GIS Ostrava 2009 25. - 28. 1. 2009, Ostrava ___________________________________________________________________ Combining Multiple Classifiers for Automatic Road Extraction from Lidar Data Farhad Samadzadegan 1 , Behnaz Bigdeli 2 1 Dept of Geomatics, Faculty of Engineering, University of Tehran, P.O Box: 11365-4563, Tehran, Iran samadz@ut.ac.ir 2 Dept of Geomatics, Faculty of Engineering, University of Tehran, P.O Box: 11365-4563, Tehran, Iran Bigdeli@ut.ac.ir Abstract: The ultimate goal of pattern recognition systems in remote sensing is to achieve the best possible classification performance for recognition of different objects such as buildings, roads and trees .Extracting of road from newer Lidar data is one of the main challenge in photogrammetry and computer vision. Roads in Lidar data appear as homogenous area in same height. In this paper the idea is to combine classifiers with different error types by a learnable combiner which is aware of the classifiers' expertise, so that the variance of estimation errors are reduced and the overall classification of road accuracy is improved. In this paper we used Naïve Bayes and Weighted Voting Method as classifier fusion methods. The results quality was assessed for each classification method with the same validation set of pixels computing the confusion matrix. Experimental results show that the proposed model outperforms results with higher accuracy rather than single classifiers. Keywords: Road Extraction, Classifier Fusion, Weighted Voting, Naïve Bayes, Diversity, Correlation 1 Introduction Extraction of roads in complex environments as urban areas is one of the challenging issues in photogrammetry and computer vision, since many tasks related to automatic scene interpretation are involved.Auclair-Fortier et.al (2000) divide road characteristics in four different types: spectral, geometric, topologic and contextual and many different road detection techniques used this characteristics for its methodology. Compared to the relatively high number of research groups focusing their work on road extraction in rural areas, only a few groups work on the automatic extraction of roads in urban environments. Different methods used different data like high and low resolution image, Lidar, RADAR… and different level of automation for its algorithm. The recognition and accurate localization of objects in digital imagery has attracted considerable attention in the past in photogrammetry and computer vision. In semi-automatic schemes an operator selects an initial point and a direction for a road tracking algorithm (McKeon and Denlinger, 1988, Vosselman and de Knecht, 1995).Snake as a most important semi-automatic method used for detection and extraction of road. A few years ago many fully automatic road extraction proposed that minimized the role of operator. The fusion of different scales helps to eliminate isolated disturbances on the road while the fundamental structures are emphasized .In the coarse resolution, roads are modeled as bright lines and in fine resolution roads are assumed to have two parallel edges that be bright, and have a homogenous texture (Mayer and Steger, 1998). Road extraction in complex urban scenes was performed by Hinz and Baumgartner [4] from multi-view aerial images with a high ground resolution. They use a road model exploiting knowledge about the radiometric, geometric, and topological characteristics of roads, making use not only of the image data, but also of a Digital Surface Model (DSM). Lidar sensor technology is evolving rapidly and now allows the acquisition of very dense point clouds in a short period of time (Kraus, 2002). Alharthy and Bethel (2003) present a simple and fast method to detect roads in urban areas from Lidar data. The main aim of the work was to exclusively use Lidar data so that limitations of availability of other sources such as ground plans could be avoided. Both the intensity and height information were used to filter the raw Lidar data and remove “noise” that was unrelated to the road. Clode [1] implement road classification in a manner similar to Alharthy and Bethel (2003) .Again, both intensity and height information are used in the classification but the idea of a local point density is introduced. The local point density is an indicator of how many neighboring Lidar points have similar spectral and geometric properties to the Lidar point in question. The fact that roads are consistent in nature is an important model assumption. Approach proposed in this paper used Lidar data in urban area for automatic extracting of road network based on multiple classifier systems.