https://orcid.org/0009-0009-6758-8666
Figures
Abstract
Urban air pollution remains a critical global issue, particularly fine particulate matter (PM), which has direct and significant impacts on human health. Understanding PM patterns is challenging due to the large number of associated variables, many of which are highly correlated and must lead to multicollinearity issues in predictive modeling. To address this problem, this study applies Principal Component Analysis (PCA) as a dimensionality reduction technique and develops forecasting models for and
concentrations. The used air pollution data were collected from 4 monitoring stations in Bangkok, Thailand. Prior to analysis, the datasets were preprocessed through imputation, transformation, and standardization to ensure consistency across variables. PCA results identified 4 principal components for each station based on cumulative variance and eigenvalue criteria. Among these, relative humidity (RH) and zonal wind direction (ZWD) consistently contributed most to the principal components. The selected components were then used as predictors in 4 machine learning models, and their performance was compared with models using all original variables. Model evaluation was conducted using mean absolute error (MAE), mean squared error (MSE), and mean absolute percentage error (MAPE). The results indicate that models based on reduced predictors consistently outperform those using the full set of variables across all stations. These findings offer an alternate effective approach for addressing multicollinearity and improving model performance in urban air pollution forecasting.
Citation: Saeyang S, Notodiputro KA, Khurram H, Lim A, Chomtee B, Wanishsakpong W (2026) Principal component-based predictive modelling of particulate matters using air pollutants and meteorological variables in Bangkok, Thailand. PLOS Clim 5(6): e0000949. https://doi.org/10.1371/journal.pclm.0000949
Editor: Jing Chen, Beijing Normal University, CHINA
Received: December 14, 2025; Accepted: May 8, 2026; Published: June 24, 2026
Copyright: © 2026 Saeyang et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The data with air pollution features is available in the paper or Supporting Information files. The population density data is publicly available on the Worldometer website: https://www.worldometers.info/world-population/thailand-population/.
Funding: This research was supported by the International SciKU Branding (ISB), Faculty of Science, Kasetsart University, Thailand, and funded by the Research and Innovation Fund (NSRF), Grant No. SAT6801104S, Thailand.
Competing interests: The authors have declared that no competing interests exist.
1. Introduction
Air pollution has become a global environmental problem that continues to intensify alongside rapid technological development and industrial expansion. It directly affects ecosystems, oceans, and human societies worldwide. Among various air pollutants, fine particulate matter (PM) as and
has been widely recognized for its significant impact on human health, especially on respiratory and cardiovascular systems. According to global estimates, approximately 86% of urban residents around 2.5 billion people live in cities where annual
concentrations exceed the World Health Organization (WHO) 2005 guideline of 10 μg/m³, contributing to an estimated 1.8 million excess deaths in 2019 [1]. Bangkok, Thailand, is one of the densely populated urban areas experiencing serious air quality challenges due to diverse pollution sources, complex atmospheric conditions, and increasing public health concerns. Over the past decade, air quality in the city has gradually deteriorated [2]. Numerous factors contribute to PM concentration levels. Among all, meteorological and air pollutant variables are considered among the most influential determinants [3,4]. Understanding the relationship between
and these factors can provide valuable insights into air pollution dynamics and support strategies to mitigate risks related to respiratory and cardiovascular diseases [5–7].
Developing predictive models can significantly enhance air quality forecasting and support environmental decision-making [8]. Although and
concentrations are known to be associated with various meteorological and air pollutant variables, directly using all these factors in modeling is often inefficient due to the large number of predictors involved. Researchers sometimes include all available variables related to the outcome in order to reduce analysis time and computational effort. However, this approach can lead to overfitting and may introduce multicollinearity, where predictors are highly correlated with one another [9]. One effective method to address these issues is Principal Component Analysis (PCA). The PCA transforms correlated variables into a smaller set of independent components while retaining most of the original variance in the data. This dimensionality reduction approach improves computational efficiency and model stability. When PCA is combined with Multiple Linear Regression (MLR), the resulting method is known as Principal Component Regression (PCR) [10]. It allows the model to utilize information from all predictors in a more efficient way [11]. In recent years, machine learning (ML) techniques have also gained increasing attention for environmental and time-series prediction because of their ability to capture complex and non-linear relationships within data [12,13]. Previous studies [14,15] have applied ML models to investigate the relationship between PM concentrations and meteorological factors in southern and central Thailand. However, these studies typically used air pollution and meteorological variables directly for
prediction without applying dimensionality reduction techniques or addressing potential multicollinearity among predictors.
