Near-surface ozone (O₃) has become a major pollutant in some urban areas in recent years, especially in spring and summer (Liu et al., 2021). It is an important secondary pollutant that is mainly produced by a series of photochemical reactions that occur between nitrogen oxides (NOx) and volatile organic compounds (VOCs) that enter the atmosphere in the presence of sunlight (Chen et al., 2022). Transient or prolonged exposure to high ozone concentrations can cause asthma and respiratory and cardiovascular diseases, which can be fatal for certain populations (e.g., pregnant women, infants, and children) (Chan et al., 2006; Li et al., 2021). The 2030 agenda of the Sustainable Development Goals released by the United Nations clearly states that by 2030, proactive and effective measures should be developed to significantly reduce the number of deaths and illnesses caused by air pollution, including O₃ (United Nations, 2015).

Fine-scale information on the spatial distribution of near-surface ozone is needed to implement precise pollution prevention and control measures. Currently, ozone monitoring based on remote sensing satellites can compensate for the shortcomings of traditional air pollution monitoring stations, but the spatial resolution of existing ozone remote sensing product data or inverse ozone data is not fine enough to support studies on small regional scales (Yang et al., 2021). Downscaling techniques can effectively address this shortcoming, and statistical regression-based downscaling techniques have been increasingly used to obtain high-spatial resolution data. The statistical regression downscaling method assumes that the scale of the relationship remains constant; the fitted relationship between O₃ and various influencing factors does not change with the scale of remote sensing images (Zhan et al., 2010). Specifically, the fitted relationship between O₃ and the considered factors obtained at a low resolution can be used to estimate the O₃ concentration. Notably, deep learning-based downscaling methods are the latest techniques for downscaling lower-resolution images to match higher-resolution images. Compared with statistical regression downscaling methods, deep learning-based methods can approximate any relationship given two datasets, namely, they can effectively learn the relationships between high- and low-resolution pairs of images (Lanaras et al., 2018; Yang WM. et al., 2019a; Mukherjee and Liu, 2021). However, in some cases, it is difficult to obtain two datasets with different resolutions, and the statistical regression downscaling method becomes more practical.

Currently, downscaling methods based on statistical regression are widely applied to surface temperature (Yang C. et al., 2019b; Wu et al., 2019; Zang et al., 2020), precipitation (Gu et al., 2009; Tian et al., 2011), soil moisture (Sun et al., 2020), aerosol optical thickness (AOD) (Zhang et al., 2022), and other data. According to the different datasets used for regression modeling, downscaling methods can be divided into global and local models. Global models apply as many datasets as possible to build a global model for spatial downscaling analysis. For example, Fan et al. (2021) established a TRMM satellite precipitation downscaling model based on multiple linear regression (MLR) and obtained a high-resolution and high-precision satellite ground fusion precipitation product. Zhu JH. et al. (2021a) constructed a random forest downscaling model to achieve multilevel resolution surface temperature data. Wang et al. (2018) applied artificial neural networks to surface temperature downscaling and found that the results were more accurate than those derived from the traditional hierarchical linear regression downscaling method.

Researchers are increasingly reporting that the local characteristics of variables are critical to obtain more accurate relationships among spatial variables, and the spatial heterogeneity of the explanatory variables should be considered in downscaling analysis (Chen et al., 2014; Zhou et al., 2016). Notably, a local model based on part of the dataset can be utilized to describe the spatially varying relationships among the different variables. Duan and Li (2016) proposed a new downscaling algorithm based on geographically weighted regression that considers the geospatial variability among variables; the results showed that its performance was superior to that of the Ts HARP global model, significantly improving the spatial resolution of ground surface temperature data. Zhu XM. et al. (2021b) introduced multiscale geographically weighted regression to consider the scale differences among the covariate factors of the surface temperature mechanism; relatively accurate results were obtained.

Although local models that consider spatial heterogeneity are more effective than global models for downscaling spatial data, the occurrence of spatial dependency among spatial data suggests that incorporating spatial dependency into local models may further improve their downscaling accuracy. Actually, local models with integrated geostatistical methods based on spatial dependency (such as kriging) have exhibited outstanding performance in spatial modeling and prediction. For instance, Kumar et al. (2012) developed a geographically weighted regression kriging (GWRK) model to study the relationship between environmental variables and soil organic carbon stocks in Pennsylvania, United States, thus improving the accuracy of soil organic carbon stock estimation. However, existing research on downscaling has mainly focused on climatic and natural factors (surface temperature, precipitation, and biomass), whereas few studies have focused on the application of downscaling techniques to address air pollutant data. In addition, the ozone pollution state in small and medium urban areas is often neglected; larger urban areas with rapid economic development and more serious air pollution are more closely monitored.

