Research On Some Key Technologies Of Automatic Building Change Detection Aided By Three-dimensional Information

Building change detection, as one of the most important components for updating a geographic information database and monitoring the geographic national conditions, is significant to urban planning, disaster assessment, urban growth monitoring, illegal building identification and geographic information database updating. Taking the illegal building identification for example, traditionally it usually requires a field survey or mass report to identify changes in buildings. These methods are time consuming and exhausting, and with large probability of omission error. In recent years, some cities, for example Beijing, attempt to identify changes with satellite imagery aided by the manual survey. However, owning to the fact that automatic image analysis technology is still not mature enough, many manual identification and validation are involved during the process. Every year, the country's land law enforcement and urban management department spent several billion of human and material resources for this task. A high degree of automation and reliable method to detect building changes is an urgent need for the market.Currently, the research of building change detection is mainly based on spectral information alone, and its precision and accuracy is relatively limited. Considering that three-dimensional information used for building change detection is of less ambiguity when compared to the spectral information alone, and it can greatly improve the precision and accuracy of the detection results. Meanwhile, thanks to the mature of the acquisition of the three-dimensional information, in particular, the development of laser scanning hardware and the breakthrough of image dense matching technology, more and more researchers in the field of Photogrammetry and Remote Sensing tend to perform automatic change detection in buildings with the use of three-dimensional information, that is, DSM-assist building change detection.In this paper, two kinds of high-resolution remote sensing data were taken as input, and the high-resolution remote sensing data were multi-temporal airborne laser scanning (LiDAR) data or multi-temporal stereo airborne images.The aim of this research is to achieve automatic building change detection and change-type determination, wherein the involved acquisition of the digital surface model, registration of multi-temporal data, filtering of dense point cloud data and automatic building change detection have made a deep study, so as to achieve the automated production of change information from multi-temporal high-resolution remote sensing data. The main contents and results obtained in this study are as follows:(1) Generation of digital surface model (DSM). Firstly, a brief introduction of airborne laser scanning system is performed, and strengths and weaknesses of the airborne LiDAR data are analyzed below, as well as its common application. Then, semi-global matching (SGM) algorithm for stereo images is detailed in this paper. On this basis, considering that point cloud data obtained by SGM are single model-based, and certain inconsistencies may exist in the overlap of different models, thus an energy optimization-based merging of multi-view stereo is proposed to obtain a more consistent large range of DSM.(2) Registration of multi-temporal data. In this paper, the basic principles of iterative closest points (ICP) algorithm is firstly introduced, and its detailed steps used for the registration of multi-temporal point cloud data in this paper is described, followed by using two different sets of multi-temporal point cloud data to demonstrate the effectiveness of this method. Secondly, the fusion of iterative closest points (ICP) algorithm and bundle adjustment for the registration of point cloud data and images is introduced to the registration of multi-temporal stereo images. Finally, two different sets of multi-temporal stereo images covering one square kilometer were used to confirm the effectiveness of the proposed method.(3) Filtering of the dense point cloud data. In order to improve the accuracy of building change detection, terrain changes must be first excluded from the object changes. In this paper, the multi-temporal digital elevation model (DEM), which is obtained by filtering of the dense point cloud data (DSMs), is used to distinguish terrain changes from object changes. In this paper, the filtering of the dense point cloud can be divided into three parts, including removal of outliers, adaptive triangulated irregular network (TIN) filtering algorithm, and scan line optimization-based filtering with dynamic programming algorithm.(4) Automatic building change detection from multi-temporal airborne LiDAR data. An automatic building change detection method that applies object-based analysis is first proposed to identify areas that have changed and to obtain from-to information in this paper. In this method, candidates of changed building objects is first extracted by using a smoothness computation, then, multi-features calculation and analysis is performed to distinguish the true changed buildings from trees, so as to achieve the building change detection and change-type determination. Furthermore, to decrease the workload of subsequent manual checking of the result, an feature-based self-diagnostic scheme is designed to guide the user to the most likely wrong detections. With this self-diagnostic scheme, the user only needs to check a part of the objects. Thus, the manual workload is significantly reduced.(5) Automatic building change detection from multi-temporal stereo aerial images. A novel highly automatic method of building change detection from multi-temporal stereo aerial images is proposed to identify changed buildings and to determine its change type. First, based on the registered multi-temporal stereo images by the above section (2), semi-global matching is applied to the images to obtain the dense old point cloud data, which is also known as DSMs. And multi-view stereo merging and morphology leveling, which is important in reducing false matching points and excluding small false objects, are applied to the generated point cloud data to form a more reliable changed objects. Then, considering the complexity of distinguishing buildings from trees in the airborne images, an image structural feature based on the histogram of the direction of lines (HODOL), which is insensitive to radiation deformation, is designed to exclude trees from buildings. Finally, two typical datasets, each covering 1 km2 area, are used to validate the proposed method.This paper mainly focused on the building change detection with the aid of three-dimensional information. It elaborates this issue from the following aspects: generation of digital surface model (DSM), registration of multi-temporal data, filtering of the dense point cloud data, building change detection from multi-temporal airborne LiDAR data and from multi-temporal airborne stereo images. And experiments demonstrate the effectiveness of the proposed method.