To address the challenges of high dimensionality and potential multicollinearity among predictors, this study applies PCA to reduce the number of variables in the air pollution dataset. The aim is to transform the original correlated predictors into a smaller set of principal components that retain most of the important information for modeling. These components are then used as inputs for predictive modeling. In addition to PCR, which serves as a benchmark model. Other three machine learning algorithms such as Extreme Gradient Boosting (XGBoost), Support Vector Regression (SVR), and Random Forest (RF) are employed to evaluate their effectiveness in predicting both and
concentrations. The dataset is divided into two subsets based on time order, with 80% of the data used as a training set for model development and the remaining 20% used as a testing set for performance evaluation. Model accuracy is assessed using statistical performance indices including Mean Absolute Error (MAE), Mean Squared Error (MSE), and Mean Absolute Percentage Error (MAPE). Through this approach, the study aims to examine whether dimensionality reduction through PCA can improve model efficiency and predictive performance in PM concentration modeling for Bangkok, Thailand.
2. Methodology
2.1. Study area and data descriptions
The Kingdom of Thailand, lying in the heart of the Southeast Asian mainland, covers an area of about 514,000 sq. km [16]. It is considered an urban heat island (UHI) due to its rapid urbanization, characterized by high population density in many main cities as shown in Fig 1(open-source data from Worldometer, see S1 Data). However, there is a very specific city, like Bangkok which has the most intense population because of its capital position, gaining most attention from the government and foreign investors compared to other main cities [17]. The data for this study were obtained from the Department of Pollution Control (DPC) in Bangkok, Ministry of Natural Resources and Environment, Thailand.
B. Study station in Bangkok. The base map was generated in R Studio using the ‘rnaturalearth’ package with data from Natural Earth.
This dataset comprises air pollution measurements and various meteorological data from hourly records of Bangkok, then averaged for all variables to daily observations. It covers the period from January 1, 2017, through April 30, 2023, with a total of 2,310 days of observations (See S2 Data). The air pollutants consist of concentration (
, µg/m³),
concentration (
, µg/m³), carbon monoxide (CO, mg/m³), sulfur dioxide (
, µg/m³), nitrogen dioxide (
, µg/m³) and ozone (
, µg/m³). The meteorological variables consist of wind speed (WS, m/s), wind direction (WD, °), temperature (Temp, °C), relative humidity (RH, %), rainfall (Rain, mm), and atmospheric pressure (Pres, mmHg). The data were collected from four air quality stations located within different districts of Bangkok: Rajabhat University Station (RU Station), Huai Kwang Station (HK Station), Nonthri Vithaya Station (NV Station), and Din Daeng Amphitheater Station (DD Station) as shown in Fig 1B. These four stations were selected to be in the center of Bangkok, which has the heaviest population and the highest PM intensity among other parts of Bangkok. However, the dataset isn’t acquired completely.
2.1.1. Handling missing data.
The proportion of missing data varied across pollutants and monitoring stations, approximately ranging from 3% to 90% (see Table 1). The highest level of missing data was observed for at station NV. An examination of the distribution of each air pollutant revealed that the missing values often occurred in consecutive sequences. This pattern is likely attributable to instrument malfunction during the measurement of air pollutant concentrations at the station. Therefore, the missing data can be classified as Missing Not at Random (MNAR) [18,19].
This study employed the Multivariate Imputation by Chained Equations (MICE) framework with five iterations to address missing data. Within this framework, the Predictive Mean Matching (PMM) method was selected to ensure that imputed values remained consistent with the distribution and trends of the observed data. Specifically, PMM imputes missing values by matching them to observed values from days with similar multivariate profiles [19]. Although several variables contained a substantial proportion of missing values, the MICE-PMM approach remained applicable, as it leverages information from other observed variables to estimate plausible values [18]. The imputation process was conducted to produce a complete air pollution dataset, which was subsequently used for all modeling analyses. Using 5 imputation number (m) with 50 iterations (maxit). The MICE method was chosen in this study after archiving the best impute accuracy compared to other relevant techniques (see S1 Table).
Regarding completed air pollution datasets. The and
concentration noticed to showing certain period with peak pollutant levels consistently happen in order every first few months of each year. To ensure mathematical continuity, only WD variable (degree unit from 0°-359°) was decomposed into two orthogonal trigonometric components: zonal wind direct from south to north (sin
) and meridional wind direct from east to west (cos
) where
is WD angle [20,21]. This transformation resolves the numerical discontinuity at the 0°/359° boundary [22]. For resulting, later recall sin
to ZWD and cos
to MWD (both using m/s as unit).
2.2. proposed methodology
2.2.1. Principal Component Analysis (PCA).
Principal Component Analysis (PCA) transforms the original 𝑘-dimensional coordinate system of the dataset into a new set of orthogonal axes, known as principal components (PCs). This transformation effectively projects high-dimensional data onto a lower-dimensional space while preserving the maximum variance. To ensure comparability across variables with different units of measurement, Z-score standardization was applied to all 11 predictor variables at each monitoring station, excluding PM10 and PM2.5. This standardization centers each variable at a mean of zero and scales it to unit variance, allowing PCA to identify the directions of maximum variance based on the correlation structure rather than the original measurement scales.
