利用近邻因子提高二氧化氮遥感反演浓度的精度-基于随机森林算法

Improving the accuracy of NO2 concentrations derived from remote sensing using localized factors based on random forest algorithm

NO_(2)是损害健康和破坏生态的主要大气污染物。本文基于NASA提供的AuraOMI遥感反演NO_(2)浓度,利用采样点8km内的经济、人口、路网和坡度数据,以及气象、植被和高程的点值数据,采用随机森林算法、地理加权回归(GWR)和多尺度GWR方法提高NO_(2)浓度的预测精度。NASA原浓度R2为0.48,以上三种模型把交叉验证R2分别提高到0.74、0.71和0.70,其中随机森林算法的精度最高,该算法的均方根误差(RMSE)和平均绝对误差(MAE)分别只有6.4μg/m^(3)和4.98μg/m^(3),且其速度远快于多尺度GWR,预测精度也高于大部分现有的同等范围研究。在浓度修正方面,局部化经济人口路网因子对预测精度提高的贡献至少为11.24%。此外,基于随机森林算法还给出全国县级城市NO_(2)浓度估计值的分布图。

NO_(2)is a main air pollutant that damages human health and ecological environment.Based on NASA's NO_(2)concentrations retrieved from Aura OMI,the prediction accuracy of NO_(2)concentration is improved in this work using the random forest algorithm,the Geographic Weighted Regression(GWR)and the Multi-scale GWR model respectively.Localized data of economy,population,road network and slope within 8 km of the sampling point,as well as the point values of meteorology,vegetation and elevation are used as correction variables in the models.It is found that the three models increase the cross validation R2 of NASA's concentrations,from original 0.48 to 0.74,0.71 and 0.70,respectively.Among the three models,the random forest algorithm is the most accurate one,with a low root mean square error(RMSE)of 6.4μg/m^(3)and a low mean absolute error(MAE)of 4.98μg/m^(3),and its speed is much faster than multi-scale GWR.In addition,the accuracy of random forest algorithm is also higher than that of most existing studies of similar extents.In terms of the concentration correction of NO_(2),it is found that the contribution of localized factors of economy,population and road network is at least 11.24%.In addition,based on the random forest algorithm,the distribution map of NO_(2)estimated concentration for county-level cities in China is also presented.