Introduction

Air quality is a significant element in climate change and causes environmental and public health problems. Globally, particulate matter (PM) concentrations with a diameter of less than 2.5 µm (PM_2.5) or less than 10 µm (PM₁₀) have become an environmental problem and have generated increased attention [1]. Therefore, PM has received much attention in many countries. Previous studies indicated the tendency to increase air pollutant concentration trends in developed and developing countries [2-4]. PM_2.5 had a significantly increasing trend in Southeast Asian countries like Bangladesh, Myanmar, Laos, and Thailand [3]. The largest source of particulate pollution in Southeast Asia is open crop residue burning. When more dust and gases are released into the air, it significantly impacts the air quality in the areas, which is particularly bad in the dry season [5].

Different studies have identified various factors associated with PM concentration at specific locations. Meteorological features such as wind speed, temperature, and relative humidity significantly affect PM concentration in different areas in Thailand [6-8]. Vehicular traffic and vegetation cover are also significant features associated with PM concentration [9, 10]. Previous studies have extensively discussed methods for modeling, predicting, and forecasting PM concentrations.

Major cities in Thailand have experienced a significant increase in air pollutant concentration over the last few years [11]. Air pollution problems have been more severe, particularly within the Bangkok metropolitan area and the northern parts of Thailand [12, 13]. The intensity and frequency of haze outbreaks have increased in the north of Thailand due to rapid changes in settlement planning [7-14]. The concentration of PM_2.5 and PM₁₀ in northern Thailand usually exceeds WHO standards during the dry seasons [15]. The hourly records of PM exceed the accepted standards of the NAAQ and the WHO [9].

The southern part of Thailand lies between the Gulf of Thailand to the east and the Andaman Sea to the west. Due to tourist activities, the region has become an important economic hub for Thailand. The south of Thailand has seasonally been under hazy conditions caused by smoke from open burning and forest fires from within the country and other neighboring countries. PM concentration in southern Thailand was higher in 2015 than in subsequent years, mostly due to transboundary atmospheric aerosol transportation [16]. Many studies have established a significant relationship between PM concentration and adverse health outcomes (see, e.g., [17, 18]). The concentration of PM in the southern part of Thailand has historically been below the national average concentration, although recent findings indicate a gradually increasing trend in PM concentration over the southern region. However, very few studies have been conducted in recent years to explore the trends of PM concentration and meteorological factors and their associations within the southern region.

Moreover, predicting PM_2.5 is also very important, as these are very small particles and are sometimes not measured by devices. Many researchers develop time series-based machine learning models for predicting PM_2.5 (see, e.g., [19-22]). Which are computationally extensive and need historical data to predict. Predicting PM_2.5 using other correlated available pollutants and meteorology variables can reduce the computational cost and massive data collection.

Therefore, this study aimed to: – Analyze the trends of PM concentration and

meteorological factors in southern Thailand. – Explore the association between PM concentration

and meteorological factors in southern Thailand. – Recommend the most accurate models for predicting

PM_2.5 concentration in cases where PM_2.5 is

unavailable at certain monitoring locations, but PM₁₀

and other meteorological factors are available.

Material and Methods

Study Area and Data Source

The study area consists of five main provinces of southern Thailand that have air pollution and weather stations. In 2022, the population of southern Thailand was 9.49 million with an area of 70,715 Km². These provinces in southern Thailand stations include Surat Thani, Phuket, Songkhla, Narathiwat, and Yala. Hourly data on PM_2.5 and PM₁₀ concentrations and meteorological factors including wind speed, temperature, relative humidity, and atmospheric pressure were obtained from the Pollution Control Department of Thailand. The locations of air pollution monitoring stations are shown in Fig. 1. The hourly data of air pollutant concentration and meteorological factors were taken from each station from 1 January 2012 to 31 December 2021 and aggregated into daily data.

Fig. 1.

The locations of the monitoring stations in the South of Thailand area.

