基于梯度提升回归树算法的地面臭氧浓度估算

将机器学习中的梯度提升回归树（GBRT）算法应用到中国地区地面O₃浓度制图中，利用地面O₃浓度观测数据，结合WRF气象数据、MODIS植被归一化指数以及高程人口数据建立训练预测数据集.通过反向变量选择法选取模型最佳特征变量对其进行训练，十折交叉验证结果：决定系数R²=0.89、均方根误差RMSE=4.75μg/m³.同时对全国O₃人口暴露水平进行评估.结果表明：在暴露强度上，我国人口加权O₃浓度值排在前5的省依次是山东、河南、江苏、河北、上海，均值浓度为94.48μg/m³.在暴露持续时间上，非达标天数最多的5个省依次是河南、山东、河北、宁夏、北京，一年内有42%的天数处于非达标的状态.

Estimation of ground-level ozone concentration based on GBRT

Gradient Boosting Regression Tree (GBRT) algorithm has been developed for application to the mapping of ground-based O₃ concentrations in China. In this study, observations of ground-based O₃ concentrations, WRF meteorological data, MODIS vegetation normalization index, elevation and population data were used to build a training prediction data set. The best feature variable of the model was selected using an inverse variable selection method. The results were evaluated using a ten-fold cross-validation method:the coefficient of determination R²=0.89and the root mean square error RMSE=4.75μg/m³. The model was used to evaluate the national O₃ population exposure level. In terms of exposure intensity, the provinces in China with the highest population-weighted O₃ concentration values are Shandong, Henan, Jiangsu, Hebei, and Shanghai, with an average concentration of 94.48μg/m³. In terms of exposure duration, the provinces with the highest number of non-compliant days are Henan, Shandong, Hebei, Ningxia, and Beijing, with an average percentage of 42% of thedays per year which are non-compliant.