Deep Ensemble Machine Learning Framework for the Estimation of Concentrations.

Area of Science:

• Environmental Science
• Data Science
• Machine Learning

Background:

• Accurate estimation of historical fine particulate matter (PM2.5) is critical for environmental health risk assessment.
• Understanding PM2.5 exposure is essential for public health policy and intervention.

Purpose of the Study:

• To develop a multiple-level stacked ensemble machine learning framework for improving the estimation of daily ground-level PM2.5 concentrations.
• To assess the performance of the developed framework against benchmark algorithms.

Main Methods:

• Developed a deep ensemble machine learning (DEML) framework with a three-stage structure involving base models (GBM, SVM, RF, XGBoost), meta-models (RF, XGBoost, GLM), and an NNLS algorithm.
• Implemented the DEML framework using data from 133 monitoring stations in Italy to predict daily PM2.5 concentrations from 2015 to 2019.
• Evaluated model performance using 10-fold cross-validation and compared it with five benchmark algorithms.

Main Results:

• The DEML framework demonstrated superior prediction performance compared to benchmark models, with higher coefficients of determination and lower root mean square error.
• DEML effectively captured the spatiotemporal variations of PM2.5 concentrations across Italy.
• The model's performance was validated through rigorous cross-validation and comparison with established algorithms.

Conclusions:

• The proposed DEML framework achieves outstanding performance in PM2.5 estimation.
• DEML can serve as a valuable tool for more accurate environmental exposure assessment.
• This advanced machine learning approach offers significant potential for improving air quality monitoring and health impact studies.