LIDAR Data Classification using Hierarchical K-means clustering Nesrine Chehata a,b , Nicolas David b , Frédéric Bretar b a Institut EGID - Université Bordeaux 3 Equipe GHYMAC 1 Allée Daguin 33607 Pessac Email: Nesrine.Chehata@egid.u-bordeaux3.fr b Institut Géographique National, laboratoire MATIS 2 Av. Pasteur 94165 St. Mandé cedex, France Email: nicolas.david@ign.fr, frederic.bretar@ign.fr KEY WORDS: Remote Sensing, LIDAR, Hierarchical Classification, DTM, Multiresolution ABSTRACT: This paper deals with lidar point cloud filtering and classification for modelling the Terrain and more generally for scene segmentation. In this study, we propose to use the well-known K-means clustering algorithm that filters and segments (point cloud) data. The K- means clustering is well adapted to lidar data processing, since different feature attributes can be used depending on the desired classes. Attributes may be geometric or textural when processing only 3D-point cloud but also spectral in case of joint use of optical images and lidar data. The algorithm is based on a fixed neighborhood size that can deal with steep relief covered by dense vegetation, mountainous area and terrains which present microrelieves. The novelty of our algorithm consists in providing a hierarchical splitting clustering to extract ground points. The number of cluster splits is used to qualify automatically the classification reliability. This point is rarely treated in previous works. Moreover landscape predictors such as slope map are used to locally refine the classification. Finally, the methodology is extended to a multiscale framework. The hierarchical clustering is processed from coarse DTM resolution to finer one. This implementation improves the algorithm robustness and ensures reliable ground estimation. Quantitative and qualitative results are presented on the ISPRS data set. 1 I NTRODUCTION Representing the Earth’s topography, that is the vegetation, the true terrain, buildings as well as any human-made infrastructures from aerial remote sensors in a 3D virtual environment has been a challenging task for scientists for many years. Recent years have seen the development of airborne scanner systems which provide dense 3D point cloud of the surface topography. This massive amount of data has to be analyzed and classified to extract per- tinent informations. A Digital Terrain Model (DTM) is a fun- damental layer for any application in a 3D virtual environment, and as a matter of course, plays a main role when dealing with natural risk management. Several methods have been developed for filtering lidar data to generate Digital Terrain Models. Algo- rithms have to process large data volumes on various and com- plex landscapes such as urban areas Dell’Aqua et al. (2001), for- est areas Kraus and Pfeifer (1998); Haugerud and Harding (2001) or mountainous areas Wack and Stelzl (2005). Many algorithms have been implemented and tested so far, but no generic solution appeared Sithole and Vosselman (2003). Existing works on lidar data labelling can be divided into three major approaches that will be briefly detailed hereby: 1. Morphological filters These filters are based on a series of 3D morphological closings and openings. Robust meth- ods against measurement errors were proposed using a dual rank filter Eckstein and Munkelt (1995). The filter pa- rameters highly depend on the terrain slope as well as on the relevancy of laser points to belong to the terrain: last pulse is not always a true ground point, especially in pres- ence of dense vegetation coverage. Vosselman (2000); Sit- hole (2001) proposed a slope based filtering. In Kraus and Pfeifer (1998), authors have proposed an iterative linear pre- diction scheme to remove vegetation points in forest areas. The potential of morphological filters to provide a good es- timate of the ground depends on the filtering window size. A small window size leads to a fine local topography pro- vided that there are enough true ground points within the neighborhood. On the contrary, a large window size tends to smooth the final DTM. To overcome these effects, some authors refine locally the window size of the filter Kilian et al. (1996); Bretar and Chehata (2007). Zhang et al. (2003) have used an iterative technique using progressive morpho- logical filters by varying the window size to estimate dif- ferent height thresholds in local regions. Others propose a repetitive interpolation of DTM in forest areas Filin and Pfeifer (2006); Kobler et al. (2007) to improve the algorithm robustness. The advantage of the morphological filters is the short computing time but they need an accurate a-priori knowledge about the terrain topology. 2. Progressive TIN densification Some points are identified as ground points and based on those, new points will be added to the ground class Sohn and Dowman (2002). In Axelsson (2000), the authors present an iterative Triangular Irregular Network generation. From a coarse triangulated surface based on the lowest points, new lidar points are in- tegrated in a Delaunay triangulation under strong angle and distance constraints. The advantages of triangulation based methods are the short computing time and the robustness. However, the TIN surface is very sensitive to negative out- liers that may shift the surface downwards. 3. Surface model filters These filters are based on robust in- terpolation of ground points Kraus and Pfeifer (1998). A coarse surface is estimated. All points are weighted by a power function of their residuals to the approximated sur- face. The surface converges toward points with negative residuals. In Elmqvist (2002), the ground is estimated by an active shape model. The drawback of this approach is that it is controlled with many parameters and is sensitive to negative outliers. In addition to the filtering process, many authors tried to orga- nize the 3D cloud into multiple classes, using essentially unsuper- vised classification methods. The input data can be only 3D point cloud. Geometric or textural attributes are used. In Elmqvist et al. (2001), the height texture is the maximal local slope and the sec- ond derivative of the pixel and the 8-neighbouring pixels. Mul- tiple echos allow the distinction between buildings and vegeta- tion. Height texture is often processed over a regular interpolated