Statistical Analysis and Modeling

The obtained data were preprocessed before analysis. We identified the outliers, and also missing values were imputed using multiple imputations, which were less than 15% in each station and for each variable. After pre-processing, we plot each PM and meteorological factor and fit a cubic spline model to understand the trend and pattern. Then we used a scatter plot matrix with a Pearson correlation coefficient. A scatter plot was used to investigate the linear relationship between these variables, and then correlation was considered to evaluate the strength of that linear relationship among variables. Finally, we compare the machine learning models to predict PM_2.5 when the station measures PM₁₀ and has meteorological factors. Data for each station were split into training and testing sets at a ratio of 80:20. The first 80% of observations were used to train the models. Then the predictive performance of these models was assessed on 20% of the observation of test data using root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). The flowchart of the study is presented in Fig. 2. All statistical analysis was performed using R programming. The details of the used models are as follows:

The cubic spline function was used to examine the seasonal patterns and trends in particulate matter variables by dividing the whole range into knots. The cubic spline model is as follows:

(1)

S(t)=a+bt+∑i=1kCi(t−ti)3

Fig. 2.

Flowchart of the Study.

where S(t) is the spline function for PM_2.5 at time t. a, b, C_k are parameters in the model and k indicates the number of knots.

Multiple Linear Regression (MLR) is a simple predictive approach that establishes a linear relationship between a continuous variable (e.g., PM_2.5 concentration) and multiple determinants. The MLR is defined as follows:

(2)

Yt=β0+∑i=1pβiXi

where, Y_t is the variable of interest, β₀ is the intercept, and ach β_i represents the coefficient of each determinant X_i.

The Support Vector Machine (SVM) was initially developed as a classification algorithm by Vapnik and Cortes [23]. However, the introduction of support vector regressors allowed SVM to be applied to numeric data. The SVM applies the kernel function to create a map of data in higher dimensional space for regression analysis [24]. The SVM model can be written as:

(3)

f(x)=w.x+σ

where x ∈ X represents a vector of independent variables, w ∈ X are weighting parameters each x ∈ X and σ indicates the distance between each hyperplane and the observed factor.

Random forest (RF) is a supervised learning algorithm that combines decision trees to create a forest based on bagging [25]. It uses the bagging technique to select variables’ random samples as the model calibration training dataset [26]. Decision trees in the forest are created using a random selection of features. Each decision tree is used to predict the training data, and the average from all decision trees is chosen as the predicted outcome.

The development of Artificial Neural Networks (ANN) is based on the human brain and the connections between neurons [27]. ANN consists of an input layer, hidden layers, and an output layer. In ANN, one input layer is mainly responsible for input parameters, one or more hidden layers, and one output layer. The networks generally have a learning system based on examples that incorporate new ones. The error estimation of the learning system is backward-spread, and the learning process is continuous. After sufficient learning is achieved, the networks are used to make predictions based on input data.

Results and Discussion

The descriptive statistics include mean, standard deviation (SD), median, maximum, and minimum of daily air pollutant concentrations and meteorological features at each station presented in Table 1. Songkhla has the maximum average exposure of PM_2.5 and PM₁₀, which were 18.2±8.24 µg/m³ and 36.3±14.2 µg/m³, respectively. While the minimum exposure of PM_2.5 and PM₁₀ was 13.9±6.42 µg/m³ and 27.5±12.2 µg/m³ in Phuket. Average levels of meteorological features varied across the southern region of Thailand. The overall average temperature in Southern Thailand was 28.9ºC in Phuket and 28.6ºC in Songkhla. The daily concentrations of PM and meteorological factors for all five monitoring stations are shown in Fig. 3. A preliminary assessment of air pollutant concentration at the five monitoring stations indicated varying characteristics in each province. We observed significant variability in the concentration of PM_2.5 and PM₁₀ at the Songkhla, Narathiwat, Surat Thani, and Phuket stations. PM concentration in Surat Thani province showed significant annual variability, with high pollutant levels in December and January. The daily concentrations of PM_2.5 and PM₁₀ in Phuket were consistent with Surat Thani province. The cubic spline model (indicated with a red curve) was used to examine the seasonal patterns and trends in PM concentration. The spline function showed significant seasonal variability in PM_2.5 and PM₁₀ concentration. Phuket has an increasing trend of both PM concentrations. Similarly, variability and trends can be seen in wind speed and atmospheric pressure in each region, while temperature has a strong seasonal pattern in each region.

