Research On Building Group Pattern Recognition For Cartographic Generalization Based On Deep Learning

Spatial pattern recognition of geographical elements,selection of cartographic generalization strategies and evaluation of results are important contents of map generalization,in which spatial pattern is one of the most important factors.As one of the important geographical feature,building group mode is the basis for the selection of map generalization strategy,and the consistency of building group mode before and after generalization is also one of the cartographic generalization quality evaluation criteria.The spatial pattern of building group is very complex,which is affected by the size,shape,direction of a single element in the building group,as well as the distance,arrangement pattern,density and other parameters of each feature.The map generalization quality evaluation of building group should be based on its pattern recognition,and consider the similarity and consistency of patterns in different scales.This paper introduces deep learning technology into building vector data processing,integrates Gestalt cognitive theory,graph theory,Delaunay triangulation and graph convolution neural network,tries to put forward a building group pattern recognition method,and based on this method,establishes a new building group cartographic generalization result evaluation model.The specific research contents are as follows:(1)Based on graph theory,Delaunay triangulation and building individual description parameters,this paper establishes a graph structure which can be applied to deep learning and can effectively express the spatial features and attribute information of building groups.The graph structure describes the individual characteristics of buildings through the size,direction,shape and other parameters,and describes the characteristics of building groups through the graph structure,which provides the data basis for the deep learning training of building group vector data.Based on this model,this paper compiles the calculation program of building group graph structure,and establishes the building group sample library which can be used for graph convolution neural network.(2)This paper presents a method of building group recognition based on deep learning.Based on the graph convolution neural network model in deep learning,we propose a spatial distribution pattern recognition model for building groups.The model aims at the building group data from 1:5000 to 1:50000.By learning the building group data from the input sample database,and constantly adjusting and modifying the parameters and input characteristics of the model,the recognition of regular pattern and irregular pattern of building group is realized.The accuracy reaches 97.45%,and the recognition effect is good.(3)Based on the pattern recognition model of building group,this paper proposes a map generalization evaluation method for building group pattern.In this method,the spatial pattern recognition model of building group is used as the evaluation tool,and the spatial pattern of building group is used as the evaluation index to evaluate the spatial pattern of building group qualitatively and quantitatively.The evaluation results provide the basis for the strategy selection of map generalization,and can be used to select the optimal map generalization strategy,which expands a new idea for the use of spatial knowledge in map generalization.