Simulation Of Land Use Change In Shanghai Based On Spatial Partitioning And Spatiotemporal Convolution

In recent decades,the acceleration of urbanization has promoted more and more scholars to study land use change patterns in order to provide support for urban planning and decision-making.Cellular Automata(CA)is widely used in the simulation of dynamic land use change(LUC)due to its simple,flexible and intuitive features.However,most previous studies have ignored the spatial heterogeneity between the sub-regions,trying to capture the evolution of land use by using the same conversion rules throughout the research area.In addition,most of the existing methods only use one data time slice to extract the neighborhood effect,while the neighborhood interaction is a long time-series geographic phenomenon with obvious temporal and spatial dependence;using only one data time slice will lead to the problem of insufficient information expression of neighborhood.In terms of the above problems,this paper proposes a land use change simulation model(PST-CA)that couples spatial partitioning,neighborhood spatiotemporal feature extraction and cellular automata.First of all,in terms of the spatial heterogeneity of urban land use changes,a self-organizing map(SOM)is proposed to divide the whole area to obtain homogeneous regions.Then a three-dimensional convolutional neural network(3D CNN)is implemented to capture the temporal and spatial dependence characteristics of neighborhoods in land use change.Shanghai was taken as the research area and we simulate the LUC process in Shanghai based on historical classification data and driving factors from 2000 to 2015.Compared with the traditional LR(Logistics Regression)-CA,SVM(Support Vector Machine)-CA,RF(Random Forest)-CA,ANN(Artificial Neural Network)-CA,the accuracy of the PST-CA model increased 4.66%?6.41%.In addition,the study found that there are three different built-up area "coverage rate-growth rate" models in each sub-areas;Moreover,the study found that the optimal time step of 3D CNN showed a relatively positive correlation with the growth rate of the built-up area.