Subsequently, a covariance matrix was computed from the standardized data, and eigenvalues were derived to quantify the variance explained by each principal component. The eigenvalues represent the proportion of total variance captured by each PC [23]. Principal components were retained based on two criteria: a cumulative explained variance exceeding 60% and eigenvalues (𝜆) greater than 1 [24]. The relationship between a set of 𝑘 observed variables and 𝑘 principal components can be expressed as follows:
where is the
principal components;
is the observed variables
;
is the weight chosen for the i observed variable to maximize the ratio of the variance of the j principal component to the total variation [25].
2.2.2. Principal Component Regression (PCR).
Once the PCAs complete, their result are PCs score that can be used to conduct the forecast model by combining with Multiple Linear Regression (MLR). The model is known as PCR. It was constructed using the selected principal components as predictors (independent variables) to predict the outcome (dependent variable). The PCR is defined as a linear combination of the independent variables with an error term as shown in equation [26].
where is the forecasted concentration of PM;
are selected PC variables;
is constant of
;
are PCs scores present as coefficient of each component;
the error term.
2.2.3. Extreme Gradient Boosting (XGBoost).
XGBoost model is based on the gradient boosting framework, designed for both classification and regression tasks. It works by building a series of decision trees sequentially, where each new tree is trained to correct the errors made by the previous ones. The model also includes a regularization term in its objective function to control model complexity and prevent overfitting. [27]. The mathematical prediction for a raw XGBoost model is the sum of the outputs from individual trees:
where is XGB regression model;
represent observed variable; and
is the structure and leaf weights of the
tree. By applying PCA to XGBoost. It splits nodes based on the PCs. These components represent broad atmospheric “modes” or “patterns” identified by the PCA [28]. The additive model of XGBoost is redefined in terms of the latent component space. The final prediction
is expressed as:
where is the forecasted concentration of PM;
is the structure and leaf weights of the
tree; and
are PCs scores present as coefficient of each component. In this study, using gradient boosted trees (gbtree) as booster. The number of sequential trees was 100, the maximum depth of each tree was restricted to 6. A learning rate or shrinkage factor of 0.1 was applied to each tree to ensure that the boosting process converges slowly and accurately.
2.2.4. Support Vector Regression (SVR).
SVR is a kernel-based machine learning algorithm that works on a different principle than standard regression [29]. Instead of trying to minimize the total squared error, SVR attempts to find a function that has at most deviation from the actual target values for all training data. The general equation for the SVR function is:
where is SVR function;
represent number of support vectors;
denotes a weight vector;
is the kernel function; and
represents the bias. In context of PCR modeling, The SVR algorithm calculates the support vectors based on these m components rather than all variables [30]. which lead final model for prediction of PM concentration (
) becomes:
where is the forecasted concentration of PM;
represent number of selected components;
denotes a weight vector of PCs;
is the kernel function applied specifically to the component space; and
represents the bias. In study, using Radial Basis Function (RBF) as kernel function and utilizes a cost parameter of 1 (C = 1). It ensures forcing the model to focus only on the significant deviations and outliers in the air pollution dataset.
2.2.5. Random Forest (RF).
RF is an ensemble learning method that constructs a multitude of decision trees during training. For regression tasks, it outputs the average prediction of all individual trees. Its strength lies in Bagging (Bootstrap Aggregating), where each tree is trained on a random subset of the data.[31]. While a single decision tree is a set of hierarchical [32], the RF equation is the ensemble average:
where is RF regression model;
The total number of trees in the forest; and
is the prediction of the
individual regression tree based on the input vector
. Regarding of PCA process earlier, the vector
was replaced by PCs which lead target variable as a function of the latent component space:
where is the forecasted concentration of PM;
The total number of trees in the forest; and
is the
regression tree built using the component scores as the splitting criteria.
are PCs scores present as coefficient of each component. The model set up used bootstrap sampling with replacement, 100 trees, and 5 nodes (To ensures the trees don’t overfit by becoming too deep).
2.3. Study procedure
The research procedure initiated with the collection of a comprehensive Bangkok air pollution dataset, where missing data points were addressed using MICE-PMM to ensure a statistically consistent and impute the missing. Later, the experimental workflow branched into two parallel paths as shown in Fig 2.
