摘要
土地利用回归(LUR)模型是目前模拟PM_(2.5)分布的常用方法,但该方法以多元线性回归进行建模,未考虑解释变量与PM_(2.5)浓度的非线性复杂关系,且易出现多重共线性。为提高模拟的准确性,该文采用随机森林(RF)算法和回归克里金插值法构建RF-LUR模型,并对长江三角洲PM_(2.5)空间分布进行模拟。研究表明:1)RF-LUR模型的拟合效果较好,修正后的模型预测值与观测值之间的R^2为0.942,经交叉检验,修正的RF-LUR模型相比LUR模型,MAE及RMSE分别减小43.0%、46.1%,IA值增加15.0%,相比普通克里金插值法,MAE及RMSE则分别减小40.8%、47.3%,IA值增加18.4%;2)2015年长江三角洲PM_(2.5)年平均浓度呈现北高南低、西高东低的总体格局,部分地区高值集聚,具有连片分布的特点;3)RF-LUR模型考虑了变量间的非线性关系,更适合模拟区域PM_(2.5)空间分布。
Land Use Regression(LUR)model is a common method to simulate spatial distribution of PM_(2.5) concentration.Since it is established through multiple linear regression,the nonlinear relationship between explained variables and PM_(2.5) concentration is ignored.In addition,it′s easy to have multiple co-linear relationships.To improve the accuracy of the simulation,in this research,Random Forests(RF)algorithm and regression Kriging method were used to build a RF-LUR model.Based on the model,the distribution of PM_(2.5) concentration in Yangtze River Delta was simulated.The research showed that:1)the RF-LUR model could provide a better fitting.Compared to LUR,the MAEand RMSEresulted from RF-LUR were reduced by 43.0%and 46.1%,respectively.Besides,the IA was increased by 15.0%.Compared to Ordinary Kriging,the MAEand RMSEresulted from RF-LUR were reduced by 40.8% and 47.3%,respectively,and the IA was increased by 18.4%.2)The spatial distribution of PM_(2.5) annual averaged concentration in Yangtze River Delta in 2015 presented an overall pattern of high values in the north and west,and low values in the south and east.Moreover,it also showed a characteristic of high value cluster in some areas with a continuous distribution.3)Since the nonlinear relationship was taken into consideration,RF-LUR was more suitable for the simulation of the spatial distribution of PM_(2.5) concentration in the regional scale.
出处
《地理与地理信息科学》
CSCD
北大核心
2018年第1期18-23,共6页
Geography and Geo-Information Science
基金
国家理科基地科研训练及科研能力提高项目(J1310028)
