Zhou et al. 2021). Daily average concentrations of NO2 and O3 were also incorporated into the model because NO2 and O3 may be associated with the variations of PM2.5 concentration. In particular, high O3 concentration as strong oxidative pollutant during summer months is able to introduce the formation of secondary particles, thereby increasing PM2.5 concentration. On the other hands, high concentration of PM2.5 during winter suppresses solar radiation, resulting in lower O3 production and concentration (Jia et al. 2017). Moreover, chemical reaction of NO2 in the air promotes the formation of secondary PM2.5 (Balamurugan et al. 2022; Wu et al. 2016). Because of COVID-19 pandemic in 2020, Thai’s government has imposed the lockdown measure to prevent the transmission of SARS-CoV-2. This policy might influence the levels of AOD and ground-level PM2.5 concentration. Thus, indicator variable of COVID-19 lockdown period was also incorporated into the model. The time series linear regression model was applied in this study because we hypothesized that the association between AOD and ground-level PM2.5 concentration varies intensively daily, but minimally spatially over the specific region (Lee et al. 2011). The final model used to examine the relationship between AOD and ground-level PM2.5 is represented by Eq. (1).wherein Log(PM2.5) indicates the logarithmic scale of ground-level PM2.5 concentration at day t and AOD, Temp, Humid, Wind, NO2, O3, and NDVI are the level of AOD, surface temperature, relative humidity, wind speed, nitrogen dioxide, ozone, and NDVI at day t, respectively. COVID is the dummy variable for COVID-19 lockdown period at day t, whereas β1,2,3,…,8 indicate the vector of regression coefficients of AOD, surface temperature, humidity, wind speed, nitrogen dioxide, ozone, NDVI, and COVID-19 lockdown, respectively. The sensitivity analysis was also performed by excluding NO2 and O3 from regression model to explore the robustness of the model performance.

The 12-fold cross-validation method was then used to assess the model performance. Specifically, we stratified the observed PM2.5 concentration into 12 groups (N = 12) by calendar month and selected a group of month as the test set and the remaining N − 1 months were utilized as the training set. The prediction of a test set was generated from the model output of training set using Eq. (1). This procedure was repeated for 12 times until the prediction was completed in all 12 months. The simple linear regression model was then fitted to explore the correlation between observed and predicted PM2.5 concentrations where R2 value, as well as the root mean square error (RMSE) and mean absolute error (MAE), was utilized to evaluate the model performance. This model performance was also separately explored in different seasons, including winter (i.e., November to February), summer (March to June), and rainy reason (i.e., July to October). All statistical analyses were performed by R (version 4.1.0).

The annual and season-specific average concentrations of PM2.5 were calculated by averaging daily predicted PM2.5 concentration for the whole year and for season-specific, respectively. Because of missing AOD data due to cloud cover and precipitation (Levy et al. 2007), the bias-corrected annual or season-specific PM2.5 concentration was applied by calculating the ratio of annual or season-specific average of predicted PM2.5 concentration for all grid cells to annual or season-specific average of daily predicted PM2.5 concentration obtained from regression model as shown in Eq. (2). The output was then used to compute the bias-corrected annual or season-specific average PM2.5 concentration by multiplying with predicted annual or season-specific average concentration of PM2.5 for each grid cell (Zheng et al. 2016). All maps were produced by ArcMap (version 10.8).wherein Corrbias is the correction factor; PM2.5grid defines annual/season-specific average of predicted PM2.5 for all grid cells; and PM2.5predicted is annual/season-specific average of daily predicted PM2.5 level.