In the first path, the 11 variables (CO, ,
,
, WS, ZWD, MWD, Temp, RH, Rain, and Pres) as predictors underwent Z-score standardization to achieve unit variance (only ZWD and MWD were already scaled during transform from WD). Later the full dataset across all station has process reduce variables by PCA separately. To extract a reduced set of orthogonal PCs will be using cumulative variance and eigenvalue selection boundaries. In the second path, the original 11 variables were retained in their raw state to serve as baseline control. Both feature sets were then utilized to forecast
and
concentrations. To maintain the temporal integrity of the air quality data, a chronological split was employed rather than a random shuffle. The dataset was partitioned into an 80% historical training set, comprising the earlier portion of timeline, and a 20% independent test set. The forecast model regression conducts four distinct machine learning architectures (MLR, XGBoost, SVR, and RF) were independently trained and validated. The final stage of the procedure involved a comparative performance analysis, where the predictive efficiency of the PCA-based hybrid models was measured against the full-variable baseline models using Mean Absolute Error (MAE), Mean Squared Error (MSE), and Mean Absolute Percentage Error (MAPE) as evaluate indices [33].
where is count of PM daily records;
is observed daily concentration of PM; and
is forecasted daily concentration of PM. The data processing and predictive modeling for this study were conducted using the R programming language (version 4.4.1). To facilitate comprehensive analysis, several specialized libraries and packages were employed. Data manipulation and preprocessing were performed using dplyr and tidyr, while e1071 was utilized for SVR and randomForest for ensemble learning. The gradient boosting models were implemented via the xgboost framework. For dimensionality reduction and statistical evaluation, the base stats package and the Metrics library were used to calculate error indices such as MAE, MSE, and MAPE. Finally, data visualizations, including the correlation matrices and time-series comparison plots, were generated using ggplot2 and corrplot.
3. Results
Initial data analysis focused on characterizing the baseline atmospheric conditions at the four monitored stations. As shown in the descriptive summary (see Table 2), there is a strong positive skew in pollutant distributions. Particularly for and
, indicating frequent episodic pollution events. For instance, the recorded value for
at DD station reached average 32.64 µg/m³ and
reached average 63.67 µg/m³. The PM pollutants in DD station are significantly higher than at the other three locations. Such disparities highlight the influence of local emission sources and urban morphology on particulate matter distribution, providing a robust foundation for the multi-station modeling performed in this study.
Following the descriptive analysis, the inter-relationships between meteorological parameters and air pollutants were examined using a Pearson correlation () matrix for each station as shown in Fig 3. To ensure statistical rigor, only correlations with a significance level of p-value < 0.05 are displayed as “*”, while non-significant pairs are “ns”.. The results indicate that both
and
exhibit significant correlations with all 11 predictors somehow, especially CO,
, and
were showing high correlate across all station. However, the matrix also reveals strong significant correlations between the predictors themselves (e.g., Temperature, Humidity, and Solar Radiation). This high degree of multicollinearity suggests that the raw variables contain redundant information, which can destabilize traditional regression coefficients. Consequently, PCA was performed to transform these correlated variables into a set of independent components, ensuring a more robust and reliable modeling process.
To address the multicollinearity identified in the correlation analysis, PCA was applied to the 10 predictors for each of the four stations from air pollution datasets. The results of the eigenvalue decomposition, including the variance proportion (%VP) and cumulative variance proportion (%VPC), are summarized in Table 3. Across all stations, the first four PCs were found to have greater than 1. Moreover, these four selected PCs collectively explain approximately above 65% of the total variance in the original dataset.
The variables contributions for four components as selected in previous section as shown in Fig 4. The first components (PC1) are the direction in the data that has the largest spread or most variance. By holding the heaviest variances, they also influence the model conducting larger than other components respectively. The red dotted line represents the contribute threshold at 8.33% (averaged from 100% / 12 PCs). However, the probability of the same variable contributing most strongly to the first component across all stations was low. At the RU station, the first principal component was mainly influenced by Temp, RH, Pres, and ZWD. Each of these variables contributed more than 10% to their respective components. At the HK station, the first principal component was dominated by ,
, RH, CO, and ZWD. At the NV station, the first principal component was primarily influenced by RH,
, ZWD,
, and CO. At the DD station, the first principal component was dominated by RH, ZWD, WS,
, and MWD.
Alternatively, when PCA all 11 variables with dividing by station can visualized through biplot as shown in Fig 5 to understand directional tendencies of air pollution and meteorological variables for all four stations along. Since all station dataset were overlapping, showing all 11 variables in air pollution data is nearly identical across Bangkok. Furthermore, the direction of arrow represents the correlation between variables. They were correlated positively whenever the direction went along in the same direction otherwise negatively.
To evaluate the efficacy of the dimensionality reduction, the performance of the four ML models (MLR, XGBoost, SVR, and RF) was using the full 10-variable suite. On the other hand, the reduced model will use 4 selected components per station as predictors. Both methods will be conducted by ML models to perform forecasts and
separately. The models were trained for the first 80% period of the dataset with different stations, while 20% test sets were used to evaluate the model’s performance efficiency. For comparison, Fig 6, 7, 8 9 presents 4 station model dividends, the upper models showing the performance of the original number of predictors while the lower are showing the reduced predictors’ performances.
