Predicting ambient PM_2.5 concentrations in Ulaanbaatar, Mongolia with machine learning approaches

Area of Science:

• Environmental Science
• Public Health
• Data Science

Background:

• Accurate assessment of individual ambient air pollution exposure is vital for epidemiological studies on adverse health effects.
• Developing countries face challenges in monitoring air quality due to sparse networks and inconsistent data.
• Fine particulate matter (PM2.5) poses significant health risks, necessitating reliable exposure assessment methods.

Purpose of the Study:

• To evaluate machine learning algorithms for predicting PM2.5 concentrations in Ulaanbaatar, Mongolia.
• To assess the feasibility and effectiveness of these models in a resource-limited, high-pollution urban environment.
• To provide robust data for epidemiological research on air pollution's health impacts.

Main Methods:

• Six machine learning algorithms were tested for predicting PM2.5 concentrations using data from 2010-2018.
• Model performance was evaluated using leave-one-location-out cross-validation and a hold-out test set.
• Random forest (RF) and gradient boosting models were identified as the top performers.

Main Results:

• RF and gradient boosting models achieved high predictive accuracy, with R² values up to 0.96.
• Spatiotemporal variations in predicted PM2.5 concentrations corresponded with emission sources and population density.
• The models demonstrated robust performance metrics, indicating reliable predictions.

Conclusions:

• Machine learning approaches are advantageous and feasible for predicting ambient PM2.5 levels in areas with limited resources and extreme pollution.
• The study provides a scalable method for estimating air pollution exposure in data-scarce regions.
• Findings support the use of ML in public health research to understand air pollution's health effects.