Results

The descriptive statistics of ground-level PM2.5, AOD, and other predictor variables in the BMR during the study period (i.e., January through December 2020) are shown in Table 1. The average of ground-level PM2.5 concentration (minimum–maximum) throughout the study period was 22.37 µg/m3 (7.23–86.67 µg/m3), which peaked during winter (November to February). The average of AOD value was 0.51 (0.03–2.28), where the highest level was observed between March and April (Fig. 2). The average of surface temperature, relative humidity, wind speed, NO2, and O3 was respectively 29.05 °C (23.59–32.87), 70.16% (41.41–93.35), 0.62 m/s (0.30–1.53), 13.50 ppb (3.71–50.07), and 36.65 ppb (12.01–78.31), whereas that of NDVI was 0.51 (0.43–0.60) during the study period. Figure 2 indicates time series variations of PM2.5 concentration and AOD during the study period, where PM2.5 concentration was peaked during winter (November to February) and the highest level of AOD was observed between March and April. The spatial distribution of AOD over the BMR is shown in Supplementary Fig. S1.

Table 1

Summary statistic for ground-level PM2.5, AOD, and other predictor variables in the BMR during January through December 2020 

Variables	Mean	SD	Min	25th	50th	75th	Max
PM2.5 (µg/m3)	22.37	12.55	7.23	13.79	18.16	27.07	86.67
AOD	0.51	0.31	0.03	0.32	0.43	0.64	2.28
Temperature (°C)	29.05	1.61	23.59	28.19	29.06	30.19	32.87
Relative humidity (%)	70.16	8.87	41.41	66.00	69.65	75.64	93.35
Wind speed (m/s)	0.62	0.18	0.30	0.48	0.59	0.73	1.53
NO2 (ppb)	13.50	7.00	3.71	7.98	11.97	17.76	50.07
O3 (ppb)	36.65	13.40	12.01	25.53	34.02	46.73	78.31
NDVI	0.51	0.05	0.43	0.47	0.51	0.55	0.60

SD is standard deviation; Min and Max are minimum and maximum values, respectively; 25th, 50th, and 75th define 25th, 50th, and 75th percentiles, respectively

Fig. 2

The concentration of PM2.5 and AOD over time during the study period, spanning from January through December 2020

Table 2 illustrates regression coefficient and its standard error (SE) of the relationship between MODIS AOD and concentration of PM2.5 controlling for many possible confounders as described in Eq. (1). In particular, a 1 unit increase of AOD was positively associated with 41.46% [(1.4146 − 1)*100] ± 5.75% increase of ground-level PM2.5 concentration and 1 ppb increase of NO2 and O3 was associated with 2.12 ± 0.25% and 1.81 ± 0.14% increase of PM2.5 concentration, respectively, whereas increasing each 1 °C of temperature, 1% of humidity, 1 m/s of wind speed, and 1 unit of NDVI was associated with 1.82% [(1 − 0.9818)*100] ± 1.08%, 0.20 ± 0.21%, 3.58 ± 9.36%, and 96.72 ± 45.63% decrease of ground-level PM2.5 concentration, respectively. Moreover, PM2.5 concentration during the time of COVID-19 lockdown was lower by 24.59 ± 5.87% compared to that during no imposed lockdown period. The multicollinearity is not likely to be a problem of the regression model in this study since the correlation among predictor variables was not very high (Supplementary Table S1). The multiple R2 of regression model is 0.71, indicating that all independent variables (i.e., MODIS AOD, surface temperature, relative humidity, wind speed, NO2, O3, NDVI, and COVID-19 lockdown period) could be able to explain 71% of the variability in ground-level PM2.5 concentration in BMR significantly (p-value < 0.05). There is high correlation between observed PM2.5 and predicted PM2.5 concentrations obtained from regression model with Pearson’s correlation coefficient of 0.82 and p-value < 0.05 (Fig. 3A). The daily variation of observed PM2.5 was highly correlated with predicted PM2.5 concentration (Fig. 3B), indicating that the prediction accuracy was improved when NO2, O3, temperature, relative humidity, NDVI, and dummy variables of COVID-19 lockdown period were integrated into the model.

