LIDAR DATA CLASSIFICATION WITH REMOTE SENSING TOOLS H. Arefi 1 , M. Hahn 1 , J. Lindenberger 2 1 Dept. of Geomatics, Computer Science and Mathematics, Stuttgart University of Applied Sciences, Stuttgart (hossein.arefi, , m.hahn.fbv)@fht-stuttgart.de 2 TopScan, Steinfurt, Germany, lindenberger@topscan.de KEY WORDS: First and Last Pulse signals, LIDAR, Feature Extraction, Classification, Remote Sensing tools, TopScan ABSTRACT: During the last decade airborne laser scanning has become a mature technology which is now widely accepted for 3D data collection. Automated processes employ the scanned laser data and the platform orientation and other parameters of the scanning system to generate 3D point coordinates. These 3D points represent the terrain surface as well as objects on top of the terrain surface. Modern airborne LIDAR systems are able to record first pulse and last pulse range measurements together with the signal strength to provide more information about the reflecting surface or object. The main goal of this paper is to investigate procedures for the classification of the LIDAR data by relying on existing standard remote sensing tools and algorithms. Particular interest in this work is given to the detection of buildings and trees. It is well known that the first and last pulse range images provide essential information to differentiate between these object classes. Rather than developing an overall workflow for general object classification the procedures for classification of buildings and trees are investigated separately. As a common basis processing is divided into pre-processing, image classification and post-processing. In the pre-processing step the input bands for classification are generated based on first pulse data, last-pulse data and a normalised difference image. Image enhancement and noise reduction is included in pre-processing. Next, some training areas are selected and Maximum Likelihood classification is applied. In the post-processing step basically morphological filtering is used to eliminate artefacts in the classification result. The LIDAR data for the experimental investigation are taken from a scanning project in which the density of scanned points is around one point per square meter. The classification results are qualitatively and quantitatively evaluated. The results show that already with those standard processes almost all trees and buildings can be detected and correctly classified in the scenes. In regions with significant natural tree cover the separation between artefacts indicating small trees and really existing small trees leads to a conflict in post-processing. 1. INTRODUCTION Airborne LIDAR systems provide digital three-dimensional information about the Earth's surface. Developed on the principles of range finding an airborne laser mapping procedure can be characterized as a largely automated measuring procedure for fully digital data collection. Based on an integrated multi-sensor system with GPS and INS an airborne laser terrain mapper provides the 3D position of each laser beam spot on the Earth's surface. The most important feature of recent airborne LIDAR systems is its ability to discriminate first and last pulse reflections. A laser pulse that is fired over vegetation usually has multiple reflections. Some particles of the laser pulse may be reflected by leaves or branches of trees often represented in the first returning pulse. Others may be reflected by the ground and the last returning pulse is most likely to be reflected by the terrain surface beneath trees. The primary goal in the early development of laser mapping was the topographic mapping of forested terrain. In the meantime, laser mapping has conquered fields of application far beyond the originally intended ones (TopScan, 2003). Topographic mapping is still a major application of laser mapping. For providing sufficiently dense elevation data of forested terrain there is still no other surveying method available. The ability to provide a very high density of directly recorded height points of the topographic surface is a significant property of LIDAR systems. The increasing demand of up-to-date elevation models for a large variety of applications requires economical procedures which are able to supply spatial data even for large areas with sufficient accuracy within short time limits. Regarding the acquisition of digital elevation models laser scanning is in competition with procedures of digital photogrammetry. An increasing number of research reports indicate that the 3D LIDAR data are a highly qualified source for GIS data acquisition. This includes examples of automated processes for 3D reconstruction of complex urban environments with buildings, trees, roads and others (Workshop 3D reconstruction, 2003). A large user group of LIDAR may have to rely on standard remote sensing and image processing packages for further processing of the data. For these purposes the 3D coordinates of scanned laser points can be interpolated to a regular grid to be available as images. Images of range and intensity data can be used conveniently for investigations by those tools. In the following we will report about investigations for classification of trees and buildings based on standard remote sensing tools even though not relying on one software package only. In the next chapter the concept of the investigation is explained. Chapter 3 describes the experiments and discusses the obtained results. Conclusions are drawn in the final chapter.