Simplification Methods For Large-scale Buildings Considering Local Features

Pattern recognition and simplification of large-scale buildings is an important topic in the field of cartographic generalization.Many scholars have made great efforts for the simplification of large-scale buildings.However,scientific research has failed to adopt a balanced strategy between shape simplification and the preservation of local features.For this reason,this paper adopts a strategy of classifying buildings according to shapes and then simplifying them according to the classified shape.The main content is to identify the pattern of the English alphabet shape of the building by Graph Convolutional Neural Network,and then adjust the local features of the template according to the classified pattern,and finally complete the simplification of the building.For the dataset part of the Graph Convolutional Neural Network,Use the skeleton line of the building as a dataset instead of the contour line,and good classification results are achieved.Finally,a simplification strategy that takes into account local features based on the shape classification.The research content of this paper and the results achieved include the following three main points.:(1)Construction of the building skeleton datasetThe geometry of a building is the most important part of the geographic elements,and the skeleton line can accurately describe the basic shape and general trend of the geometric elements of a building.In this paper,the building skeleton line is constructed by first encrypting the building vector data with nodes and then performing a constrained Delaunay triangulation,which is divided into class 1,2 and 3 triangles in turn.Based on a two-dimensional matrix structure,a two-dimensional skeleton line matrix is constructed using the number of triangular edges of the three types as columns.For the intersection of small skeleton line segments,a node selection algorithm is used to select their positions and connect them to form the skeleton line that best matches the pattern characteristics,thereby achieving a fully automatic extraction of the skeleton line,and the node selection algorithm is also the innovation of this paper in extracting the skeleton line dataset part.(2)Graph Convolutional Neural Network shape classification of building based on skeleton linesThe representation of buildings as an extremely important component of geographical space and their shape recognition is a fundamental theory in the field of cartographic generalization.This paper combines the skeleton line vector data of buildings and Graph Convolutional Neural Networks to propose a building shape classification method,which has clear and well-defined steps,first extracting the skeleton lines of buildings to form a dual graph structure with edges as nodes,and then extracting their basic geometric features to obtain the graph representation information of building shapes;The high-dimensional shape information is extracted using multiple rounds of extraction and aggregation of building skeletal lines and their shape information using a Graph Convolutional Neural Network.Classification of high-dimensional shape information is accomplished through a fully connected layer.The experimental results shown that the method effectively extracts the building shape information and obtains accurate classification results,eliminates the lack of in-depth description of building shape by contour lines,fully reflects the advantages of Graph Convolutional Neural Networks,and achieves a method of extracting and classifying building shape information with high accuracy.The model was tested in the experiment using a sample of 240 Lanzhou City.The overall classification accuracy was 93.3%,with the F-measure values of 96.5%,92.3% and 100% for E-,F-and H-types respectively.The experimental results show that the shape classification model of the buildings in has a strong classification capability,also shown that buildings with complex alphabetic shapes are better classified than those with simple alphabetic shapes.(3)Local feature changes in buildings based on matched templatesOnce the second part of the large-scale building shape recognition was completed,Extraction of the feature matrix for the original building,and the templates were shaped according to the feature matrix;The original building is replaced by the shaped template to complete the simplification.This approach takes good consideration of the local features of the building.Ultimately,the evaluation indexes were measured and tested in comparison with other methods,all of which showed great results.This paper designs a training set and using the advantages of graph convolutional neural networks for the recognition of data in the spatial cognition field,construction of a model for building shape recognition,selects the shape of the English alphabet to which the building belongs based on this model,extraction of the feature matrix for the fully classified buildings.The template is transformed according to the feature matrix for local features.The building is eventually replaced by the transformed template,thereby completing the simplification of the building.The algorithms in this paper explores the idea of building simplification based on combining deep learning and traditional algorithms.