Consequently, in this study, the Chang-Zhu-Tan urban agglomeration is selected as the study area, and a downscaling model for ozone remote sensing products is proposed based on the multiscale geographically weighted regression kriging (MGWRK) method. China High Air Pollutants (CHAP) ozone remote sensing data are selected as the data source, and ozone precursor emissions and their meteorological factors are used as covariates to achieve a downscaled ozone remote sensing product from 10 km × 10 km–1 km × 1 km based on the relational scale invariance.

Currently, downscaling is increasingly applied in the field of air pollution analysis; examples include the downscaling of AOD products (Li et al., 2020; Liang et al., 2022), estimation of the PM_2.5 concentration (Zhang and Pan, 2020), and generation of long-term hourly 0.25° global PM_2.5 datasets (Valencia et al., 2022). Existing statistical downscaling methods are used in the downscaling of O₃ remote sensing products. However, traditional statistical downscaling methods assume that the relationship between the dependent variable and the downscaling factor remains spatially constant (Zhan et al., 2018), ignoring the spatial heterogeneity in environmental variables and the variability of different environmental variables across different scales. Therefore, in this study, the MGWR method was proposed for the spatial downscaling of satellite-based ozone datasets by combining this method with kriging.

Compared with the MLRK model, the GWRK and MGWRK methods can capture the spatially varying relationships between the O₃ concentration and its associated variables. Notably, if spatial heterogeneity exists among the different variables, the MLRK and GWRK methods can yield better fitted results than the MLRK model; conversely, they cannot provide accurate prediction results. Furthermore, compared with the GWRK model, the MGWRK method can reveal multiple scale effects of different independent variables on the O₃ concentration. Moreover, if the scale effect does not exist, the prediction accuracy of the MGWR model may be similar to that of the GWRK model. The experimental results for the Chang-Zhu-Tan urban agglomeration, China, show that the downscaling accuracy of the MGWRK and GWRK models is higher than that of global regression (MLRK). This indicates that spatially varying relationships can be identified between the O₃ concentration and its associated variables in the case study. The MGWRK method achieves the best fitting accuracy on both annual and seasonal scales, which demonstrates that the consideration of multiple scale effects can more effectively capture the complex relationships in downscaling O₃ datasets.

Actually, both MGWR and the other two models assume linear relationships. This assumption determines their intrinsic inadequacy in characterizing complex non-linear relationships (Fotheringham et al., 2017). Notably, the MGWR model cannot effectively describe complex non-linear relationships among spatial variables. In this case, machine learning or deep learning is an alternative technique to address non-linear relationships. However, machine learning and deep learning usually require samples with independent and identical distributions, and they cannot manage spatial data with scale effects and spatial heterogeneity well (Yang et al., 2018b). Moreover, most machine learning models or deep learning models require many samples in the training process. Therefore, statistical regression downscaling methods, such as the MGWR model, cannot be replaced. The experimental results also prove that the statistical regression downscaling method can achieve a relatively satisfactory accuracy and is practical.

However, there are still certain uncertainties in the O₃ dataset that may lead to errors. First, the resampling of environmental variables from a higher to a lower spatial resolution may lead to smoothing effects. In this study, a downscaling model is constructed based on their relationship at the 10-km scale. When the environmental variables are resampled, bilinear interpolation replaces the value of each environmental variable with the average value of the surrounding neighboring pixels, thus losing some characteristic information of the variables and producing a relative smoothing effect. Second, a total of 68 air quality monitoring stations were used to validate the downscaling results in this study, and the sparsely distributed validation data may have affected the accuracy of the estimates. Finally, other environmental variables are neglected. There are other environmental variables that may also affect the spatial distribution of O₃, such as insolation, solar radiation, and DEM. In this study, only precipitation, humidity, temperature, wind speed, NDVI, AOD, and emission inventory data (NO_X and VOCs) are considered, which may lead to uncertainty in the downscaling results.

In this research, the MGWRK model was proposed for the spatial downscaling of satellite-based ozone datasets. This method considers the spatial dependency of variables and can be used to account for the spatially varying relationships between the O₃ concentration and its associated variables at different spatial scales. The method was assessed for scaling down the ozone remote sensing products in 2017 from a spatial resolution of 10 km × 10 km–1 km × 1 km in the Chang-Zhu-Tan urban agglomeration via comparisons with MGWR and MLRK. The MGWRK model achieved the most accurate predictions of ozone. Moreover, this study demonstrated that the MGWRK model, with the consideration of spatial effects, captured the complex relationships between the ozone concentration and its associated variables. This downscaling method can also be used for spatial downscaling of other spatial data products. Additionally, the derived 1-km spatial resolution annual and seasonal ozone remote sensing data products provide useful information for other relevant studies of the area, such as environmental and public health and air pollution prevention and control.