The performance of the four machine learning models (MLR, SVR, RF, and XGBoost) was compared across two distinct predictor configurations for both and
concentrations. Table 4 presents the error metrics using the original 11 independent variables, while Table 5 details the performance using only the first four PCs derived from the PCA. Across all monitoring stations (RU, HK, NV, and DD), models trained on the first four PCs yielded lower MAE, MSE, and MAPE values compared to those trained on all variables as predictors. This trend was consistent for both
and
forecasts, suggesting that the dimensionality reduction successfully concentrated the predictive signal while filtering out redundant atmospheric noise.
4. Discussion
The increasing sophistication of environmental monitoring has led to the generation of massive, high-dimensional datasets. As datasets size grow and complex, the likelihood of correlations among variables also increases. In modeling processes, such correlated predictors can lead to biased estimates and instability, which is commonly referred to as multicollinearity. To address this issue, this study employed PCA to reduce the dimensionality of air pollution datasets collected from four monitoring stations located in central Bangkok. It was selected as a study target. However, since the dataset has over 50% Standard Deviation (SD) in several variables. It is represents the authentic daily and seasonal fluctuations of the Bangkok metropolitan area.[34]. Prior to dimensionality reduction, the air pollution dataset contained missing values across multiple variables. To handle this issue, the MICE-PMM method was applied.
Even so, the primary objective of imputation in this study was to obtain a complete dataset. Therefore, further sensitivity analysis of imputation accuracy was not conducted. Additionally, because variables were measured in different units, wind direction (WD) is expressed in degrees and is not directly suitable for linear modeling. To adjust, it the required transformation of WD. It was decomposed into two continuous components: the zonal component (ZWD), representing the sine function, and the meridional component (MWD), representing the cosine function. This transformation allowed wind-related variables to be incorporated effectively into the analysis. Correlation analysis revealed that relative humidity (RH) was negatively associated with particulate matter (PM) concentrations across all stations, consistent with previous studies [35,36]. In contrast, carbon monoxide (CO), nitrogen dioxide (), and ozone (
) showed positive correlations with PM, aligning with findings from earlier research [14,37–39]. The observed positive relationship between CO and
can be explained by their common emission sources, particularly vehicular and combustion-related activities in urban environments [40]. For dimensionality reduction, 11 variables (CO,
,
,
, WS, ZWD, MWD, Temp, RH, Rain, and Pres) were selected. While
and
were excluded as they served as response variables in subsequent modeling. Prior to PCA, all variables were standardized using Z-score normalization to ensure comparability. Except for ZWD and MWD, which were inherently scaled during transformation. PCA was then applied, and the first 4 components were selected in all stations based on cumulative variance exceeding 60% and eigenvalues greater than one. Over 60% cumulative variance is acceptable since we focus on achieving the lowest number of components which is the purpose of data reduction [40]. Therefore, the retained components in this study are sufficient to capture the major variability of the dataset by showing all model performances were better than full dimension models. Although PCA groups variables based on variance structure, resulting in differences across stations. The RH and ZWD appeared as major contributors to the first principal component in all stations. High humidity levels facilitate the conversion of gaseous precursors (such as
and
from traffic) into solid particulate matter. This process significantly increases the mass concentration of
during humid periods [41]. Since Bangkok is located close to the ocean, the ZWD represents the east-west movement of wind. When winds originate from the sea, they may either disperse pollutants or trap them against urban structures in the city center. This ventilation effect is a primary factor influencing PM concentrations. [42]. This results suggests that the interaction between relative humidity and zonal air transport is the most significant latent factor influencing the variance of the Bangkok atmospheric system [43].
