Research On Coastline Extraction And Feature Classification Based On LiDAR Point And Image

Coastal Zone is a composite geographical unit connected by the marine and terrestrial, and the most favorable environmental and resource area for human activity. Because of the geographical environment and the practical factors, many island coastal zones should be investigated in our country, including shoreline, intertidal, vegetation, land use and wetlands, ect. In which, the most important element is the land utilization for the coastline and coastal zone.Currently, there are many algorithms based on Remote Sensing Images data for coastline extraction and feature classification. Some of them have yielded results. But using a single remote sensing data is difficult to obtain high accuracy and high automation degree coastline terrain classification, so we need to introduce other auxiliary data source. LiDAR data and images have both advantages and disadvantages, they are highly complementary.Therefore, then new technology aims at combining airborne LiDAR accurate3D point cloud data with digital images for shoreline extraction and feature classification, is becoming a new research hotspot of island coastal zone investigation.This study includes two main topics-shoreline extraction and coastal feature extraction, by discussing it, a number of related key technologies is discussed in depth.(1) Summarizes the advantages and disadvantages of existing point cloud index algorithm. Studies a two-level index algorithm massive LiDAR point clouds based on Hilbert permutation code and R-tree. Optimizes the first-level index by clustering and M value of R-tree’s degree; Using Hilbert R tree as a second index, to effectively control the height of a two-level R tree. Meanwhile the point cloud can be increased and updated partly, and achieves the efficient management of mass point cloud. In order to verify the efficiency of the algorithm, we use the larger point cloud data to establish complete KD tree, quadtree memory index and the index described in this article. KNN query, window query, and based on the gradient changes of filtering processing were carried out in these three indexes, results showed that the two-level index described herein is optimal on query efficiency and overall performance.(2) Summarizes the characteristics of LiDAR data, which includes:elevation difference, gradient, normal, strength, echo times, flatness. Also explored a multi-feature filtering algorithm, which is improved by filtering algorithm, based on the gradient of the LiDAR data. According to the results of the gradient filter, and then to flatness filter as supplemental judgment condition, in this way, it will improve the precision of the filter. Experimental results show that the method of filtering is effective, robust and the Classification accuracy is very high (>80%), it meets the classification accuracy of routine application. Simultaneously, the test results of the use of the classification thresholds are default threshold is obtained by automatic classification, by adjusting the threshold value, the higher classification accuracy can be obtained theoretically.(3) Through the repeated studies of typical experimental zone, we propose a water edge extraction method of a joint LiDAR point clouds and images, starting from the following key technical approach:1) water extraction method which combines image segmentation and Normalized Difference Water Index;2) initial water edge extraction and a variety of point cloud features refined water edge extraction method;3) water edge smoothing method.And the experimental results testify the effectiveness of this method.(4) Summarizing the existing method of terrain feature classification, aiming at the present problems, this dissertation is based on random forest, puts forward a object-oriented feature selection and classification method bases on airborne LiDAR data, and evaluates the correlation of coastal zone target object’s geometry, spectrum and texture features, etc, selects the appropriate features used for urban terrain classification. Via the elimination of reverse iteration can quantitatively select the features which is the most relevant with the target, rely on the SVM and RF classifier experiments can prove that the RF classification precision is equal to the accuracy of the SVM classification after the feature selection, and that, the SVM classifier is in the use of the characteristics which have been selected compared to use all the features, by feature selection eliminates part of the feature classification accuracy slightly better than the uses of all the characteristics of the classification accuracy. Finally, in order to improve the phenomenon which is severely mixed and separated between buildings and other land feature when the environment is located complex building region, we are based on the fuzzy sets theory, studies an method of category membership classification. This method is based on image fusion and LiDAR point cloud. This experimental results show that with the aid of LiDAR point cloud, it can effectively improve the mixed and separated phenomenons, which spectral information classification results appearing in buildings, bare land, roads and water, etc. ground between classes. And it can be able to provide more accurate classification results in complex building region.