Deep ensemble machine learning with Bayesian blending improved accuracy and precision of modelled ground-level ozone for region with sparse monitoring: Australia, 2005–2018

Ground-level ozone (O 3 ) is a significant public health concern. We developed maps of monthly average 1-h maximum O 3 concentrations in New South Wales, Australia (2005–2018), a region with sparse monitoring. For the first time Bayesian Maximum Entropy (BME) blending was used within a Deep Ensemble Machine Learning (DEML) framework for air pollution predictions. The DEML combined geographical predictors in random forest (RF), extreme gradient boosting (XGBoost), and gradient boosted machine (GBM) models with three meta-models. BME blending incorporated observed O 3 data into posterior predictions. We generated 2.5 km × 2.5 km resolution gridded surfaces. The DEML estimates achieved an R 2 of 0.89 and RMSE of 2.3 ppb in the held-out test dataset at monitors. DEML grid cell predictions (R 2 : 0.84, RMSE: 3.03 ppb) were improved by BME blending (R 2 : 0.89, RMSE: 2.49 ppb). Mean bias reduced from −0.7 ppb to −0.4 ppb. This demonstrates high accuracy and precision in a sparsely monitored region. • This is the first use of Bayesian Maximum Entropy with Deep Ensemble modelling. • Staged DEML with BME achieved high accuracy and precision for ozone modelling. • The research focused on ground-level ozone in a region with sparse monitoring data. • Extreme Gradient Boosting was the best base model, with slight improvement from DEML. • BME blending enhanced predictions by integrating measurement data.