Intercomparison of global ground-level ozone datasets for health-relevant metrics

Ground-level ozone is a significant air pollutant known for its detrimental effects on human health and agricultural productivity. Several methods have been used to estimate ground-level ozone concentrations globally, including chemical reanalyses, geostatistical methods, and machine learning, but the similarities and differences in these datasets have not been analyzed systematically. We compare six ground-level ozone datasets produced by different methods against one another and relative to observations, considering ozone long-term trends and spatial distribution. Specifically, we compare three global chemical reanalysis datasets, two machine learning datasets, and one geostatistics dataset, for the ozone season daily maximum 8-hour average mixing ratio (OSDMA8), focusing on 2006 to 2016. The results show that there are significant differences among these datasets in global average ozone estimates as large as 5-10 ppb, as well as in multi-year trends and regional distributions. In East Asia, all datasets show an increasing trend in ozone, in North America, five datasets show a decreasing trend, and in Europe, the three chemical reanalysis datasets show an increasing trend while the other datasets show no increasing trends. Among the six datasets, the population exposed to over 50 ppb of ozone varies widely: from 60.8% to 99% in East Asia, 17% to 88% in North America, and 9% to 77% in Europe (average OSDMA8, 2006–2016). These differences are large enough to cause important differences in applications, such as estimating the ozone burden on health, evaluating social inequities and planning public health policies. We also evaluate the performance of each dataset with respect to ground-level observations from the latest Tropospheric Ozone Assessment Report (TOAR) II dataset. From 2006 to 2016, most of the gridded datasets overestimate ozone, particularly at lower observed concentrations. Although some datasets perform well in terms of the monthly average MDA8, they may be less reliable for the yearly OSDMA8 metric. In 2016, across all stations, R² values range among the six datasets from 0.35 to 0.63, and RMSE values from 5.28 ppb to 13.49 ppb. Performance further declines when considering only stations with observations above 50 ppb. Regionally, the R2 is from 0.21 to 0.63 in Europe and from 0.23 to 0.71 in the United States, while only below 0.1 in Japan and South Korea. It is not clear to what extent differences between datasets result from differences in methods or input data. Although some databases share some of the same input data, we find that important differences exist among these datasets, with considerable uncertainties in yearly estimates, and highlighting the importance of continued research on global ozone distributions.