Researches On Classification Techniques Of Airborne LiDAR Data With Auxiliary Aerial Imagery

With sustainable social and economic development, the demands for speed and accuracy of spatial data acquisition are increasing. Photogrammetry, as the primary means of acquiring spatial data, is becoming an active research area. The airborne LiDAR technology has been widely used in object extraction and surface three-dimensional (3D) reconstruction as a new type of progressively wide measurement system, because of its fast access to massive spatial data. The chief task of airborne LiDAR data interpretation and modeling is land-use/land-cover classification. The classified ground points can be used for topographic mapping, engineering survey, environmental planning, etc. Buildings, various types of vegetation and other non-ground points are available for 3D digital city construction, city planning, or the reconstruction of 3D building models and urban vegetation analysis in GIS database, and so on. Therefore, the rapid and accurate interpretation of massive LiDAR point cloud data has attracted much of the concern of many researchers around the world.There are some problems in land-use or land-cover classification using only LiDAR point cloud data, due to its discrete, irregular and uneven characteristics of point distribution and the lack of spectral feature information. However, with the development of hardware technology, many of airborne laser scanning systems include high-resolution digital cameras that can obtain high-resolution aerial images at the same time as the laser scan data. The high-resolution aerial images can provide a wealth of spectral, texture information and compensate for the weakness of LiDAR data for land-use classification effectively. It is beneficial to fast and more efficiently post-process the airborne LiDAR point cloud data by fusing LiDAR data with aerial images, because they naturally have the complementary properties of each other. [A Habib 2004; Habib 2005; Habib 2006].In order to improve classification accuracy by using airborne LiDAR point cloud data with aerial images, in this paper, we have studied and researched several critical theories and technologies related to land-use classification, which is listed as follows:1) Several ways of querying three-dimensional discrete LiDAR point cloud data have been studies, including two-dimensional sub-block index, quad-tree index and KD-Tree that can query quickly the nearest neighbor k points and area in 3-dimensional space. The traditional quad-tree indexing method is improved to process point cloud data, by proposing a dynamic calculation method of quad-tree optimal depth. The key technology of revised quad-tree method to speed up the indexing and query process is proposed using stack to simulate the recursion, and the smallest outsourcing rectangle, which solved the organization and visualization issues of the point cloud data, to some extent.2) According to the terrain characteristics of mountainous areas covered by dense vegetation, we propose a novel filtering algorithm from LiDAR data by a section-cross method. First, the whole LiDAR point cloud is divided into a series of voxels, each of which continue to be halved. Then multi-returns and intensity information is used to obtain the original ground points in each halved voxel. Finally, the left three-dimensional points in each halved voxel are projected into two-dimensional planes to obtain ground points by using point-to-point geometric relations. Experiment results show the proposed section-cross method is feasible to extract digital elevation model (DEM) form LiDAR point cloud in the densely wooded mountainous areas.3) We present a new registration method of the high-resolution aerial image and airborne LiDAR point cloud data in this paper. First, the misalignment errors between CCD camera and the IMU are analyzed. Meanwhile, the corresponding points in overlapping images are collected as control and connection points; the LiDAR elevation data is introduced for high-precision three-dimensional spatial points based on multi-baseline forward intersection with LiDAR elevation constraint; these spatial points are reversely calculated to corresponding 2D image points by adding the misalignment error matrix the traditional collinear equations. Finally, the exterior orientation elements can be refined by adjustment calculation with the distance constraint from the corresponding 2D image points to manually selected image points for the registration of LiDAR data and aerial images. Experimental results confirm the validity of the proposed method.4) We first analyze the spectral information and geometric information from the high-resolution aerial imagery and LiDAR point cloud data, respectively. Then we propose a semi-supervised multi-level multi-feature classification method fusing the LiDAR point cloud with the aerial image data. The classification task is divided into three levels:in the low-level, LiDAR point cloud is classified as ground and non-ground points; in the mid-level, using high-precision training data, vegetation is extracted from the ground points, and vegetation and buildings are separated in non-ground points. The shape index is used to refine the classification results. Finally, in high-level, according to the image classification results, rough classification results, elevation information, spatial dispersion, first-and last-returns and intensity information, the LiDAR data is classified into grass, buildings, bare ground, vegetation and miscellaneous class composed by points that don’t belong to the first classes. The accuracy of classification is improved by fusing the LiDAR data with aerial imagery.5) Concerning the emerging issues caused by various complicated processing methods and massive LiDAR data in the land-use classification task using the integration LiDAR data and imagery, we design a parallel data processing algorithm based on a heterogeneous cluster computer system to solve them, such as the message communications, task decomposition, distribution and load balancing. Thus, this proposed parallel processing system can effectively combine a variety of algorithms and hardware resources to improve the processing efficiency.In combination with photogrammetry, digital image processing, data structures and parallel computing method, we focus on utilizing the high-resolution aerial image and LiDAR point cloud for land-use classification researches, including point cloud data organization, DEM generation, registration and classification of LiDAR data and imagery, parallel processing, and give solutions to specific problems related to classification. The terrain classification by the integration of LiDAR and aerial image data is a complex task, so that how to improve the terrain classification accuracy, level of automation and accuracy of the model is still needed a lot and in-depth research.