Diurnal Spatiotemporal Variations of Near-Surface Ozone Concentrations Across China

Ozone, an important atmospheric trace gas, predominantly resides in the stratosphere. It serves to safeguard human health and the Earth's environment by absorbing ultraviolet light. However, ozone near the surface acts as a significant atmospheric pollutant and a characteristic gas of urban photochemical pollution. This is mainly triggered by precursor pollutants such as nitrogen oxides (NO_x) and volatile organic compounds (VOCs), emitted by human activities. Therefore, the concentrations of near-surface ozone raise considerable concern because of their impacts on the atmosphere, human health, and even cash crops.

Many scientists have generated surface-level maps of O₃ concentration using a range of datasets, including outputs from Chemical Transport Models (CTMs), assimilated data (such as meteorological variables), and observations from remote sensing (such as O₃ column density). The majority of these studies utilized machine learning techniques to develop comprehensive estimation models by identifying the connections between multi-source data and in-situ measurements. However, traditional retrieval algorithms that depend on ozone absorption in ultraviolet (UV) radiation have only made the availability of optical satellite products in daytime, leaving a notable gap in understanding the spatio-temporal variations of ozone concentrations at night. In fact, several studies have identified a phenomenon known as nocturnal ozone enhancement (NOE). These events lead to increased ozone levels at night, thereby boosting the atmospheric oxidation capacity during nighttime hours. In some cases, NOE events have caused the ozone peak surpassing the daytime peak. Therefore, it is of significance to adopt a new framework to predict both daytime and nighttime ground-level O₃ concentrations with fine spatial and temporal resolutions.

Currently, the bright temperature in the thermal infrared band of satellites has become another potential choice for near-surface ozone estimation, as ozone also absorbs thermal infrared radiation in specific bands. To obtain high temporal resolution ozone concentration estimates, geostationary satellite observations with minute time intervals are chosen as the primary data source for implementing ML-based retrievals. In this work, hourly bright temperatures with a resolution of 0.05° in band 8 to band 16 provided by Advanced Himawari Images (AHI) of Himawari-8 (H8), along with other auxiliary data (e.g., solar zenith angle, satellite zenith angle, meteorology from ERA5 including 2-m temperature, relative humidity, surface pressure, boundary layer height, eastward/northward wind speed, and surface ozone mixing ratio, as well as the normalized difference vegetation index provided by MODIS, etc), are utilized to estimate diurnal near-surface ozone concentrations across China in 2020. The hourly ozone concentrations obtained from 2012 ground stations are used as the ground truth output for training. Multi-source data are matched around each station within a variable spatial window radius and a 0.5-hour time radius.

Cloud screening is essential to mitigate the impact of cloud cover on bright temperatures and only clear-sky grid samples are selected for model training and prediction. However, like most satellites, H8 uses optical channels to infer cloud pixels, thus only providing daytime cloud products. The extraction of nighttime cloud pixels becomes significant. Taking the daytime cloud product of H8 as the reference, we use its thermal infrared imagery, along with ERA5 meteorological parameters as inputs, to train an ensemble of boosted trees for diurnal cloud detection. The model is utilized with automatic hyper-parametric optimization. Optimization attempts to minimize the cross-validation loss (error) by varying the parameters until the optimal model parameters are found. The cloud detection process is conducted day by day. For January, a total of 31 models were generated with an average 5-fold cross-validation accuracy of 0.88.

After data alignment, a total number of about 500,000 matched samples during three time periods (i.e., daytime from 9 a.m. to 4 p.m., nighttime from 9 p.m. to 4 a.m., and dawn-dusk from 5 to 8 a.m. and 5 to 8 p.m.) are generated for each month. The random forest (RF) models are trained and then used to predict the surface ozone concentration at each grid for each hour at three time periods. The results show that the average cross-validation R² of the estimated model over all samples is 0.92 and that the daytime model outperforms the nighttime and dawn-dusk models, especially in June, July, and August. The estimated near-surface ozone concentrations show a consistent 24-hour pattern, with higher and more fluctuating daytime concentrations, with peaks around 16:00 LT, and lower and more stable nighttime concentrations, with minimums between 5:00 and 7:00 LT at dawn. 24-hour variations are most pronounced in May, June, and July. Although the highest national average ozone concentration is recorded during the daytime in summer (95.29 ± 20.4 μg/m³), the highest value at night occurred in spring (66.06 ± 10.8 μg/m³), which may be due to frequent horizontal and vertical mixing and ozone transport in spring.