A data-driven supervised machine learning approach to estimating global ambient air pollution concentrations with...

Area of Science:

• Environmental Science
• Data Science
• Atmospheric Chemistry

Background:

• Global air pollution monitoring relies on sparse, unevenly distributed stations.
• Temporal data gaps are common due to technical and power issues.
• Accurate, comprehensive air quality data is crucial for public health and environmental assessments.

Purpose of the Study:

• To develop a scalable, data-driven machine learning framework to address data gaps in air pollution monitoring.
• To generate a comprehensive global dataset of key air pollutants (NO2, O3, PM10, PM2.5, SO2) with high spatial and temporal resolution.
• To provide prediction intervals for each estimate to quantify uncertainty.

Main Methods:

• Developed a supervised machine learning framework for imputing missing air quality data.
• Trained models to estimate concentrations of NO2, O3, PM10, PM2.5, and SO2.
• Generated global concentration estimations at 261,377 locations with 0.25° spatial resolution and hourly intervals.

Main Results:

• Created a comprehensive global dataset of air pollutant concentrations.
• Provided hourly estimates with prediction intervals for over 260,000 locations worldwide.
• Examined model performance across diverse geographical regions.

Conclusions:

• The machine learning framework effectively imputes missing air quality data, creating a valuable resource.
• The high-resolution dataset supports detailed downstream assessments for various stakeholders.
• Analysis of performance provides insights for optimizing future air quality monitoring station placement.