Pattern Recognition Of Building Linear Alignments Based On Graph Convolutional Neural Network

Spatial distribution patterns of buildings refer to the salient arrangements and forms that groups of buildings exhibit collectively in space,which can be recognized visually and named semantically.It contains abundant information of the structures and functions of the city region,which is of great significant in cartographic generalization,multiple expression of spatial data and spatial-temporal data mining.Building alignments patterns,which are the basic and important components of the distribution patterns of buildings,not only widely exist on maps,but also constitute other patterns as a basic unit.The identification and classification of building alignment patterns is a complicated problem.Firstly,because the selection of metric factors and the classification of linear alignment have an impact on the recognition and classification results.Secondly,the identification and classification of spatial object distribution structure and patterns are inherently uncertain and closely related to human visual cognition process,which puts forward higher requirements on the selection of recognition methods and the determination of thresholds.Therefore,in order to solve the problems that the classification of the building alignment patterns in the current research is not fine enough to fully meet the needs of retaining the spatial distribution characteristics in the process of cartographic generalization and multiple expression,and the feature factor weights and classification thresholds are difficult to determine in the pattern recognition method,this paper explored and studied the theory and method of extracting and pattern recognition the building alignment by intelligent computing,which aims at making pattern recognition results more in line with human cognition,and supporting for the cartographic generalization and multiple expression of the spatial data.The main contents of the research are as follows:1)Division of building alignments.Based on the analysis of the research status of building spatial distribution patterns,this paper raised a further division method for building alignment patterns according to the characteristics of the buildings in alignments.First,the building alignment pattern was divided into single-connected and multi-connected.Then,the single-connected alignment was divided into linear and curvilinear alignment.Finally the linear and curvilinear alignment were divided into standard linear alignment,central linear alignment,boarder linear alignment,oblique linear alignment and standard curvilinearalignment,central curvilinear alignment,boarder curvilinear alignment respectively.2)Description model of building alignments.This paper proposed a description model for the patterns of building alignments consisting proximity model,similarity model and structure model based on the proximity,similarity and common orientation of the Gestalt principle.The proximity model was built up considering the proximity relationships determined by the Delaunay triangulation of the buildings.The similarity model was built up by the similarity of individual spatial attributes of buildings,which includes size,shape,orientation and density.The structure model was built up by the global and local characteristics of building alignments including orientation consistency of buildings and alignment path,deviation of buildings and consistency of building borders.3)Extracting method of building alignments.The extraction method for building alignments.First the MST of buildings was generated based on the proximity model of building alignment using the Prim algorithm.A threshold value was determined by the muli-connected buildings' edge value of the MST.The buildings were clustered roughly according to this threshold.Then a combination method was carried out to combine the rough clusters.Finally alignments were extracted from the buildings.4)Pattern recognition method of building alignments.The pattern recognition method for building alignments.The alignments extracted before were classified by voting,and parameters of the building in these alignments were calculated.Then the sample set were constructed.A graph convolutional neural network model was established by combining the B-P neural network and graph convolution operation.The parameters in the model were trained by the supervised learning process with the sample set.Experiments were conducted with the data provided by Open Street Map,and the results showed that the pattern recognition model of building alignments based on the graph convolutional neural network can predict the patterns of building alignments proposed by this paper.