Land use regression model established using Light Gradient Boosting Machine incorporating the WRF/CMAQ model for highly accurate spatiotemporal PM2.5 estimation in the central region of Thailand

The level of fine particulate matter (PM2.5) in central Thailand has exceeded the national air quality standard in the dry season for the years. The limited number of monitoring stations makes it difficult to obtain spatiotem-poral PM2.5 distributions in the region. In this study, we first developed a land use regression (LUR) model using a non-linear machine learning (ML) algorithm incorporating the Weather Research and Forecasting and Com-munity Multiscale Air Quality (WRF/CMAQ) model to predict daily ambient PM2.5 levels in Thailand. We aimed to evaluate the PM2.5 estimation performance of the integrated LUR modeling approach in developing countries. The Light Gradient Boosting Machine (LightGBM) was used as the non-linear ML algorithm. CMAQ-simulated PM2.5 concentrations, WRF-simulated meteorological parameters, population density, and land use variables were used as predictors of the LUR model. A 5-fold site-based cross-validation (CV) technique was performed to evaluate the prediction performances of the LUR model using the coefficient of determination (R2). The daily PM2.5 concentrations were estimated by the LUR model at a 1-km grid resolution. Our LUR model exhibited a CV -R2 of 0.71. Moreover, the LUR model effectively illustrated PM2.5 distributions at a high spatiotemporal resolution over the central Thailand. Our findings demonstrate the advantages of the integrated LUR model for accurately estimating the daily ambient PM2.5 level, which is influenced by transboundary and local pollution. This modeling approach could be implemented for future air pollutant estimation in Southeast Asia and other developing countries.

Automated summary

In this study, a land use regression model was developed using a non-linear machine learning algorithm to predict daily PM2.5 levels in central Thailand. The model exhibited high prediction accuracy and effectively showed the spatiotemporal distribution of PM2.5.