Research On Ground Objects Classification Based On Airborne LiDAR Data And Hyperspectral Data

UAV Airborne LIDAR point cloud data or hyperspectral image data has become an important data source for high-precision classification of ground objects.However,the ground objects classification that only uses a single data source still has a certain degree of deficiencies,and the object classification that fusion of multi-source data is emerging.This paper focuses on the key technologies of data fusion,feature selection and extraction,classification algorithm and so on for LiDAR point cloud data and hyperspectral image of UAV.First,conduct pretreatment research on two kinds of data,use the triangulation filtering method to filter the point cloud data and use the irregular triangulation interpolation method to generate the pit-free DSM,and perform the preprocessing research on the hyperspectral image such as atmospheric correction.Then the geometric registration of DSM and hyperspectral images is realized based on the image registration method;then extracting the point cloud features and the hyperspectral image features;and construct the fused single point feature vector.Finally,the forest classifier is used to complete the classification of ground objects in the research area,and the accuracy of the classification results is verified.The main research contents and results of the thesis are as follows:(1)Introduces the pretreatment process of UAV LIDAR point cloud and hyperspectral image.Denoising and filtering the airborne LiDAR point cloud data to generate DSM grayscale images;and draw on the idea of image registration to realize the registration between the LiDAR point cloud and the hyperspectral image.The experimental results show that the registration error was less than 0.8m,which meets the experimental requirements.(2)Aiming at the characteristics of the two kinds of data,the elevation and structure features of point cloud data and the spectral and texture features of hyperspectral images are extracted respectively.Based on the point cloud data,the hyperspectral features are assigned to the corresponding single point on the point cloud data.A set of single-point feature vectors was constructed which combines point cloud features with Hyperspectral features.(3)For the fused set of single point feature vectors,use the random forest classification method to classify the ground objects in the study area(Huilai Experimental Station of the Chinese Academy of Sciences).It has realized the high-precision classification of 11 types of ground objects in the research area,including soybeans,corn,crabapple seedlings,buildings,luan trees,poplars,arborvitae,Chinese pine,low grass,towers and ground points.The overall classification accuracy is84.39%,which is 4.78% higher than the classification accuracy using only point cloud data features.