Estimates
of ground-level ozone concentrations are necessary to
determine the human health burden of ozone. To support the Global
Burden of Disease Study, we produce yearly fine resolution global
surface ozone estimates from 1990 to 2017 through a data fusion of
observations and models. As ozone observations are sparse in many
populated regions, we use a novel combination of the M3Fusion and Bayesian Maximum Entropy (BME) methods. With M3Fusion, we create a multimodel composite by bias-correcting and weighting
nine global atmospheric chemistry models based on their ability to
predict observations (8834 sites globally) in each region and year.
BME is then used to integrate observations, such that estimates match
observations at each monitoring site with the observational influence
decreasing smoothly across space and time until the output matches
the multimodel composite. After estimating at 0.5° resolution
using BME, we add fine spatial detail from an additional model, yielding
estimates at 0.1° resolution. Observed ozone is predicted more
accurately (R
2 = 0.81 at the test point,
0.63 at 0.1°, and 0.62 at 0.5°) than the multimodel mean
(R
2 = 0.28 at 0.5°). Global ozone
exposure is estimated to be increasing, driven by highly populated
regions of Asia and Africa, despite decreases in the United States
and Russia.