Multidecadal ozone trends in China and implications for human health and crop yields: A hybrid approach combining chemical transport model and machine learning

Surface ozone (O₃) poses great threats to both human health and crop production worldwide. However, a multi-decadal assessment of O₃ impacts in China is lacking due to insufficient long-term continuous O₃ observations. In this study, we used a machine learning (ML) algorithm to correct the biases of O₃ concentrations simulated by the chemical transport model from 1981–2019 by integrating multi-source datasets. The ML-enabled bias-corrected O₃ concentrations improve the estimates of O₃ impacts on human health and crop yields. The warm-season increasing trend of O₃ in Beijing-Tianjin-Hebei and its surroundings (BTHs), Yangtze River Delta (YRD), Sichuan Basin (SCB) and Pearl River Delta (PRD) regions are 0.32, 0.63, 0.84, and 0.81 μg m^–3 yr^–1 from 1981 to 2019, respectively. In more recent years, O₃ concentrations experience more fluctuations in the four major regions. Our results show that only BTHs have a perceptible increasing trend of 0.81 μg m^–3 yr^–1 during 2013–2019. The estimated annual all-cause premature deaths induced by O₃ increase from ~55,900 in 1981 to ~162,000 in 2019 with an increasing trend of ~2,980 deaths yr^–1. The annual premature deaths related to respiratory and cardiovascular disease are ~34,200 and ~40,300 in 1998, and ~26,500 and ~79,000 in 2019, having a rate of change of –546 and +1,770 deaths yr^–1 during 1998–2019, respectively. Using AOT40-China exposure-yield response relationships, the estimated relative yield losses (RYLs) for wheat, rice, soybean and maize are 17.6%, 13.8%, 11.3% and 7.3% in 1981, and increases to 24.2%, 17.5%, 16.3% and 9.8% in 2019, with an increasing rate of +0.03% yr^–1, +0.04% yr^–1, +0.27% yr^–1 and +0.13% yr^–1, respectively. Currently, estimating ozone-induced crop production losses still faces great uncertainties in magnitudes and/or spatial patterns when using different approaches, particularly in large-scale studies involving diverse ecological and climatic conditions. The averaged national annual mean RYLs for wheat are estimated to range from 4.3 to 24.6%, considering most available exposure metrics, including concentration-based and flux-based metrics. Our study, for the first time, used ML to provide a robust O₃ dataset over the past four decades in China, enabling a long-term evaluation of O₃-induced health impacts and crop losses. These findings are expected to fill the gap in the long-term O₃ trend and impact assessment in China.

Figure 1. Density scatter plots and linear regressions between O₃ measurements and predictions of LightGBM and GEOS-Chem model at (a1, a2) daily level and (a3, a4) hourly level, respectively. The annual averaged MDA8 O₃ concentrations of LightGBM bias-corrected predictions and corresponding anomalies from 1981 to 2019: (b1) BTHs, (b2) YRD, (b3) SCB, and (b4) PRD.  (c1) The mortality (thousand) for different health endpoints; (c2) The province-based mortality (thousand) attributed to different health endpoints; (c3) The annual province-based population (million). (d) Bar plot of the RYLs for crops using different metrics from 1981-2019: (left panel) LightGBM, and (right panel) GEOS-Chem. 

Figure 2. Spatial distribution of averaged annual RYLs (%) for wheat: (a) AOT40, (b) FBB, (c) DO₃SE_LRTAP, and (d) DO₃SE_Feng. The spatial correlation coefficients (r) of estimated RYLs using different metrics (Table).