Classification of LiDAR data with point based classification methods

Nowadays, airborne LiDAR data is used frequently in various applications such as object extraction, 3D modelling, change detection and revision of maps with increasing point density and accuracy. In this application, LiDAR point cloud should be classified as a first processing step. To achieve high accuracy, accurate classification of LiDAR point cloud is needed. In this study, possibilities of raw LiDAR point classification with automatic point based classification are investigated. The available methods for classifying LiDAR point cloud are examined and researches are summarized in national and international arena for this purpose. The problems in the accurate automatic classification of LiDAR data with intensity and multiple returns are analysed. Furthermore, the problems encountered in the classification of LiDAR data, which was converted from point cloud to the regular grid, were investigated. In this study, the point based classification of LiDAR point cloud is proposed to eliminate problems in classification of regular grid. The automatic point based classification approach composed of hierarchical rules has been established to solve problems (class mix, etc.) may arise depending on the characteristics of LiDAR data. The analyses of the parameter in the hierarchical rules have been performed and two point based classification approach were proposed which have different hierarchy. The hierarchical classification rules have been established for Standart Approach, which was used frequently in point based classification, in order to test the performance of the proposed approach. The point clouds for three different sites in Zekeriyaköy, Istanbul have been classified with proposed approach and three main classes (ground, buildings and vegetation) were acquired. The accuracy assessments for point based classification have been performed for Test Site 1 (light urban area) and Test Site 2 (urban area). The overall classification accuracy was computed as %54 for Standard Approach and %80 for Approach 1 and Approach 2 in Test Site 1. For Test Side 2, the computed overall classification accuracy was %59 for Standard Approach and %83 for Approach 1 and Approach 2. In Standard Approach, ground points were mostly classified as vegetation with in light urban area, urban area and forested area. The obtained classification results were quite close to each other in light urban area, urban area and forested area with Approach 1 and Approach 2. The improvements have been obtained substantially for the problems encountered with Standard Approach such as ground points classified as vegetation, points in roofs classified as vegetation and some points in building corners classified as vegetation. The 3D building models were generated for buildings in Test Site 2 to test usage of the classified point clouds for 3D building model generation with Approach 1 and Approach 2. The obtained results showed that the 3D building models were generated successfully. The results of the accuracy assessment of the point based classification and generated 3D building models proves that the research aims are successfully achieved.