To evaluate the effectiveness of dimensionality reduction, the selected principal components were used as predictors in four models to forecast and
concentrations. Model performance was compared against models using all 11 original variables. The dataset was split chronologically, with 80% used for training and 20% for testing. Model performance was assessed using MAE, MSE, MAPE. The results showed that models using the reduced set of principal components consistently outperformed those using all original variables, demonstrating the advantage of PCA in improving model efficiency and predictive accuracy. Among the models, the PCA-SVR approach achieved the best performance for
prediction, while the PCA-MLR performed best for
. These findings differ from some previous studies, where PCA-XGBoost models showed superior performance [34,44]. It’s highlighted that model effectiveness may vary depending on dataset characteristics and regional conditions. Overall, this study emphasizes the importance of addressing high-dimensionality and multicollinearity in environmental datasets. The proposed approach demonstrates that applying PCA prior to modeling can enhance efficiency and maintain strong predictive performance. In this study, PCA was applied to the full dataset prior to obtain stable and consistent principal components across all observations. However, this approach may introduce data leakage. Preprocessing steps such as dimensionality reduction should ideally be fitted using only the training data and then applied to the test set [45,46]. Therefore, the reported model performance in this study may be slightly optimistic. Despite this limitation, the approach was chosen to preserve the global structure and stability of the principal components across stations which is important for interpretability and comparison. Future work should consider implementing a leakage-free pipeline by fitting PCA exclusively on the training dataset. By focus on evaluating these models during more consistent periods. Particularly, when the PM concentrations are highest in haze season (January–March). By applying PCA to reduce the number of input variables, the models can maintain strong predictive accuracy. This approach would enable the Bangkok Metropolitan Administration (BMA) to focus on a smaller set of key variables, reducing the time and cost associated with collecting a full range of air pollution data.
5. Conclusion
This research establishes a high-performance framework for urban air quality forecasting by demonstrating that strategic dimensionality reduction consistently outperforms high-dimensional data modeling. The study utilized a comprehensive multivariate air pollution dataset from 4 monitoring stations in Bangkok, Thailand. It encompasses 13 meteorological and air pollutants features (,
, CO,
,
,
, WS, WD, Temp, RH, Rain, and Pres). Only WD has transformed into two components (ZWD and MWD). To address data incompleteness, the MICE-PMM method was applied to impute missing on dataset. This process can ensure a continuous, high-fidelity input matrix. The application of PCA effectively mitigated multicollinearity, with the first 4 PCs being selected based on an eigenvalue larger than 1 and a cumulative variance threshold of over 60%. The PCs serve as the optimized predictor set for modeling. The model used to forecast
and
concentration separately; Standard machine learning models such as MLR, XGBoost, SVR, and RF were considered. Results showing the reduced predictor models achieved significantly lower MAE, MSE, and MAPE compared to the 11 predictor suites. Notably, PCA-SVR and PCA-MLR emerged as the most accurate architectures for
and
forecasting respectively. This study underscores the necessity of dimensionality reduction when processing large-scale multivariate datasets, ultimately enhancing the efficiency of urban atmospheric modeling by reducing computational costs and analysis time.
Supporting information
S1 Data. Showing population intensity of many main cities in Thailand.
This data contains three columns, the rank, the main city names in Thailand, and population intensity. https://www.worldometers.info/world-population/thailand-population/
https://doi.org/10.1371/journal.pclm.0000949.s001
(CSV)
S2 Data. Included the air pollution dataset were used as main data in this study.
It contains 4 stations and each station had 12 variables with 2,310-day observations.
https://doi.org/10.1371/journal.pclm.0000949.s002
(XLSX)
S1 Table. Showing the imputation accuracy using Support Vector Regression (SVR), K-Nearest Neighbor (KNN), and Multivariate Imputation by Chained Equations (MICE) to impute the missing.
The imputed performance was evaluated by Root Mean Square Error (RMSE) in RU station.
https://doi.org/10.1371/journal.pclm.0000949.s003
(PDF)
S2 Table. Showing the imputation accuracy in HK station.
https://doi.org/10.1371/journal.pclm.0000949.s004
(PDF)
S3 Table. Showing the imputation accuracy in NV station.
https://doi.org/10.1371/journal.pclm.0000949.s005
(PDF)
S4 Table. Showing the imputation accuracy in DD station.
https://doi.org/10.1371/journal.pclm.0000949.s006
(PDF)
Acknowledgments
The authors would like to thank all contributors to this study. We gratefully acknowledge the Pollution Control Department of Thailand for providing the air quality data used in this research, and Worldometer for supplying the Bangkok population density data.
References
- 1. Southerland VA, Brauer M, Mohegh A, Hammer MS, van Donkelaar A, Martin RV, et al. Global urban temporal trends in fine particulate matter (PM2·5) and attributable health burdens: estimates from global datasets. Lancet Planet Health. 2022;6(2):e139–46. pmid:34998505
- 2. Ahmad M, Panyametheekul S, Thaveevong P, Ngamsritrakul T, Tassaneetrithep B, Supasri T, et al. Long-term monitoring of PM2.5 and PM10: Implications for air quality and public health in urban Bangkok, Thailand. Environmental Challenges. 2025;21:101312.
- 3. Pan L, Lu L, Fu P, Nitivattananon V, Guo H, Li Q. Understanding spatiotemporal evolution of the surface urban heat island in the Bangkok metropolitan region from 2000 to 2020 using enhanced land surface temperature. Geomatics, Natural Hazards and Risk. 2023;14.
- 4.
Wang G-M, Li W, Li Z. The factor analysis of PM2.5 calculating and source structure in Jiangsu province. 2015.
