Unmasking the sky: high-resolution PM_2.5 prediction in Texas using machine learning techniques.

Area of Science:

• Environmental Science
• Atmospheric Science
• Data Science

Background:

• Fine particulate matter (PM2.5) is a significant air pollutant in Texas.
• Limited regulatory monitoring stations hinder effective PM2.5 assessment.

Purpose of the Study:

• To develop a daily, fine-scale prediction model for PM2.5 concentrations in Texas.
• To integrate satellite-derived Aerosol Optical Depth (AOD) with meteorological and land-use data for enhanced PM2.5 prediction.

Main Methods:

• Utilized a comprehensive dataset from 2013-2017 in Texas, including ground-level PM2.5, MODIS AOD (1-km resolution), weather, and land-use data.
• Developed predictive models using gradient boosted trees and random forest algorithms, trained separately for each year.
• Evaluated model performance through in-sample and out-of-sample validation techniques.

Main Results:

• Models achieved excellent in-sample prediction accuracy, with R² values ranging from 0.81 to 0.97.
• Out-of-sample R² values were 0.44-0.75, indicating good predictive capability.
• Observed a decreasing trend in predicted PM2.5 concentrations in Eastern Texas over the study period.

Conclusions:

• Machine learning, particularly gradient boosting, effectively predicts PM2.5 concentrations in Texas.
• Satellite AOD and diverse environmental data improve spatial and temporal PM2.5 estimations.
• The developed models offer valuable tools for environmental studies and decision-making in areas with sparse monitoring.