LAUSR

Abstract Ground-level ozone (O3) is a primary air pollutant, which can greatly harm human health and ecosystems. At present, data fusion frameworks only provided ground-level O3 concentrations at coarse spatial (e.g., 10 km) or temporal (e.g., daily) resolutions. As photochemical pollution continues increasing over China in the last few years, a high-spatial–temporal-resolution product is required to enhance the comprehension of ground-level O3 formation mechanisms. To address this issue, our study creatively explores a brand-new framework for estimating hourly 2-km ground-level O3 concentrations across China (except Xinjiang and Tibet) using the brightness temperature at multiple thermal infrared bands. Considering the spatial heterogeneity of ground-level O3, a novel Self-adaptive Geospatially Local scheme based on Categorical boosting (SGLboost) is developed to train the estimation models. Validation results show that SGLboost performs well in the study area, with the R2s/RMSEs of 0.85/19.041 μg/m3 and 0.72/25.112 μg/m3 for the space-based cross-validation (CV) (2017–2019) and historical space-based CV (2019), respectively. Meanwhile, SGLboost achieves distinctly better metrics than those of some widely used machine learning methods, such as eXtreme Gradient boosting and Random Forest. Compared to recent related works over China, the performance of SGLboost is also more desired. Regarding the spatial distribution, the estimated results present continuous spatial patterns without a significantly partitioned boundary effect. In addition, accurate hourly and seasonal variations of ground-level O3 concentrations can be observed in the estimated results over the study area. It is believed that the hourly 2-km results estimated by SGLboost will help further understand the formation mechanisms of ground-level O3 in China.