- 5. Choochuay C, Pongpiachan S, Tipmanee D, Deelaman W, Iadtem N, Suttinun O, et al. Effects of agricultural waste burning on PM2.5-bound polycyclic aromatic hydrocarbons, carbonaceous compositions, and water-soluble ionic species in the ambient air of Chiang-Mai, Thailand. Polycyclic Aromatic Compounds. 2020;42:1–22.
- 6. Wanishsakpong W, McNeil N, Notodiputro KA. Trend and pattern classification of surface air temperature change in the Arctic region. Atmospheric Science Letters. 2016;17(7):378–83.
- 7. Xu G, Ren X, Xiong K, Li L, Bi X, Wu Q. Analysis of the driving factors of PM2.5 concentration in the air: A case study of the Yangtze River Delta, China. Ecological Indicators. 2020;110:105889.
- 8. Patel P, Patel S, Shah K, Desai K, Patel S, Shah M, et al. A systematic study on PM2.5 and PM10 concentration prediction in air pollution using machine learning and deep learning model. Environmental Chemistry and Ecotoxicology. 2025;7:1401–15.
- 9. Kim JH. Multicollinearity and misleading statistical results. Korean J Anesthesiol. 2019;72(6):558–69. pmid:31304696
- 10. Jolliffe IT, Cadima J. Principal component analysis: a review and recent developments. Philos Trans A Math Phys Eng Sci. 2016;374(2065):20150202. pmid:26953178
- 11. Wetchayont P. Investigation on the impacts of COVID-19 lockdown and influencing factors on air quality in Greater Bangkok, Thailand. Advances in Meteorology. 2021;2021:1–11.
- 12. Adami G, Orsolini G, Fassio A, Viapiana O, Sorio E, Benini C, et al. POS0474 Factors Associated With Erosive Rheumatoid Arthritis, A Multimarker Principal Component Analysis (pca) And Principal Component Regression (PCR) Analysis. Annals of the Rheumatic Diseases. 2023;82:497–8.
- 13.
Katal A, Singh N. Artificial neural network: models, applications, and challenges. 2022:235–57.
- 14. Lim A, Ammatawiyanon L, Khurram H, Tongkumchum P, McNeil D. Spatial occurrence-intensity modeling of dengue incidence in southernmost provinces of Thailand. PLoS Negl Trop Dis. 2025;19(7):e0013347. pmid:40700458
- 15. Lim A, Owusu BA, Thongrod T, Khurram H, Pongsiri N, Ingviya T, et al. Trend and Association Between Particulate Matters and Meteorological Factors: A Prospect for Prediction of PM2.5 in Southern Thailand. Pol J Environ Stud. 2024;34(5):5215–23.
- 16.
Teerarojanarat S, Tingsabadh K. A geographical look at the word geography of Thailand. 2020.
- 17. Iamtrakul P, Padon A, Chayphong S. Quantifying the impact of urban growth on urban surface heat islands in the Bangkok metropolitan region, Thailand. Atmosphere. 2024;15.
- 18. Austin PC, White IR, Lee DS, van Buuren S. Missing Data in Clinical Research: A Tutorial on Multiple Imputation. Can J Cardiol. 2021;37(9):1322–31. pmid:33276049
- 19. Heymans MW, Twisk JWR. Handling missing data in clinical research. J Clin Epidemiol. 2022;151:185–8. pmid:36150546
- 20. Bernacki J. Particulate Matter Forecasting Using Hybrid Autoencoders: The Role of Meteorological Data. Annales Universitatis Mariae Curie-Sklodowska sectio B – Geographia Geologia Mineralogia et Petrographia. 2025;80:265–83.
- 21. Blazakis K, Perrakis S, Karapidakis E, Mavromatakis F, Katsigiannis Y. ML-Based Multi-Horizon Wind Speed and Wind Direction Forecasting for Aviation and Energy Applications in Coastal Crete. Electronics. 2025;14(24):4856.
- 22. Murphy E, Huang W, Bessac J, Wang J, Kotamarthi R. Joint Modeling of Wind Speed and Wind Direction Through a Conditional Approach. Environmetrics. 2025;36(3).
- 23. Gajbhiye S, Sharma SK, Awasthi MK. Application of Principal Components Analysis for Interpretation and Grouping of Water Quality Parameters. IJHIT. 2015;8(4):89–96.
- 24. Barakat A, El Baghdadi M, Rais J, Aghezzaf B, Slassi M. Assessment of spatial and seasonal water quality variation of Oum Er Rbia River (Morocco) using multivariate statistical techniques. International Soil and Water Conservation Research. 2016;4(4):284–92.
