Research On Classification Method Of Land Cover By Fusing Aribone LiDAR Point Clouds And Aerial Imagery

In view of the current advantages and disadvantages of object classification based on a single data source, this paper mainly research on the key techniques such as airborne LiDAR point cloud data filtering and remote sensing image registration feature classification.First,research on the relative contents of point cloud data preprocessing and propose two filtering methods in view of the poor robustness of current filter algorithms,which include a filter method of LiDAR point cloud based on the change curves of skewness and surtosis or differential morphological mrofiles.Then, after pretreatment of LiDAR point cloud data interpolation DSM grayscale, corner points of construction detected by SUSAN algorithm, and complete the DSM grayscale and remote sensing image registration based on the two levels of affine transformation model. Finally, aobject-oriented hierarchical classification method is used to complete features accurate classification.The major work and conclusions of this dissertation as follow:(1) Introduce airborne LiDAR measurement principle and analyze the obtained data.Through compared with the advantage and disadvantage of existing data organization mode, this paper adopted a virtual grid and rule grid and in view of the different filtering methods,choose the appropriate grid size to grid-handling of the original point cloud to improve the efficiency of data processing. Improved single threshold method was used to get rid of coarse points in the acquisition process of airborne LiDAR.(2) Conduct the thorough research to the classic filtering algorithm and propose the filter method of LiDAR point cloud based on the change curves of skewness and surtosis in view of the influence of the threshold value for filtering precision.This method introduced the skewness and kurtosis balance thought, using the statistical moment equilibrium principle to filter feature points.To solve the problem which the terrain adaptability of morphological filtering algorithm is not strong, the improved method is proposed based on differential morphological mrofiles. It introduced the gaussian convolution operation in image processing and used threshold discriminant function to identify the ground point. Through ISPRS provided official datas to test of two kinds of algorithms,experimental results show that two methods can obtain good filtering effect and has higher precision and robustness.(3) By using image registration thought, translate the airborne LiDAR points clouds with remote sensing image 3D-2D registration into DSM grayscale images and remote sensing image 2D-2D registration. To this end, the natural neighborhood method is firstly used to analyse the LiDAR points clouds interpolation for DSM grayscale.Then,adaptive SUSAN corner detection algorithm is utilized to extract building corner,and use multistage affine transformation model to complete the registration.Experiments show that this method is simple and high operation efficiency and the registration precision meet the demand of the experiment, the effect is better.(4) For the problem of LiDAR points clouds and remote sensing image feature classification, this paper puts forward the object-oriented hierarchical classification method.Using multi-scale segmentation technology and according to the rule set of classification characteristics,adopts the fuzzy classification algorithm of decision support system to complete feature segmentation.Through the example,the accuracy of this method is improved obviously compared with based on pixel classification,which has better classification effect.