Eighteen years of daily PM2.5 predictions (2005–2022) for a region of western Canada: Machine learning and satellite inputs for applications in rural health

Northeastern British Columbia is a rural and remote region in Western Canada that is experiencing rapid growth in unconventional oil and gas development (UOGD) and increasingly severe wildfire impacts. Air quality is a concern for the communities in the region, but there is very limited air pollution monitoring. To address this gap, this study explores the application of machine learning to the satellite-based estimates of aerosols and meteorology from the Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2) to estimate daily fine particulate matter (PM_2.5) at a spatial and temporal resolution relevant for ongoing health studies. The random forest model was trained and validated on the ground-level air quality monitoring network from 2013 to 2022 and then used to predict and backcast daily concentrations in the study area at a 50-km resolution from 2005 to 2022. The predictions were then compared to global and provincial health guidelines and analyzed for annual trends. Our model achieved a 10-fold cross validation root mean square error (RMSE) and R² of 3.89 μg/m³ and 0.77, and test scores of 3.02 μg/m³ and 0.78. Between 2006 and 2021, the number of days exceeding PM_2.5 guidelines increased by 122 %, and the person-days exceeding the guideline increased by 166 %. At our spatial resolution, we find that wildfire is a more important variable in predicting daily PM_2.5 concentrations compared to UOGD. We demonstrate the application of satellite reanalysis products to estimate ground-level PM_2.5 in a rural area of Canada with minimal ground monitoring stations.