Machine learning maps Thailand’s aerosol pollution decline as incomes rise

Researchers Thanakhom Srisaringkarn and Kentaka Aruga applied four machine learning algorithms to provincial satellite data spanning 2000 to 2023 and found an inverted U-shaped relationship between aerosol pollution and economic growth, the classic signature of the Environmental Kuznets Curve.

The study, published in Discover Atmosphere, is the first to analyze regional heterogeneity in Thailand’s aerosol pollution using machine learning forecasting combined with the Environmental Kuznets Curve framework. Rather than relying on official government monitoring stations, the authors drew their pollution data from satellite image processing on the Google Earth Engine platform, covering 76 of Thailand’s 77 provinces. Buengkan province was excluded because its dataset before 2011 was unavailable. The satellite-derived aerosol measure, which serves as a proxy for dust, smog, PM2.5 and PM10, has a reported 67% accuracy in reflecting real-world ambient air conditions in the study areas.

Aerosol pollution is a chronic annual problem in Thailand, driven by human activity and climate conditions, with severe consequences for public health and the economy.

Four algorithms competed on the forecasting task: Linear Regression, Multivariate Adaptive Regression Splines, Random Forest and Support Vector Regression. The dataset was split chronologically, with 2000 to 2018 used for training and validation and 2019 to 2023 reserved as unseen test data, a period deliberately excluded from forecasting because the COVID-19 pandemic produced outlier conditions. Hyperparameters were tuned with a Grid Search method before evaluation using R², MAE, MSE and RMSE. Random Forest emerged as the most robust performer, achieving an R² of 0.729 on the test set with region features included. Linear Regression actually scored slightly higher on the test period alone, at 0.748, but across the train and validation periods its error metrics were consistently worse than Random Forest’s. Support Vector Regression followed at 0.703 and MARS at 0.685.

Region features proved decisive. When the models were run without them, Random Forest’s test R² dropped from 0.729 to 0.657, and every algorithm performed significantly worse. The authors conclude that region-specific characteristics directly affect forecasting accuracy, reflecting how topography, environmental conditions and economic activity differ across the country. The Random Forest configuration itself was aggressive: trees grew until each leaf contained a single observation, the sample fraction was set at 0.8, bootstrap sampling was disabled, and only two randomly chosen predictors were evaluated at each split. These settings allow the algorithm to handle mixed discrete and continuous inputs, perform automatic variable selection and remain robust to outliers.

To interpret what the winning model had learned, the researchers turned to Shapley Values and Partial Dependence Plots, calculated on the 2000 to 2018 data to avoid pandemic-era distortions. Land surface temperature at daytime and population density, by contrast, push pollution upward.

Shapley analysis revealed that both the log of GPP per capita and its squared term have positive effects on aerosol pollution when their values are low but negative effects when they are high, producing the inverted U-shaped pattern that defines the Environmental Kuznets Curve. Partial dependence plots from both the Linear Regression and Random Forest models confirmed the shape. The turning point was identified at a log GPP per capita of 11.492, equivalent to international $9,335 adjusted by Thailand’s purchasing power parity at 2024, using the LOESS smoothing method. The authors caution that the bivariate statistical evidence carries relatively low explanatory power, with R² values of 0.1810 and 0.2010, because climatic factors and regional heterogeneity generate substantial variance in the relationship.

The forecast through 2050, chosen as the endpoint because Thailand targets net-zero greenhouse gas emissions by that year, shows a diminished downward trend in most regions. The Northeastern region declines the most, by −0.0569% to −0.0990% annually with an approximate average of −0.0826% per year, followed by the Northern region at an average of −0.0749% and the Bangkok-led region at −0.0679%. The Random Forest model’s average annual growth rate tracks closest to the overall average across all algorithms. Yet the authors warn against complacency: the annual rate of decline is marginal, less than 1%, meaning aerosol pollution will remain a persistent issue without appropriate mitigation strategies and could intensify in the future. They also note that forecasting accuracy decreases as the horizon lengthens, so models must be updated continuously with new data. Because machine learning cannot report statistical significance for individual features, the authors recommend pairing their approach with econometric models or causal machine learning methods such as Causal Forests in future work, alongside analysis at the district or sub-district level where topographic and environmental differences can be captured more precisely.

Subject of Research: Machine learning analysis of the relationship between aerosol pollution and economic growth in Thailand using the Environmental Kuznets Curve framework

Article Title: Aerosol pollution and economic growth in Thailand with a machine learning analysis of the environmental Kuznets Curve

Article References: Srisaringkarn, T., & Aruga, K. (2026). Aerosol pollution and economic growth in Thailand with a machine learning analysis of the environmental Kuznets Curve. Discover Atmosphere, 4(1), Article 24. https://doi.org/10.1007/s44292-026-00092-8

Image Credits: AI Generated

DOI: 10.1007/s44292-026-00092-8

Keywords: aerosol pollution, Thailand, environmental Kuznets Curve, machine learning, Random Forest, air quality, economic growth, SHAP values, satellite data, PM2.5, forecasting, green finance