Application of satellite remote sensing data and random forest approach to estimate ground-level PM2.5 concentration in Northern region of Thailand

Wongnakae, Pimchanok; Chitchum, Pakkapong; Sripramong, Rungduen; Phosri, Arthit

doi:10.1007/s11356-023-28698-0

Cited by 13 publications

(4 citation statements)

References 71 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Other studies have also centered on estimating PM 2.5 concentrations based on the AOD of the MODIS-Terra platform, which provides a spatial resolution of 10 × 10 km 2 . These studies employed multiple or multivariate linear regression techniques for PM 2.5 concentration estimation (Kanabkaew, 2013;Wei et al, 2019;Amnuaylojaroen, 2022;Wongnakae et al, 2023). Another AOD product used for estimating particulate matter is the MAIAC product, offering a higher spatial resolution than other products.…”

Section: Introductionmentioning

confidence: 99%

Estimation of hourly one square kilometer fine particulate matter concentration over Thailand using aerosol optical depth

Punpukdee,

Winijkul,

Kyaw

et al. 2024

Front. Environ. Sci.

View full text Add to dashboard Cite

In the recent years, concentration of fine particulate matter that are 2.5 microns or less in diameter (PM2.5) in Thailand has consistently exceeded the national ambient air quality standard. Currently, the measurement of PM2.5 concentration relies on air quality monitoring stations operated by the Pollution Control Department of Thailand (PCD). However, these stations are insufficient, particularly in rural areas, where agricultural open burning are major sources of pollution after harvesting period. This study aims to enhance the monitoring of PM2.5 concentration by leveraging cost-effective technologies. We propose the integration of satellite data, specifically Aerosol Optical Depth (AOD) from Multi-Angle Atmospheric Correction (MAIAC) product and Himawari-8 satellites, with the Weather Research and Forecasting Model (WRF) data, to provide supplementary data to the ground-based monitoring. Hourly 5 × 5 km2 AOD data from Himawari-8 were downscaled to a high-resolution of 1 × 1 km2, leveraging the AOD distribution pattern of the concurrent MAIAC product using eXtreme Gradient Boosting (XGBoost) model. Notably, during Thailand’s rainy season (May to August), the study observed a relative reduction in the training model’s R-square value. This phenomenon is attributed to temporal discrepancies between Himawari-8 and the MAIAC products during this period. The predictive models of PM2.5 concentrations with the identification of pertinent variables through Pearson’s correlation analysis and recursive feature elimination, driven by the robust XGBoost model. Subsequently, the downscaled AOD, wind speed, temperature, and pressure were identified as predictors for the estimation of hourly PM2.5 concentration. This comprehensive approach enabled the projection of PM2.5 levels across Thailand, encompassing over 600,000 grids at 1 × 1 km2 resolution. The developed models, thus, offer a valuable tool for robust and high-resolution PM2.5 concentration estimation, presenting significant implications for air quality monitoring and management in Thailand.

show abstract

Section: Introductionmentioning

confidence: 99%

Estimation of hourly one square kilometer fine particulate matter concentration over Thailand using aerosol optical depth

Punpukdee,

Winijkul,

Kyaw

et al. 2024

Front. Environ. Sci.

View full text Add to dashboard Cite

show abstract

“…Early applications used AOD/PM2.5 ratios computed with a chemical transport model (CTM) to infer surface PM2.5 from observed AOD (Liu et al, 2004;van Donkelaar et al, 2006;van Donkelaar et al, 2021) but this may be affected by CTM biases. More recent applications have used machine learning algorithms to train satellite AODs on PM2.5 network measurements (Guo et al, 2021;Pendergrass et al, 2022;Wongnakae et al, 2023). Commonly used machine learning algorithms include XGBoost and Random Forest (RF), both based on decision trees, and neural networks; precision tends to be similar across algorithms (Di et al, 2019;Kulkarni et al, 2022).…”

Section: Introductionmentioning

confidence: 99%