Research On Geographically Weighted Regression Considering The Global And Local Spatial-temporal Non-stationarity Difference

Geographically and temporally weighted regression(GTWR)is an effective approach to detect the spatial-temporal non-stationarity of spatial panels.A lot of research has been done which focus on the estimation,the spatial-temporal non-stationary test and application of GTWR,but lack of model constructing by selecting characteristic variables,the spatial-temporal stationary and non-stationary analysis,or the heteroscedasticity elimination.In anattempt to solve the above problems,we proposed the characteristic variables selection method based on GTWR,mixed geographically and temporally weighted regression(MGTWR),local polynomial geographically and temporally weight regression(LPGTWR).Besides,an application to estimate the PM2.5 concentration in the Beijing Tianjin and Hebei region was shown in this paper.The main research contents are as follows.?.The characteristic variables selection based on GTWR Because of the characteristic variables selection method of multivariate linear regression does not take into account the spatial-temporal non-stationary,it can not be directly applied to GTWR.In order to solve this problem,this paper proposes a characteristic variables selection method for GTWR,which is based on the greedy algorithm,stepwise regression,and Akaike information criterion(AIC).The experimental results show that the proposed mothed can choose the characteristic variables with strong correlation under the premise of considering the spatial-temporal non-stationarity.?.Mixed Geographically and Temporally Weighted Regression Geographically and temporally weighted regression can explore the spatial-temporal non-stationarity but without considering the global spatial-temporal stationarity.In order to solve this problem,we proposed a mixed geographically and temporally weighted regression.The two-stage least squares estimation method of MGTWR is given in part IV.The experimental results show that the MGTWR method can achieve better performance compared with the mixed geographically weighted regression and GTWR on the applicability,the coefficients estimation and the dependent variable fitting value.?.Local polynomial geographically and temporally weight regression In order to eliminate the heteroscedasticity of GTWR,we put foward local linear geographically and temporally weighted regression.The weighted least squares estimation method based on Taylor series expansion were proposed for LPGTWR.The experimental results show that compared with the local geographically weighe regression and GTWR,the proposed can not only reveal the spatial-temporal non-stationarity,but also can effectively eliminate the heteroscedasticity to improve the accuracy of analysis.?.The estimation of PM2.5 concentration in Beijing Tianjin and Hebei region The experiment of PM2.5 concentration estimation in Beijing Tianjin and Hebei region was made.We used the characteristic variables selection based on GTWR to construct the MGTWR and PLGTWR model.The MGTWR model was used to estimate the the monthly PM2.5 concentrations and the PLGTWR model was used to estimate the daily PM2.5 concentrations.Besides,we discussed the spatial distribution,spatial variation and seasonal variation of PM2.5 in Beijing Tianjin and Hebei region.