| Buildings are one of the most important artificial features in the basic geodatabase,and the automatic extraction of buildings is of great value for image interpretation,urban planning,3D modeling,military reconnaissance,etc.Therefore,getting building information in remote sensing images has turned out to be among the significant study subjects in remote sensing domain.For the past few years,the acquisition capability of high-resolution remote sensing images(the following is referred to as "high-resolution images")has been increasing,providing massive data support for the extraction of buildings from remote sensing images,but its rich feature detail information has also caused an increase in the variance within features and a decrease in the variance between feature classes,resulting in the limitation of pixel-based extraction method,which is not suitable for high-resolution image.While the advantages of object-oriented extraction methods are more outstanding in high-resolution image information extraction domain.Furthermore,it is difficult to obtain buildings with 3D information from high-resolution images,while the rich 3D information of Digital Surface Model(DSM)data can make up for the lack of high-resolution image information,so the paper uses high-resolution images and DSM data with height information to be the foundation,integrates the advantages of both data,and takes the object-oriented analysis method as the core idea to develop the method of building extraction research.The main research contents and results are as follows.(1)Aiming at the problem of significant computational difference in the current methods for determining the optimal scale,on the basis of analyzing and summarizing previous studies,it is considered that in image segmentation,it is necessary to achieve as much homogeneity within the object as possible and as much heterogeneity between adjacent objects as possible.Therefore,the paper presents a way to identify the optimal scale on the basis of object homogeneity and heterogeneity indexes and combined with the mean change point method,and constructs the variation curve between each evaluation index and the segmentation scale.The viability of this way to identify the optimal scale in multi-scale segmentation is verified by experiments,and the method is simple and efficient without manual adjustment of parameters.A satisfactory result can be obtained.(2)Aiming at the problem of "dimensionality disaster" of features in object-oriented analysis method,the paper proposes to analyze the features affecting the extraction of buildings from high-resolution images by using the geographic detector model from the selective principle of feature selection approaches.Firstly,the factor determination force of each feature is calculated by the factor detector,and then the feature with strong determination force under the factor detector is detected by the detector in the risk area,so as to reveal the difference of various features in the distribution of distinct ground object types,and the experimental verification is conducted on the data of different types of regions.The results show that this method has certain advantages for feature selection in object-oriented analysis methods.The experimental results also show that the height information provided by DSM data is indeed beneficial to the extraction of high-resolution image buildings,which further confirms the feasibility of this paper combining high-resolution image and DSM data to achieve the extraction of high-resolution image buildings based on the core idea of object-oriented analysis method.(3)Aiming at the problem of irregular building edges after initial extraction of buildings,based on the abrupt change in the elevation of building edges,this paper uses a neighborhood elevation comparison algorithm to regularize the edges of buildings,and designs experiments to verify that this method has good applicability for building data in rural and urban areas,and can effectively optimize the contour of buildings,The overall accuracy and effect of building information extraction are relatively ideal. |
