The Theory And Method Of Geographically And Temporally Neural Network Weighted Regression

Regression analysis of geographical relations is a hot topic in the study of space-time modelling.Developing new spatio-temporal regression methods to improve the capabilities of geographical analysis and data mining is significantly important for the understanding of social processes and geographical phenomena.Spatio-temporal non-stationarity is an intrinsic property of geographical relations and its estimation is the key to the modelling of non-stationary relationships.Existing non-stationary modelling methods,however,have disadvantages such as inadequate expression of space-time proximity,difficulties in the complicated construction of weighting kernels,and incapability of the accurate estimation of non-stationarity.To resolve the above-mentioned problems,artificial intelligence is introduced in this thesis for its superior computing and fitting capacities.In general,we aim to establish an innovative theory of geographically and temporally neural network weighted regression(GTNNWR)for the accurate modelling of non-stationary relationships.The theory was applied in the modelling of the red tide disaster in Zhejiang coastal waters and the experimental results verified its effectiveness.The main contents of this thesis are as follows:(1)A multi-level space-time proximity deep neural network is constructed to unify the expression of space-time proximity and a geo-weighted neural network is proposed to accurately estimate spatio-temporal weights.We also develop an efficient training framework for the estimation of spatio-temporal non-stationarity and designed a series of statistical diagnostic methods for the examination of models.Accordingly,a new theory of GTNNWR is established.(2)Based on the theory of GTNNWR,a geographically neural network weighted regression(GNNWR)model is proposed for the modelling of spatial relations.After experimentation of the spatial non-stationary relationship modelling of the red tide disaster in Zhejiang coastal waters,the feasibility and efficiency of the GNNWR model are confirmed.In addition,the results further indicate that the spatial weighted neural network(SWNN)is more powerful than the spatial kernel of the traditional geographically weighted regression(GWR)model.(3)Based on the theory of GTNNWR,a GTNNWR model is put forward for the modelling of spatio-temporal relations.After experimentation of the spatio-temporal non-stationary relationship modelling of the red tide disaster in Zhejiang coastal waters,the feasibility and efficiency of the GTNNWR model are also verified.Furthermore,the results demonstrate that the spatial and temporal proximities neural network(STPNN)can effectively handle the non-linear fusion effects of temporal proximity and spatial proximity,and considerably enhance the performance of the spatio-temporal non-stationarity estimation.(4)Based on the theory of GTNNWR,a generalized geographically and temporally neural network weighted regression(GGTNNWR)model is proposed for the modelling of complicated spatio-temporal relations.After experimentation of the complicated spatio-temporal non-stationary relationship modelling of the red tide disaster in Zhejiang coastal waters,the feasibility and efficiency of the GGTNNWR model are confirmed.Moreover,the capability of spatial and temporal proximities deep neural networks(STPDNN)for fitting the complex space-time proximity of geographical processes is fully evaluated.In summary,this thesis expects to make theoretical innovations and method breakthroughs in the modelling of non-stationary relationships,and promote the development of spatio-temporal statistical methods.