A Simplification Model For Building Shapes Based On Machine Learning

The simplification of the shape of spatial objects is an important research topic in the map generalization,regarding with the core problems of Geographic Information Science such as spatial cognition and spatial similarity measurement.As a geographical element,buildings that is widely distributed in maps,the simplification of their shape has always been a crucial part in the large-scale map generalization.Simplification of shapes has two major steps: description and simplification.Among them,the geometric shapes of buildings are complex and variable and thus are difficult to describe.Besides,different description methods correspond to different simplification strategies.Therefore,the simplification of the shape of a building is also the most difficult task to do.When measuring the geometry of a building element,traditional methods use a number of shape descriptors to measure a certain feature of the shape or,construct a strict mathematical function and statistical method to approximate the shape contour.However,these methods do not consider humans' factors very well.The visual factors in the process of spatial cognition,and the identification methods of them also need manual settings and interventions.In addition,for these traditional shape description methods,their simplification strategies also do not take into account visual factors and have certain limitations.In view of this,based on the related algorithm of machine learning,this paper constructs a geometric description model that can recognize the visual factors of the building shapes,and proposes the shape simplification method corresponding to this model.The research achievements and innovations made in the dissertation include the following three aspects:(1)Based on the characteristics of self-learning of image features based on deep Convolutional Neural Networks,combined with the unsupervised learning ability of Auto Encoder,a self-supervised learning description model for building geometry is constructed.The basic idea is that,in the structure of the automatic encoder,a deep convolutional neural network is embedded so that the encoder and the decoder of the Auto Encoder can use the convolution operations to detect the shape features of the buildings.(2)Based on the model constructed in(1),a method for describing the shape of a building based on spatial and visual recognition of human beings is proposed.Specifically,an encoder and a decoder of the shape description model are used to perform the feature detection and reconstruction on a shape data set,respectively,thereby obtaining a shape feature set calculated by the encoder.This feature set is an abstract feature description of the input simplification shape,which provides a good reference for the simplification model.(3)According to the spatial visual cognition law,a building shape simplification model based on cognitive template matching is proposed.The basic idea is to analyze the feature elements of visual cognition and construct the building simplification template shapes according to the cognitive classification method;then,use the similarity measurement between the shape feature sets to measure the matching template object;finally,use the matching template to treat simplification.The shape is then set and simplified.In this paper,we use a big data of real building shapes to train the self-supervised shape description model.Then the two models are tested and analyzed.Experiments show that the shape description model constructed in this paper can detect the shape feature set of any buildings very well.The results are in line with human visual cognition.It overcomes the shortcomings of human intervention for feature definition to a certain extent,and it has high shape distinguishability.The template-based building shape simplification model constructs several template objects that conforms to spatial visual cognition.The similar template object measured according to the shape feature set has a good visual match,and most of the shapes in the experimental area are well-simplified.The effects are relatively ideal.The experimental results of the two models are in line with the spatial visual cognition habits.This combination solves the shape simplification problem of buildings in large-scale maps in an intuitive way,providing a certain reference for the application of the machine learning in map generalization.