Using a land use regression model with machine learning to estimate ground level PM_2.5.

Area of Science:

• Environmental Science
• Air Quality Monitoring
• Spatio-temporal Modeling

Background:

• Ambient fine particulate matter (PM2.5) is a major global health risk factor.
• Accurate spatial-temporal modeling of PM2.5 is crucial due to limited monitoring stations.
• Existing methods require enhancement for high-resolution PM2.5 variation capture.

Purpose of the Study:

• To assess the spatial-temporal variability of PM2.5 using machine learning-based land use regression (LUR).
• To compare the performance of conventional LUR, Hybrid Kriging-LUR, and integrated machine learning algorithms.
• To identify the most effective model for estimating PM2.5 exposures in Taiwan.

Main Methods:

• Utilized daily PM2.5 data from 73 monitoring stations in Taiwan (2006-2016).
• Integrated diverse datasets including land use, meteorological, and environmental resources.
• Employed conventional LUR, Hybrid Kriging-LUR, deep neural network, random forest, and XGBoost algorithms.
• Validated models using data splitting, 10-fold cross-validation, and external, seasonal, and county-based verification.

Main Results:

• Conventional LUR and Hybrid Kriging-LUR models explained 58% and 89% of PM2.5 variations, respectively.
• Incorporating the XGBoost algorithm boosted explanatory power to 73% (LUR) and 94% (Hybrid Kriging-LUR).
• The Hybrid Kriging-LUR model combined with XGBoost demonstrated superior performance over other methods.

Conclusions:

• The Hybrid Kriging-LUR model integrated with the XGBoost algorithm effectively estimates PM2.5 spatial-temporal variability.
• This combined approach significantly enhances the accuracy of air pollution exposure assessment.
• The study highlights the value of advanced machine learning techniques in environmental monitoring.