Geographically Convolutional Neural Network Weighted Regression And Its Empirical Analysis In PM2.5 Modeling

Spatial non-stationary analysis is of great significance for revealing the spatial distribution of geographical elements and understanding social phenomena and environmental dynamic processes.The core of spatial non-stationary relationship modeling method is the expression of spatial proximity and the solution of spatial weight.The traditional spatial non-stationary relation modeling method based on mathematical analytical expression is restricted by the mathematical structure of the weighted kernel function and the solving method based on the weighted least squares.The spatial weight solving ability is limited and the spatial proximity expression is limited to the sample space;The geographically and temporally neural network weighted regression method?GNNWR?introduces a neural network to improve the spatial weight solving ability to some extent,but its spatial proximity expression is still limited to the sample space,and the expression adequacy is affected by the sample quality,resulting in low spatial non-stationary relationship modeling accuracy.To this end,this paper aims to break through the constraints of sample space,construct a global spatial proximity network to realize the expression of global spatial proximity;further fully utilizes the feature extraction and dynamic learning ability of convolutional neural networks to construct a spatially weighted convolutional neural networks,which achieve accurate solution of spatial weights and improve the modeling accuracy and applicability of spatial non-stationary relationship modeling methods.The main contents and results of the thesis are as follows:?1?Breaking through the restriction of spatial proximity expression in sample space by existing spatial non-stationary relationship modeling methods,a global spatial proximity network that expresses the global spatial proximity is constructed,and further.The spatial weighted convolutional neural network of considering the characteristics of the global spatial proximity network is designed and the statistical inference method are given,the framework of the geographically convolutional neural network weighted regression?GCNNWR?is formed to realize the innovation of the spatial non-stationary relationship modeling method.?2?Combining the existing excellent deep neural network training algorithm,the optimal training process and overall strategy of GCNNWR model are constructed.The activation function and parameter initialization method and neural network optimization training algorithm are designed.The training framework of GCNNWR model is built up to improve the performance of spatial non-stationary relationship modeling.?3?Taking the spatial non-stationary relationship model of the PM_2.5 concentration as an example,a comparative experiment of GCNNWR model and the classical regression analysis model was designed to verify the reliability and applicability of GCNNWR model.?4?Using GCNNWR model to complete the estimation of the 2017 quarterly and annual average PM_2.5 concentration of 3km resolution in China,and carry out the spatial scale analysis and time scale analysis of the distribution characteristics of PM_2.5 concentration in China.To a certain extent,the temporal and spatial variation of PM_2.5 concentration is revealed.