Comparison of High Spatial Resolution PM_2.5, PM₁₀, and NO₂ Estimates Using a Deep Ensemble Machine Learning Framework...

Area of Science:

• Environmental science
• Data science
• Atmospheric chemistry

Background:

• Machine learning (ML) and ensemble models show promise for air pollution estimation.
• Evidence for their added value in sparsely monitored regions is limited.
• This study addresses this gap in New South Wales, Australia.

Purpose of the Study:

• To compare the performance of ML and ensemble models for estimating PM2.5, PM10, and NO2.
• To evaluate models in a sparsely monitored, low-pollution, large geographic region.
• To assess the impact of different cross-validation (CV) methods on model performance.

Main Methods:

• Developed and compared three ML models, a supervised linear regression (SLR) model, and a three-stage Disaggregated Ensemble Machine Learning (DEML) framework.
• Utilized 236 predictors including land use, population, traffic, and satellite data.
• Employed 10-fold CV and hold-out validation for model evaluation using R² and RMSE.

Main Results:

• DEML models provided the best fit for all pollutants (PM2.5, PM10, NO2).
• Improvements of DEML over base ML models were modest.
• Base ML models are valuable for low-pollution settings with heterogeneous monitoring density due to lower implementation costs.
• CV methods significantly impacted model performance.

Conclusions:

• While DEML models offer the best fit, basic ML models are a cost-effective alternative for air pollution estimation in sparsely monitored, low-pollution regions.
• Careful consideration of CV methods and setting constraints is crucial when selecting modeling approaches for air quality assessment.