Spatial Interpolation Based On Spatial Auto-regressive Neural Network And Its Verification Of Simulation And Measured Data

Spatial interpolation is one of the important methods of spatial analysis to predict unknown spatial data from known spatial data.The core of spatial interpolation is based on the calculation of spatial correlations and spatial weights.Traditional spatial interpolation methods generally calculate spatial weights based on Euclidean spatial distances,which has limitations of representing the complex non-linearity between spatial distances and spatial weights;the existing spatial interpolation methods based on neural network have improved spatial interpolation accuracy to some extent,but they are mostly optimized for existing methods,whose migration ability and adaptability need to be improved.For this reason,this thesis proposes a spatial autoregressive neural network,which makes use of the analysis and calculation abilities of neural network to realize the accurate calculation of spatial weights.To verify the accuracy and adaptability of the method in different dimensions and scenarios,this thesis applies SARNN to the interpolation of spatial simulation data and marine environment measured data which have different dimensions and characteristics.The main contents of this thesis are as follows:(1)To solve the problems that traditional spatial interpolation methods cannot fully express the spatial correlation,a generalized spatial distance calculation method which takes into account the spatial direction heterogeneity is proposed,and a neural network is used to fit the nonlinear relationship between the generalized spatial distance and the spatial weight.Then construct SARNN.For the problems of model training and optimization,the network training process and optimization algorithm are given,and a set of SARNN-based training framework is established,which breaks through the limitation of the traditional space distance lacks of the ability to represent the change trend of different directions in the multi-dimensional space,realizes the accurate calculation of spatial weights,and provides theoretical ideas and route support for spatial interpolation methods.(2)Based on the two-dimensional spatial simulation data and Zhejiang coastal phosphate data,a comparative experiment between SARNN and traditional spatial interpolation methods is designed.The results show that the method proposed in this thesis improves the interpolation accuracy,smooths the interpolation of the trend surface of interpolation result,and solves the problems of sawtooth phenomenon,"bull's eye" and transition bands in traditional methods.(3)Implement distance calculation and weight fitting in three-dimensional space,and design comparative experiments between SARNN and traditional spatial interpolation methods based on complex three-dimensional space simulation data and Argo measured temperature data.The results show that the method in this thesis has a good prediction fitting ability for the three-dimensional data with local discontinuity and strong mutation,and it is more accurate for the extreme point interpolation results,which has migration ability and adaptability in three-dimensional space.The SARNN method proposed in this thesis can more accurately predict the spatial distribution characteristics of the data and reflect the spatial distribution.It solves the problems such as "bull's eye" and sawtooth phenomenon in traditional interpolation methods,and has certain migration ability and adaptability,which provides a new idea in spatial interpolation and has certain scientific and practical significance.