Research On Automatic Vectorization Of Buildings From Remote Sensing Images

With the development of remote sensing observation technology and remote sensing satellite system,high-resolution remote sensing image contains more and more abundant ground object information.It can accurately describe various geospatial objects,such as vehicles,ships,aircraft,roads,etc.remote sensing image plays an important role in economic and social development and national defense construction.Building information is an important kind of ground object information contained in remote sensing images.With the development of economy and society,buildings occupy a very important position in modern cities.Accurate building information plays an important role and significance in urban planning,engineering drawing,urban construction and monitoring.Facing the problem of extraction and analysis of remote sensing image data,this dissertation studies the model and method of automatic vectorization of buildings in remote sensing image.It mainly extracts buildings from remote sensing image,and then regularizes the extraction results and extracts the boundary to carry out automatic vectorization of buildings.The main research work of this dissertation includes:1.Analysis and modeling of automatic vectorization of buildings from remote sensing images.Facing the problem of automatic vectorization of remote sensing image buildings,this dissertation combs the existing remote sensing image data and characteristics,and analyzes the characteristics of grid data and vector data in GIS and the conversion methods between them.The data and problems are defined and described,the framework model of building automatic vectorization for remote sensing image is proposed,and the relevant evaluation indexes are given.2.Object oriented vectorization method for building semantic segmentation of remote sensing images.According to the characteristics of building Vectorization in remote sensing image,a building object meaning segmentation model based on deep neural network is established.The coding decoding structure is adopted,and the hole convolution is added to the feature extraction part,which can increase the receptive field and reduce the image size.At the same time,the attention module is added to effectively extract the context information of the image.The boundary loss is added to the loss function,which can pay more attention to the quality of the boundary when extracting buildings,so as to prepare for the subsequent vectorization.On this basis,the building extraction method of remote sensing image is improved based on GCdeeplabv3 +.The results show that the extracted building boundary is closer to a continuous straight line,which improves the smooth corners of the extraction results and the adhesion between dense buildings.3.Vectorization of building automatic extraction results based on generating countermeasure network.For the automatic extraction results of buildings in grid form(building mask),an accurate and efficient vectorization transformation model and method are studied.Aiming at the problems that the extracted building results do not fit the geometric shape,obvious sawtooth edges and excessively smooth corners,a regular model of building extraction results based on generating confrontation network is proposed.The generated building mask is closer to the right-angle polygon and the edge is smoother.A building mask contour vectorization algorithm based on the main direction is designed to effectively generate polygonal building contours that fit the actual geometric structure.