The relation between PM constraints and meteorological features was determined using a scatter plot and Pearson correlation analysis in Fig. 4. The diagonal presents the histogram of each series. Lower diagonals present the scatter plot to examine the linear or non-linear relation between variables, and the upper diagonal presents the Pearson correlation coefficient with significance present with steric signs. Scatter plots and correlation measures show that both PM concentrations are positively, strongly, and linearly related. Temperature is positive but very weaker related to PM in Songkhla, Phuket, and Narathiwat. Windspeed has a negative relation with PM in Songkhla but is positively related to Yala and Surat Thani. Relative humidity is negatively associated with PM concentrations at all stations, so it can decrease the concentration of PM in the air. Higher atmospheric pressure accumulates more particles so positively associated with PM concentrations except in Narathiwat.

The performance of RF, SVM, MLR, and ANN for predicting PM_2.5 was evaluated and presented in Fig. 5 for each station with different evaluation measures. Each graph presents the observed PM_2.5, which is test data, and the predicted values using each model along with RMSE, MAE, and MAPE. The results revealed that the MLR and ANN have almost the same and lowest values of RMSE, MAE, and MAPE measures for each station except Narathiwat, where SVM performs a bit better than other models. But if we compare the computational complexity and performance of all models, then MLR is the least complex, has good results, and is suggested to be used for predicting PM_2.5 in these regions.

Fig. 3.

PM Concentration and Meteorological features from five stations in Southern Thailand. The red line is the cubic spline fitted line.

Fig. 4.

Correlation between PM Concentration and Meteorological features from five stations in Southern Thailand.

This study has investigated the trends of PM_2.5 from five air quality monitoring stations in southern Thailand. The concentration of PM_2.5 was predicted at each station by combining PM₁₀ with meteorological features, including atmospheric factors such as temperature, pressure, wind speed, and relative humidity, using three machine learning models (support vector machines, random forests, and artificial neural networks) and the multiple linear regression model. The present study has also predicted PM_2.5 at these stations using machine learning models and multiple linear regression.

Results from the correlation analysis indicated a significant correlation between PM_2.5 and PM₁₀ at each station. The significance of meteorological factors (temperature, relative humidity, wind speed, and atmospheric pressures) differed by location. Theoretically, a higher atmospheric water concentration in the atmosphere dissolves particulates such as fine particulates, reducing PM concentrations. Consequently, the present study found negative correlations between relative humidity, the concentration of PM₁₀, and PM_2.5 at all stations in Southern Thailand. These results agree with findings from studies within the Southeast Asia region [28, 29]. However, similar studies have reported a positive correlation between relative humidity and PM concentration. The effect of relative humidity on pollutants may vary due to the type of aerosols in different atmospheres [30]. High temperatures cause convection, which may lead to the dispersion of air pollutants. However, this study reports no correlation between temperature and PM concentrations across the stations. As these regions have tropical weather, therefore, the relatively stable temperature could mitigate any positive correlation with particulate matter. However, similar studies found a significant correlation between temperature and PM [31, 32]. The predictive accuracy of the MLR model for predicting PM_2.5 is higher than other machine learning models. This might be due to the nature of the relation between these variables being linear. The results are location-specific, indicating that geographical factors could affect the concentration and distribution of particulate matter. However, the methodology can be used in further research.

Fig. 5.

Prediction Performance of PM_2.5 Concentration Using Different Machine Learning Models. A) RF Model. B) SVM Model. C) MLR Model. D) ANN Model. Performance Evaluations are based on Test Data.

Conclusions

All the provinces have high exposure to PM concentration as compared to WHO. While Songkhla has very high exposure, Phuket has an increasing trend of PM concentration, which needs more attention from the government and stakeholders in making policies for the future. Humidity is negatively affecting PM concentration in all these regions, while atmospheric pressure causes more PM particles. For the stations where PM_2.5 is missed, it can be predicted using PM₁₀ and other variables due to their significant relationship. Future studies can use this methodology for analyzing the pattern and predicting PM_2.5 in other countries and regions. Moreover, these results help local governments to make strategies and scientists to better understand the ground condition and suggest the best possible solution to policymakers that overcome the health issues that arise due to the exposure.