Building Pattern Analysis Supported By Deep Convolutional Learning

As an important part of the geospatial vector database,building is one of the foundations for representing spatial phenomena,supporting spatial analysis and providing spatial services.Whether perceived individually or as a group,building has significant pattern characteristics.However,as a two-dimensional planar vector data,the formal and quantitative representation of the building is still diff-icult to define and unify;what's more,as a high-level concept,the pattern has strong ambiguity and uncertainty.Although many studies have been carried out from multiple perspectives,multi-level and multi-field,and some achievements have been made,it is desiderating to explore more effective technologies to extract the features hidden behind the data.In recent years,as a deep supervised learning architecture,convolutional neural networks(CNNs)have shown excellent performance in various fields.The excellent performance of CNNs is attributed to the local correlation preservation,which is consistent with the local correlation and multi-scale features of geospatial data such as buildings.However,deep learning does not resonate in the analysis task of building patterns.The main reason for restricting its development is that the buildings stored in vector form cannot meet the regularizing requirements of deep learning for input data.Aiming at this problem,the study introduces a novel graph convolution operation based on the graph Fourier transform and convolution theorem,and constructs a graph convolutional neural network(GCNN)by combining it with the computational mode of the neural network.The main work is as follows:(1)The scientific problems of building pattern research were discussed from two dimensions of space size and space type,and the pattern characteristics at different scales and the basic classification from different perspectives were summarized.Based on the Gestalt psychology theory,the cognitive variables of individual buildings such as position,size,orientation,shape were summarized,as well as the cognitive variables of building groups such as density and distance.The geometric graphs for modeling the relationship between buildings based on the Delaunay triangulation(DT),Voronoi diagram and minimum spanning tree(MST)were introduced.(2)The graph Fourier transform with the orthogonal Laplacian eigenvectors as the decomposition bases was introduced,and a novel graph convolution was defined by converting it from the vertex domain into a point-wise product in the Fourier domain on the basis of this transform.Furthermore,a fast and localized graph convolution based on polynomial approximation was introduced,which offers many advantages,such as fast computation,a small number of parameters,and spatially local connectivity.Based on these properties,a graph convolutional neural network(GCNN)was proposed by the operation with the computational mode of the neural network,which can be used to solve the classical building pattern analysis problems.(3)Aiming at the typical template characteristics of individual building distributions and representation,a template matching and simplification method for buildings was proposed.The main idea is to extract some stable,typical and repetitive shapes as templates to replace the original buildings.Four different types of templates were summarized from the characteristics and creation ways;two approaches for measuring of shape similarity were proposed,namely turning function and graph convolutional autoencoder(GCA);a template matching approach based on the least square was proposed.A series of experiments were conducted to compare and discuss the performance of the turning function and the GCA model in shape similarity measurement.Large building data with different scales was used to verify the effectiveness of the template matching simplification,and the results were evaluated qualitatively and quantitatively;The advantages and disadvantages of different template matching approaches were compared,and the influence of template diversity and hierarchy on the simplification results was further discussed.(4)For the clustering of buildings,a supervised GCNN clustering model was proposed.By taking the features of nodes and the states of edges as learning objects respectively,this study proposed two different clustering methods:Node-clustering and Edge-clustering.· Node-clustering uses the center point of cluster as the annotation information,and converts the clustering problem into the feature mapping problem,and finally uses the K-means algorithm to obtain the clustering result;· Edge-clustering uses the distribution differences and spatial autocorrelation of node features in the left and right neighborhoods as well as their own geometric properties as inputs,and finally predicts the state of retention or deletion for each edge.The study used the supervised method to evaluate the clustering results,and the adjusted Rand index(ARl)was used to quantize the difference between the labeled and predicted clustering results.Experiments showed that both node-and edge-clustering could extract the features for descripting the clustering patterns between buildings,and the Edge-clustering was superior to Node-clustering in fitting ability and generalization ability.Meanwhile,this study compared and discussed the proposed GCNN clustering with two existing methods,and concluded that the supervisory information in the GCNN model can effectively integrate a large number of cartographic experiences and knowledge,which enables it more suitable for the data of different scales and regions.(5)For the classification of building groups,a GCNN-based method was proposed.The method is similar in thought to image classification,that is,constructing the relationship between buildings with a geometric graph,and treating individual building as a processing unit,combining and transforming the descriptive variables thereof by a convolution operation,finally outputting the class of the overall group through the full connected layer.From the perspective of visual cognition,the paper classified the building groups at the block level into two types,namely regular pattern and irregular pattern;and used different regions of data to test the effectiveness,robustness,and generalization ability of the proposed GCNN method.Experiments showed that this method was superior to the traditional machine learning methods such as support vector machine and random forest in classification accuracy and generalization ability.Further,it is an end-to-end solution,that is to say,it does not need to extract features for describing the building groups manually.Meanwhile,it overcomes the shortcomings of rule-based classification methods such as too strict rules,manual parameter setting,and limited model adaptation.By considering the building pattern analysis as a case,this study highlighted the potential of the GCNN to analyze graph-structured spatial data.In the future,additional efforts should be devoted toward improving the GCNN with deep learning technologies and applying the GCNN to complex geospatial cognitive analysis,geographical intelligent computing and knowledge modeling.