Table 2

Regression coefficient and its standard error (SE) of the relationship between MODIS AOD and concentration of PM2.5 controlling for many possible confounders

Variables	β coefficients	Exp(β)	Standard error (SE)	Exp(SE)
Intercept	4.3454	77.1193	0.4462	1.5624
AOD	0.3469	1.4146	0.0559	1.0575
Temperature	− 0.0183	0.9818	0.0107	1.0108
Relative humidity	− 0.0020	0.9980	0.0021	1.0021
Wind speed	− 0.0364	0.9642	0.0895	1.0936
NO2	0.0209	1.0212	0.0025	1.0025
O3	0.0179	1.0181	0.0014	1.0014
NDVI	− 3.4171	0.0328	0.3759	1.4563
COVID-19 lockdown	− 0.2822	0.7541	0.0570	1.0587

Regression coefficient is interpreted per 1 unit change of AOD and NDVI, 1 °C change of temperature, 1% change of humidity, 1 m/s change of wind speed, and 1 ppb change of NO2 and O3, whereas that of COVID-19 lockdown is interpreted by comparing between with and without lockdown period

Fig. 3

The correlation between observed and predicted PM2.5 concentrations (A) and time series plot of observed and predicted PM2.5 concentrations (B) obtained from the regression model 

The results of 12-fold cross-validation are depicted in Fig. 4. Specifically, the association between observed PM2.5 concentration and predicted PM2.5 concentration obtained from cross-validation for the whole year and for season-specific was indicated in terms of correlation plot, R2, MAE, and RMSE. Based on correlation plot, the closer the points are, the higher model accuracy is. Moreover, higher R2 and lower MAE and RMSE indicate higher model accuracy, and vice versa. The R2 of cross-validation for the whole year period, as well as for summer, winter, and rainy season was 0.48, 0.55, 0.21, and 0.52, respectively, which are lower than the R2 from the regression model (i.e., 0.71). The RMSE and MAE of cross-validation in winter (13.72 µg/m3 and 6.64 µg/m3) were greater compared to those in summer (6.97 µg/m3 and 4.24 µg/m3) and rainy season (4.01 µg/m3 and 2.94 µg/m3), respectively. Therefore, the prediction accuracy during summer and rainy season was higher than that in winter and over the whole year period. Moreover, the performance of regression model yielded reasonable prediction for the whole year and for season-specific as indicated by the slopes, intercepts, and R2 values. In particular, the model underestimates PM2.5 concentration by 0.5% and 12.4% for the whole year period and in winter, respectively, while overestimates PM2.5 concentration in summer and rainy season by 18.1% and 2.9%, respectively (Table 3). Figure 5 indicates the spatial distribution of the bias-corrected annual and season-specific average PM2.5 concentration over the BMR, where dark blue and red indicate lower and higher concentrations of PM2.5, respectively. The annual average concentration of PM2.5 was higher in the downtown area of Bangkok compared to the rest of the study area (Fig. 5A). The average concentration of PM2.5 during winter (Fig. 5B) was also higher in the city center of Bangkok and that was higher than that during summer (Fig. 5C) and rainy season (Fig. 5D). The additional information of bias-corrected annual and season-specific average of PM2.5 concentration over the BMR is shown in Supplementary Table S2.

Fig. 4

The simple linear regression between observed and predicted PM2.5 concentrations obtained from 12-fold cross-validation for a year round, winter, summer, and rainy season. The solid lines indicate the regression line, whereas dashed lines represent the 0:1 (intercept:slope) line

Table 3

The 12-fold cross-validation results on ground-level PM2.5 prediction

Variables	Intercept	Slope	R2	RMSE	PM2.5 (mean ± SD) (µg/m3)
					Predicted	Observed
Full model
Whole year	9.98	0.55	0.48	9.16	22.25 ± 9.97	22.37 ± 12.55
Summer	4.27	0.95	0.55	6.97	21.68 ± 9.14	18.36 ± 7.14
Winter	19.29	0.30	0.21	13.72	29.43 ± 9.45	33.60 ± 14.48
Rainy	2.64	0.86	0.52	4.01	15.74 ± 5.68	15.30 ± 4.78
no-NO2 and O3 model
Whole year	14.98	0.29	0.19	11.63	21.39 ± 8.22	22.37 ± 12.55
Summer	12.72	0.51	0.19	8.98	22.04 ± 8.27	18.36 ± 7.14
Winter	24.59	0.05	0.01	17.49	26.40 ± 8.31	33.60 ± 14.48
Rainy	11.99	0.25	0.14	4.63	15.83 ± 3.16	15.30 ± 4.78

