The Method Of Global Spatial-Temporal Complex Proximity Neural Network Weighted Regression

Spatiotemporal regression model is a hot topic in the research of spatial-temporal modeling.The development of new spatiotemporal regression model is a practical significance for revealing the spatial-temporal distribution and occurrence of geographic phenomena.The spatial-temporal proximity is one of the core contents of spatial-temporal non-stationarity relationships.However,due to the diverse and complex relationships between time and space,it can't adequately express the spatial-temporal proximity.As a result,there are some limitations in the accurate construction of proximity in the spatial-temporal regression model.To resolve the mentioned problems,the global spatial-temporal proximity expression theory that geographic attribute space is fused in this thesis.We establish a global spatial-temporal proximity cube.Then we propose an innovative theory of global spatially and temporally convolutional neural network weighted regression for accurate modelling of spatial-temporal proximity.The theory is applied in the modelling of national air pollution disaster to conduct method verification.The research is summarized as follows:(1)Global spatial-temporal proximity theory is proposed.Completed the fusion expression of the proximity relationship in three dimensions of space,time and spatial attribute space.Break through the proximity relationship measure of the sample space,and build the spatial-temporal proximity cube.To achieve a unified expression of the spatial-temporal complex proximity relationship.(2)We propose a global spatial-temporal complex proximity neural network weighted regression theory,which combines a three-dimensional convolutional neural network and a training distributed framework to construct an accurate expression of global spatial-temporal proximity.We develop a general training framework and design statistical diagnostic analysis methods.And we combine the existing neural network weighted regression model with the global spatial-temporal proximity theory.(3)We propose global spatial convolutional neural network weighted regression(GSCNNWR)model and global spatial neural network weighted regression(GSNNWR)model to estimate spatial proximity.After conducting experiments on the spatial non-stationary relationship of the air pollution disaster of China area,the efficiency and feasibility of the GSCNNWR and GSNNWR model are also verified.(4)We propose global spatial and temporal convolutional neural network weighted regression(GSTCNNWR)model and global spatial and temporal neural network weighted regression(GSTNNWR)model to estimate spatial-temporal proximity.In order to reveal the spatial-temporal non-stationary relationship of the air pollution disaster of Jingjinji area.We conducted experiments on the GSTCNNWR and GSTNNWR models and verified the accuracy of the models.It proved that the global spatial-temporal proximity theory can accurately reveal the spatial-temporal proximity relationship between geographical elements.In summary,this thesis expects to make method breakthroughs and theoretical innovations in the modelling of spatial-temporal proximity,and develope a new spatialtemporal statistical method.