Long-term exposure to ambient ozone
(O3) can lead to
a series of chronic diseases and associated premature deaths, and
thus population-level environmental health studies hanker after the
high-resolution surface O3 concentration database. In response
to this demand, we innovatively construct a space–time Bayesian
neural network parametric regressor to fuse TOAR historical observations,
CMIP6 multimodel simulation ensemble, population distributions, land
cover properties, and emission inventories altogether and downscale
to 10 km × 10 km spatial resolution with high methodological
reliability (R
2 = 0.89–0.97, RMSE
= 1.97–3.42 ppbV), fair prediction accuracy (R
2 = 0.69–0.77, RMSE = 5.63–7.97 ppbV), and
commendable spatiotemporal extrapolation capabilities (R
2 = 0.62–0.76, RMSE = 5.38–11.7 ppbV). Based
on our predictions in 8-h maximum daily average metric, the rural-site
surface O3 are 15.1±7.4 ppbV higher than urban globally
averaged across 30 historical years during 1990–2019, with
developing countries being of the most evident differences. The globe-wide
urban surface O3 are climbing by 1.9±2.3 ppbV per
decade, except for the decreasing trends in eastern United States.
On the other hand, the global rural surface O3 tend to
be relatively stable, except for the rising tendencies in China and
India. Using CMIP6 model simulations directly without urban–rural
differentiation will lead to underestimations of population O3 exposure by 2.0±0.8 ppbV averaged over each historical
year. Our original Bayesian neural network framework contributes to
the deep-learning-driven environmental studies methodologically by
providing a brand-new feasible way to realize data fusion and downscaling,
which maintains high interpretability by conforming to the principles
of spatial statistics without compromising the prediction accuracy.
Moreover, the 30-year highly spatial resolved monthly surface O3 database with multiple metrics fills in the literature gap
for long-term surface O3 exposure tracing.