Full model is including the following variables into the model as showed in Eq. (1): AOD, surface temperature, relative humidity, wind speed, NO2, O3, NDVI, and indicator of COVID-19 lockdown period

Fig. 5

The spatial distribution of annual average bias-corrected PM2.5 concentration (A), and the mean during winter (B), summer (C), and rainy season (D) over the BMR. The black dots define air quality monitoring stations situated in the BMR

Figure 6 illustrates the results of 12-fold cross-validation by means of excluding NO2 and O3 from the model. Finding indicates that the accuracy of model performance was lower than that of the model with NO2 and O3 (as indicated in Fig. 4), in which R2 of the model without NO2 and O3 was lower and RMSE was greater for both whole year and season-specific periods. Specifically, incorporating NO2 and O3 into the model considerably improved the model performance, where the slope increased by 0.26 and intercept declined by 5.00 for the whole year period. This finding was also observed for season-specific analysis, where including NO2 and O3 into the model substantially increased the slope by 0.44, 0.25, and 0.61, and the corresponding intercept decreased by 8.45, 5.30, and 9.35 in summer, winter, and rainy season, respectively (Table 3).

Fig. 6

The simple linear regression between observed and predicted PM2.5 concentrations obtained from 12-fold cross-validation without NO2 and O3 for a year round, winter, summer, and rainy season. The solid lines indicate the regression line, whereas dashed lines represent the 0:1 (intercept:slope) line

Discussion

Many previous studies have revealed that regression model with the MODIS AOD measurement can be used as predictor variable to estimate the spatial ground-level PM2.5 concentration, taken into account many potential confounders (Guo et al. 2021; Hu et al. 2014; Xu and Zhang 2020; Zheng et al. 2016). This method would benefit for exposure assessment for epidemiological research, especially in the areas with no monitoring station network that is commonly used to explore the association of PM2.5 with morbidity and mortality. In this study, we found that linear regression model would dramatically improve the prediction accuracy when air pollutants, meteorological variables, greenness indicator, and dummy variable for COVID-19 lockdown period were integrated into the model with the overall R2 of 0.71, but cross-validation R2 was 0.48, suggesting that the overfitting was found in the model. This finding is similar to many previous studies (Guo et al. 2021; Li et al. 2015; Wang et al. 2019; Zheng et al. 2016), implying that the overfitting is commonly existed in the model. The model performance on investigating the association between MODIS AOD and ground-level PM2.5 concentration varied by season (Xin et al. 2014). Specifically, the prediction accuracy observed during winter was lower than that during summer, rainy season, and a year round, where the R2 from 12-fold cross-validation was 0.21, 0.55, 0.52, and 0.48 for winter, summer, rainy season, and a year round, respectively. The RMSE from 12-fold cross-validation during winter (13.72 µg/m3) was also higher than that in summer (6.97 µg/m3), rainy season (4.01 µg/m3), and the whole year (9.16 µg/m3). This finding might be contributed from the substantially higher concentration of PM2.5 observed during winter, where multiple-linear regression model may not be able to predict the high concentration of PM2.5 accurately (Zheng et al. 2016). In addition, the concentration of PM2.5 was the lowest in rainy season with some missing grid cells as we found in Fig. 5D. This finding was caused by missing MODIS AOD data due to substantial cloud cover observed during rainy season. The slope of simple linear regression between observed PM2.5 and predicted PM2.5 concentrations obtained from 12-fold cross-validation for the whole year, winter, summer, and rainy season was respectively 0.55, 0.30, 0.95, and 0.86, indicating that the predicted PM2.5 concentration from linear regression model during summer was highly accordant with the observed concentration since the slope is close to 1. This result suggests that the model performance during summer is more accurate, to some extent, compared to that during other seasons, although regression model was overfitted the data.

