【摘要】 针对PM2.5浓度估算中时空特征考虑不足和样本量较少的问题,该文将协同训练和时空地理加权回归相结合,提出了协同时空地理加权回归。采用两个不同参数的时空地理加权回归模型作为回归器,利用一个回归器训练另一个回归器的未标注样本,选择最优结果作为标注样本加入标注样本,通过不断学习扩大标注样本量提升模型的回归性能。以京津冀地区2015年3-7月的PM2.5浓度数据为实验数据,利用气溶胶光学厚度产品、温度、风速和相对湿度进行建模,采用不同核函数的时空地理加权回归作为对比方法进行实验。结果显示,协同时空地理加权回归性能比基于Gauss核函数时空地理加权回归提升了10%,比基于bi-square核函数时空地理加权回归提升了6.25%,证明该文方法能够提升时空样本数量不足时的PM2.5浓度估算精度。更多还原

【Abstract】 In view of the lack of spatial and temporal characteristics and the small sample size in the estimation of PM2.5concentration,a co-training geographically and temporally weighted regression(COGTWR)approach was proposed by combining co-training with geographically and temporally weighted regression(GTWR).Two different GTWR models were established by using the labelled data in the COGTWR.One model was used to train unlabelled data and some good results were chosen into another model.After being trained many times,the trained sample size was expanded,which was useful to improve the accuracy of regression.AOD,temperature,wind speed and relative humidity were used to estimate PM2.5concentration from March 2015 to July 2015 in Beijing,Tianjin and Hebei region.COGTWR was compared with GTWR.The result showed that COGTWR improved the performance by 6.25%,10%relative to GTWR by bi-square kernel function,GTWR by Gauss kernel function.It was proved that COGTWR could improve the accuracy of PM2.5concentration estimation when the number of spatio-temporal samples is insufficient.更多还原