LIKELIHOOD AND ACCURACY ANALYSES OF 3D BUILDING MODELS FROM AIRBORNE LASER DATA Fabio Crosilla, Alberto Beinat, Domenico Visintini, Barbara Fico, Emiliano Sossai Department of Georesources & Territory, University of Udine, via Cotonificio, 114 I-33100 Udine, Italy crosilla@dgt.uniud.it, interreg.cart@uniud.it KEY WORDS: laser scanning, building model, comparison, quality, accuracy. ABSTRACT: Airborne laser scanning can be employed today for the production of large scale maps and the creation of detailed 3D city models. For this purpose, several solutions have been proposed, also available in commercial software, to make easier and semi-automated the extraction and the geometric modeling of buildings from LIDaR data. As a consequence, while 3D urban modeling is gaining popularity among architects, engineers and urban designers, proper investigations are addressed to evaluate the likelihood and the accuracy of this new kind of survey products. In this paper, after a concise analysis of the possible errors affecting the building model construction, and a brief description of the principal methods to achieve it, we illustrate our experience in the realization of a detailed 3D model of the city of Gorizia (NE Italy), using the TerraScan software on two different laser scan datasets taken at different sampling densities (2 points/m 2 and 15 points/m 2 , respectively). Finally, we discuss the results of a test involving one set of 25 buildings, extracted from the city model and topographically surveyed by reflectorless EDMs and GPS equipment, in order to establish their effective size and shape. The comparison between the models obtained from the LIDaR and the real ones from the conventional survey outline that the likelihood and accuracy of the models derived from a 15 points/m 2 laser surveys are equivalent to those obtained by topographic measures. 1. INTRODUCTION Nowadays, the airborne laser scanners represent the most advanced technique for the field survey thanks to their excellent levels of automation, efficiency and resolution, consisting in a sample frequency up to 100.000 3D-points/s, a decimetre accuracy and a point density of 10-20 per square meter. The features, the capabilities and the products of the laser surveys are well-known and widely reported in the literature. The first comprehensive collection was published in 1999 as a Special Issue on “Airborne Laser Scanning” of the ISPRS Journal of Photogrammetry and Remote Sensing (Wehr and Lohr, 1999). Among the different aspects, the evaluation of likelihood and accuracy of the laser surveying has been deeply analyzed in these years (e.g. Baltsavias, 1999, Ahokas, Kaarttinen and Hyyppä, 2003; Kraus, Briose, Attwenger and Pfeifer, 2004). In this regard, the test of a cartographic product is not a new topic in the surveying field, but the check of the laser data used to this purpose concerns some new aspects that are not taken into account when testing a digital map from photogrammetric images. The laser output is in fact a cloud of “randomly shot points” and this conditions the procedures and the response of the tests. First of all, the height and the planimetric components have to be necessarily separated. Furthermore, while the elevation accuracy can be estimated, although only in an indirect way by means of horizontal planes, it is practically impossible to compute planimetric precision, unless to suitably exploit ramps of known slope (Casella, 2001). Moreover, if some years ago the final output of a laser survey was simply the point cloud, or the corresponding Digital Surface Model (DSM), nowadays we consider the 3D-model of the surveyed buildings. In fact, the current software for processing laser data allows the production of building models (e.g. TerraScan), and so an increasing number of 3D city models have been already accomplished. The evaluation of likelihood and accuracy of these laser-derived building models, subject of this paper, is an emerging problem, recently analyzed by Ahokas, Kaarttinen and Hyyppä (2004) and by Casella, Grampella and Franzini (2004). On the same argument, an international test “Building extraction from aerial images and/or airborne laser scanning” has been proposed in 2004 by the EuroSDR. The results will be presented during the incoming First International Workshop on “Next Generation 3D City Models” (Bonn, June 2005). The likelihood and the accuracy of the 3D building models depend on the propagation of the errors made during each step of the laser processing. In the following, these phases are shortly reported, so to better understand how many possible causes can affect the final three-dimensional positioning of the building model edges. 1. Airborne data acquisition. The acquisition stage is characterized by some main geometrical parameters, as (e.g. Baltsavias, 1999) acquisition frequency, relative height of the flight, number of echoes, and maximum swath angle. The combination of these values define the sampling density, and has a strong influence on the surveying productivity, on the risk of occlusions, and on the points accuracy in particular way. 2. Data geo-referencing. By exploiting the so-called direct orientation of the measurement sensor, this step involves a combination of 3D-rotation and 3D-translation transformations to refer the generic scanned point with respect to a global frame (generally WGS84) via the laser-frame and the INS-frame (e.g. Schenk, 2001). The position depends on the combination of six three-dimensional quantities, four of them instantaneously variable (GPS position, INS attitude, laser distance, laser rotation): the final accuracy can be very irregular, also within the same strip, e.g. it is poorer along the sides of the strip. 3. Scan strips data adjustment. Since the laser data geo-referencing is not over-determined, systematic errors in final positioning are possible: they can be minimized by making use of analytical constraints or by recalibrating the laser system (e.g. Burman, 2002).