- 25. Allee KD, Do C, Raymundo FG. Principal Component Analysis and Factor Analysis in Accounting Research. Journal of Financial Reporting. 2022;7(2):1–39.
- 26. Pohlman J, Leitner D. A comparison of ordinary least squares and logistic regression. The Ohio Journal of Science. 2003;103.
- 27. Ahmad M, Al-Mansob R, Reza KK, Keawsawasvong S, Sabri M, Jamil I. Extreme Gradient Boosting Algorithm for Predicting Shear Strengths of Rockfill Materials. Complexity. 2022;2022:1–10.
- 28. Zhai M, Huyan Z, Hu Y, Jiang Y, Li H. Improvement of orbit prediction accuracy using extreme gradient boosting and principal component analysis. Open Astronomy. 2022;31(1):229–43.
- 29. Jiang Z, Gu J, Han Y, Fan W (David), Chen J. Modeling Actual Dwell Time for Rail Transit Using Data Analytics and Support Vector Regression. J Transp Eng, Part A: Systems. 2018;144(11).
- 30. Vincent B, Duhamel C, Ren L, Tchernev N. A PCA and SVR based method for continuous industrial process modelling. IFAC-PapersOnLine. 2018;51(11):1604–9.
- 31. Breiman L. Random Forests. Machine Learning. 2001;45(1):5–32.
- 32. Ali J, Khan R, Ahmad N, Maqsood I. Random Forests and Decision Trees. International Journal of Computer Science Issues. 2012;9.
- 33. Fernández JD, Menci SP, Lee CM, Rieger A, Fridgen G. Privacy-preserving federated learning for residential short-term load forecasting. Applied Energy. 2022;326:119915.
- 34.
Moursi A, Shouman M, Hemdan EE-D, El-Fishawy N. PM2.5 Concentration Prediction for Air Pollution Using Machine Learning Algorithms. 2019.
- 35. Dahari N, Latif MT, Muda K, Hussein N. Influence of Meteorological Variables on Suburban Atmospheric PM2.5 in the Southern Region of Peninsular Malaysia. Aerosol Air Qual Res. 2020;20(1):14–25.
- 36. Sirithian D, Thanatrakolsri P. Relationships between Meteorological and Particulate Matter Concentrations (PM 2.5 and PM 10) during the Haze Period in Urban and Rural Areas, Northern Thailand. Air, Soil and Water Research. 2022;15.
- 37. Huang R-J, Wang Y, Cao J, Lin C, Duan J, Chen Q, et al. Primary emissions versus secondary formation of fine particulate matter in the most polluted city (Shijiazhuang) in North China. Atmos Chem Phys. 2019;19(4):2283–98.
- 38. Ketjalan A, Humphries U, Suadee W. A study on PM2.5 concentration in Bangkok, Thailand: A case study of Bang Na Station. Int J Adv Appl Sci. 2023;10(10):55–61.
- 39. Xie Y, Zhao B, Zhang L, Luo R. Spatiotemporal variations of PM2.5 and PM10 concentrations between 31 Chinese cities and their relationships with SO2, NO2, CO and O3. Particuology. 2015;20:141–9.
- 40. Energy, Environment PI, Lit-Ken T, Sie-Meng O, Quen L, Yee-Siang G. Principal Component Analysis on Meteorological Data in UTM KL. Progress in Energy and Environment. 2017.
- 41. Oanh NT, Bich TL, Tipayarom D, Manadhar BR, Prapat P, Simpson CD. Characterization of particulate matter emission from open burning of rice straw. Atmospheric Environment. 2011;45(2):493–502.
- 42. Wimolwattanapun W, Hopke PK, Pongkiatkul P. Source apportionment and potential source locations of PM2.5 and PM2.5–10 at residential sites in metropolitan Bangkok. Atmospheric Pollution Research. 2011;2(2):172–81.
- 43. Hoque M, Ashraf Z, Kabir M, Nasrin S. Meteorological influences on seasonal variations of air pollutants (SO2, NO2, O3, CO, PM2.5 and PM10) in the Dhaka megacity. American Journal of Pure and Applied Biosciences. 2020;2.
- 44. Dukuly LP, Gupta M, Ghani S, Akram W. PCA-integrated machine learning framework for predicting rock fragmentation in blasting operations. Multiscale and Multidiscip Model Exp and Des. 2025;8(9).
- 45. Correa J, Tavakoli H, Vogel S, Gebbers R. Overfitting due to data leakage in soil sensor calibration: Examples from lab-based and in-situ soil NIR spectroscopy. Computers and Electronics in Agriculture. 2025;239:110920.
- 46. Rosenblatt M, Tejavibulya L, Jiang R, Noble S, Scheinost D. Data leakage inflates prediction performance in connectome-based machine learning models. Nat Commun. 2024;15(1):1829. pmid:38418819