19 Predicting daily PM2.5 in Mexico City: A hybrid spatiotemporal modeling approach

Abstract

Objectives/Goals: In recent years, there has been growing interest in the development of air pollution prediction models, particularly in low- and middle-income countries that are disproportionately impacted by the effects of air pollution. Recent methodological advancements, particularly in machine learning, provide novel opportunities for modeling efforts. Methods/Study Population: We estimate daily ground-level fine particulate matter (PM2.5) concentrations in the Mexico City Metropolitan Area at 1-km2 grids from 2005 to 2023 using a multistage approach. Spatial and temporal predictor variables include data from the moderate resolution imaging spectroradiometer (MODIS), Copernicus Atmosphere Monitoring Service (CAMS), and additional meteorological and land use variables. We employed machine-learning-based approaches (random forest and gradient boosting algorithms) to downscale satellite measurements and incorporate local sources, then utilized a generalized additive model (GAM) to geographically weight predictions from the initial models. Model performance was evaluated using 10-fold cross-validation. Results/Anticipated Results: On average, the random forest, gradient boosting, and GAM models explained 75, 82, and 83% of variations measured in PM2.5 concentrations. PM2.5 levels were generally higher in densely populated urban centers and lower in suburban and rural areas. Important predictors of ground-level PM2.5 included wind (both u and v components), 2-meter mean air temperature, elevation, and the normalized difference vegetation index (NDVI). Discussion/Significance of Impact: Using novel machine learning-based approaches, we developed robust models with fine-scale spatial (1-km2) and temporal (daily) variations of PM2.5 in Mexico City from 2005 to 2023. The predicted PM2.5 concentrations can further advance public health research on air pollution in Mexico City and beyond.