We also found that incorporating various meteorological variables and greenness indicator into the model would improve the model performance on the prediction of ground-level PM2.5 concentration, which is in agreement with previous studies (Kloog et al. 2011; Liu et al. 2007; Ma et al. 2014; Wang et al. 2019; Zheng et al. 2016). Furthermore, we compared the prediction performance of regression model between a model with and without NO2 and O3, where incorporating NO2 and O3 into the model substantially improved the prediction accuracy, in which the slopes of cross-validation were increased and the intercepts were dramatically decreased. This finding is similar to a previous study (Zheng et al. 2016). Therefore, the local air pollutants might also influence the model performance on investigating the association between MODIS AOD and ground-level PM2.5 concentration.

The principal limitation of this study is a lacking of MODIS AOD data in a certain day during the study period that is commonly caused by high cloud cover or high surface reflectance (Hu et al. 2014). However, the main objective of this study was to develop the statistical model to predict ground-level PM2.5 concentration using temporal AOD data, controlling for other potential confounders, because we assumed a priori that the association between AOD and ground-level PM2.5 varies intensively daily, but minimally spatially over the BMR. Therefore, the availability of daily AOD value could be applied to estimate ground-level PM2.5 concentration, to some extent. Moreover, we replaced the missing AOD data using EM algorithm, where the additional measurement error might be introduced, but a previous study revealed that this method is appropriate for time series data that is fitted to the data of this study (Junger and Ponce de Leon 2015). We also interpolated AOD data from 3- to 1-km resolution using the bilinear interpolation technique, which might also introduce the measurement error. Therefore, further studies are needed to estimate the finer spatial scales of PM2.5 concentration using the finer resolution of AOD.

AOD products have been widely used to estimate ground-level PM2.5 concentration in many previous studies using linear mixed effect model and geographically weighted regression (GWR) models (Lee et al., 2011; Ma et al., 2014; Zheng et al., 2016). However, integration of the MODIS AOD, air pollutants, meteorological variables, and greenness indicators in the current study provides strength to accurately estimate ground-level concentration of PM2.5 using multiple-linear regression model. This finding indicates that although the complicated models or algorithms are not utilized, linear regression model with appropriate predictor variables can be used to accurately estimate surface concentration of PM2.5. Moreover, this study is the first study that comprehensively estimates ground-level PM2.5 concentration in Bangkok by integrating the MODIS AOD, air pollutants, meteorological variables, and greenness indicator using linear regression, and the prediction accuracy was improved, to some extent. Therefore, model coefficients of various predictor variables generated from this study can be utilized to estimate the concentration of PM2.5 in an area with no air quality monitoring station in Bangkok, Thailand.

Conclusion

The multiple-linear regression model was used to estimate ground-level PM2.5 concentration using the MODIS AOD as predictor, controlling for various meteorological variables, greenness indicator, local air pollutants, and dummy variable of COVID-19 lockdown period. The annual average concentration of predicted PM2.5 (22.25 ± 9.97 µg/m3) was similar to that of observed values (22.37 ± 12.55 µg/m3), indicating the accurate results. The accuracy of model performance was varied by seasons, where the most accurate finding was observed during summer with R2 and RMSE of the cross-validation of 0.55 and 6.97 µg/m3, respectively. The model performed better when the daily average ambient NO2 and O3 concentration was incorporated into the model. Finding from this study suggests that linear regression model with appropriate predictor variables could be used to estimate ground-level PM2.5 concentration that might be applied further for epidemiological studies.

Below is the link to the electronic supplementary material.

Supplementary file1 (DOCX 878 KB)