Ground-level ozone (O 3 ) pollution poses significant risks to ecosystems and public health, particularly in China. Current satellite-based retrieval approaches largely rely on total-column ozone and precursor products derived from radiative transfer models, which can introduce uncertainties due to assumptions regarding surface and atmospheric conditions. To address these limitations, we developed a one-step retrieval framework to directly estimates ground-level O 3 concentrations from Ozone Monitoring Instrument (OMI) hyperspectral imagery using an Extremely Randomized Trees (ERT) model. The ERT was trained by integrating OMI spectral observations with ground-based O 3 observations, enabling automatic identification of informative wavelength features through a hybrid data-driven and physics-informed strategy. Independent validation demonstrates its strong predictive performance, with a high sample-based ten-fold cross-validation correlation coefficient (CV-R) of 0.92 and a low root mean square error (RMSE) of 13.55 μg/m 3 . It also exhibits robust spatial extrapolation capability (space-based CV-R = 0.91) and moderate temporal transferability (time-based CV-R = 0.76) for regions and periods without training samples, although performance slightly declines in densely vegetated areas and under elevated boundary layer conditions. Our analysis demonstrates the critical importance of OMI observations, with model accuracy decreasing by up to 15% when OMI data are excluded. Leveraging these advantages, the model effectively captures characteristic spatiotemporal O 3 patterns across China and successfully identifies severe summer pollution hotspots in the North China Plain (NCP) and Yangtze River Delta (YRD) regions, which are driven by distinct factors. This is a first demonstration of direct ground-level O 3 retrieval from OMI hyperspectral imagery without dependence on its secondary products, potentially establishing a new paradigm for satellite-based air quality monitoring. • A one-step machine learning model is developed to directly estimate O 3 from OMI imagery. • Hybrid data-physics feature selection identifies ozone-sensitive bands and enhances model interpretability. • The model achieves high accuracy across China with strong spatial extrapolation capability (CV-R = 0.91). • OMI hyperspectral observations effectively improve model performance by up to 15% in R and 36% in RMSE. • Model effectively reveal seasonal hotspots and regional drivers in the North China Plain and Yangtze River.
Sun et al. (Fri,) studied this question.