【摘要】 采用随机森林（RF）和地理加权随机森林（GWRF）对长三角地区2003～2019年的PM_2.5浓度及其驱动因素数据进行训练、校验与测试,并探讨它们之间的关系.结果表明:（1）相比RF模型,GWRF模型对PM_2.5浓度的训练和预测更优,其各项模型评估指标均优于RF模型,且GWRF模型残差的空间自相关性更低.（2）GWRF模型预测2019年PM_2.5浓度分布优于RF模型,与实际观测浓度分布基本一致,但两个模型均存在北部高估南部低估的情况,且高估区域大于低估区域.（3）RF模型在研究PM_2.5浓度分布最重要且显著的驱动因子方面是全局性的,而GWRF模型得到干旱、气温、温差、风速以及人类干扰对PM_2.5分布的影响是局部性的.在大尺度下,这种局部性的效应对于PM_2.5精细化防控更具实际性的指导意义.此外,在全球气候暖干化和区域气候空间异质性的背景下,把干旱融入PM_2.5预测并建立具有局部效应的模型有助于环境监管机构及决策者制定防控措施.更多还原

【Abstract】 This study used Random Forest （RF） and Geographically Weighted Random Forest （GWRF） to train,verify,and test the PM_2.5 concentration and data of its driving factors in the Yangtze River delta region from 2003 to 2019 and explore their relationships.The results show that:（1） Compared with the RF model,the GWRF model performs better in training and predicting PM_2.5concentration.The evaluation indicators of the GWRF model are better,and the residual spatial autocorrelation is lower than the RF model.（2） The GWRF model predicts a better PM_2.5 concentration distribution in 2019 than the RF model,which is basically consistent with the actual observed concentration distribution.However,overestimation in the north,underestimation in the south,and the area of overestimation are more significant than the underestimation for the two models.（3） The RF model is global in studying the most significant driving factors of PM_2.5concentration distribution.In contrast,the GWRF model’s interpretation of PM_2.5 influence by drought,temperature,temperature difference,wind,and human interference shows a localized effect.On a large scale,this localized effect has practical guidance for preventing and controlling PM_2.5.In addition,in the context of global climate warming and spatial heterogeneity of regional climate,integrating drought into PM_2.5 prediction and establishing models with localized effects can help environmental regulatory agencies and decision-makers formulate prevention and control measures.更多还原