基于随机森林模型的中国近地面NO2浓度估算

NH₃针对传统近地面NO₂浓度空间模拟过程中NO₂浓度与其影响要素之间关系的复杂非线性机制解释不充分的缺陷，本研究基于随机森林（RF）算法、融合多源地理要素开展了近地面NO₂浓度空间分布模拟研究.以卫星OMI对流层NO₂柱浓度数据和多源地理要素（道路交通、气象因子、土地利用/覆盖、地形高程、人口数量）为输入变量，近地面NO₂浓度为输出变量，利用RF算法构建近地面NO₂浓度反演模型.通过对比地面观测数据与传统土地利用回归模型（LUR）检验RF模型的有效性，基于所构建的最优RF模型在不同时间尺度下模拟分析中国大陆地区近地面NO₂浓度空间分布特征.结果表明：（1）集成多源地理要素的RF回归模型精度高，月均模型整体拟合度R² 0.85，RMSE 6.08μg/m³，交叉验证的R² 0.84，RMSE 6.33μg/m³，显著高于LUR模型（拟合R² 0.53，RMSE 10.48μg/m³，交叉验证的R² 0.53，RMSE 10.49μg/m³）； （2）地面NO₂浓度与预测变量呈现显著的复杂非线性与时间尺度依赖关系，卫星OMI柱浓度对模型影响程度最大，重要性指标IncMSE介于97.40%~116.54%，多源地理特征变量对RF模型同样具有不可忽视的贡献力（IncMSE在23.34%~47.53%之间）；（3）中国大陆地区NO₂污染程度较高，年均模拟浓度为24.67μg/m³，存在明显季节性空间差异，NO₂浓度冬季（31.85μg/m³） > 秋季（24.86μg/m³） > 春季（23.24μg/m³） > 夏季（18.75μg/m³），呈现以华北平原为高值中心、向外围逐渐减轻的空间分布格局.较已有研究揭示对流层NO₂柱浓度宏观分布特征，本研究对近地面NO₂污染特征的研究成果对于合理制定污染防控策略、降低居民暴露健康损害具有指导意义.

Estimating ground-level NO2 concentrations across mainland China using random forests regression modeling

In order to capture the complex and nonlinear relationship between ground-level NO₂ concentrations and predictor variables, random forest (RF) models combined with multiple types of geographic covariates were developed to estimate ground-level NO₂ concentrations. In this process, satellite-based OMI NO₂ tropospheric columns and multi-source geographic covariates (i.e., road network, meteorological factors, land use/cover, DEM and population density) were used as potential predictor variables and ground-level NO₂ concentrations were used as the dependent variable for RF models construction. The reliability of the RF models was validated by comparison with ground-measured NO₂ concentrations and typical linear land use regression (LUR) models. Afterwards, the spatial distribution characteristics of NO₂ concentration mapped by RF models across time scales in mainland China were assessed and analyzed. Results showed that RF modeling outperformed LUR modeling with obvious higher model fitting-based R² and lower RMSE, which were 0.85 and 6.08μg/m³ for monthly RF models compared with 0.53 and 10.48μg/m³ for LUR models. This was confirmed by the cross-validation-based R² and RMSE with values of 0.84 and 6.33μg/m³, while those of LUR models were 0.53 and 10.49μg/m³. The partial dependence of RF models suggested that the actual relationships between ground-level NO₂ concentrations and predictor variables were nonlinear and time-dependent. OMI NO₂ tropospheric columns contributed most strongly to the RF models of NO₂ concentrations, which had largest percentage of IncMSE (ranged from 97.40% to 116.54%). Meanwhile, the importance of different geographic variables could not be disregarded, which had values of IncMSE between 23.34% and 47.53%. Additionally, the NO₂ concentrations simulated by RF models showed that the annual average NO₂ concentrations across mainland China during the study period were 24.67μg/m³, which had significant seasonal variations with value of 31.85, 24.86, 23.24 and 18.75μg/m³ in winter, autumn, spring and summer, respectively. Spatially, higher concentrations of simulated NO₂ concentrations occurred in the North China Plain and decreased to the periphery. Compared with the existing studies focusing on tropospheric NO₂ column density, this study sheds new light on accurate monitoring of spatial-temporal distribution of ground-level NO₂ pollution. Findings from this study will provide new implications for policy making for future national prevention and control of air pollution to reduce the population health burden in China.