LIDAR POINT CLOUD BASED FULLY AUTOMATIC 3D SINGL TREE MODELLING IN 
FOREST AND EVALUATIONS OF THE PROCEDURE 
Yunsheng Wang 3 ,* , Holger Weinacker 3 , Barbara Koch 3 , Krzysztof Sterenczak b 
a Dept. of Remote Sensing and Landscape Information Systems(FeLis), University of Freiburg, Tennenbacher Str. 4, 
79106 Freiburg, Germany 
- (yunsheng.wang, holger.weinacker, barbara.koch)@felis.uni-freiburg.de 
b ZSIPiGL, Faculty of Forestry, Warsaw University of Life Sciences (SGGW), 
Ul. Nowoursynowska 159, bud. 34, 02-776 Warsaw 
- Krzysztof.Sterenczak@wl.sggw.pl 
KEY WORDS: LIDAR, Modelling, Aerial Survey, Feature Extraction, Forestry, Inventory, Precision 
ABSTRACT: 
A whole procedure of fully automatic 3D single tree modelling based on LIDAR point cloud is introduced in the paper. The 
evaluation of the procedure is then delivered by verifying the modelling results with field collecting data in sample plots. With the 
procedure, individual trees are extracted not only from the top canopy layer but also from the sub canopy layer, 3D shape of the 
extracted individual tree crowns are reconstructed, from which important parameters such as crown height range, crown volume and 
crown contours at different height levels can be derived. For the evaluation of the performance, the procedure is implemented with 
LIDAR data of 25 sample plots where detailed field inventories have been accomplished. Results of the procedure such as the 
number of individual trees in each sample plot, the location of the detected individual trees are verified by a statistical comparison 
with the field collecting data. Further analysis on the evaluation results is delivered at the final. 
1. INTRUDUCTION 
The individual tree levelled forest investigation has always been 
of high interest in forest management. As a relative new 
member of remote sensing instruments, airborne laser scanning, 
namely LIDAR, is especially suitable for reproducing the 
three-dimensional (3D) structure of forest stand due to its 
capability of 3D measurements with high accuracy. Several 
approaches of LIDAR based individual tree extraction in forest 
have been achieved during the past few years. The majority of 
the existed algorithms are DSM (Digital Surface Model) based 
(Hyyppa and Inkinen 1999; Persson et al. 2002; Koch et al. 
2006). Trees are delineated according to the features of crowns 
on the DSM, thus the individual trees in the lower canopy layer 
whose crowns are covered by the top canopy layer cannot be 
detected. Beside the detection of individual trees, Pyysalo and 
Hyyppa (2002) has provided a process for reconstructing tree 
crowns, with a pre knowledge of the location and the crown size 
of single tree, raw points belong to the tree are extracted, the 
height of the tree, the height of the crown and the average radius 
of the crown at different heights are derived. Further more, a 
new full-waveform based algorithm for detecting tree stems has 
been delivered by Reitberger et al. (2007). 
The procedure of LIDAR raw point cloud based 3D single tree 
modelling has been firstly published by Wang et. al. (2007), an 
improved version will be introduced in the second chapter of 
this paper. Another main interest of this paper is to evaluate the 
performance of the procedure. Accuracy verifications of the 
processing results in several sample plots are presented in the 
third part of the paper. Chapter 4 is for the analysis of the 
evaluation results. The LIDAR data as well as the field 
inventory data in sample plot being used in this paper was 
collected for a project on the utilization of LIDAR technology 
in forest inventory, financed by Polish General Directorate of 
State Forests and coordinated by Faculty of Forestry, Warsaw 
University of Life Sciences. 
2. 3D SINGLE TREE MODELLING 
2.1 Pre-Processing of LIDAR Raw Point Could 
The Pre-processing contains two steps. Firstly, the subdivision 
of data in area of interest, secondly, the generation of 
normalized point could. 
Due to the computational cost of memory and storage, it is not 
necessary to analyse all the points in a large area in one step. 
The study area can be divided into several small grids 
accordingly (Figure 1 -(a)). Further processes are concentrated 
with each cell separately. The favored size of a single study cell 
can be flexible from 10m* 10m to 200m*200m according to our 
own experiences. 
(a) (b) 
Figure 1. (a) Subdivision of a study area, study cells are marked 
with yellow lines over a DSM; (b) Comparison between original 
LIDAR raw point cloud and normalized point cloud; Left: 
original raw point cloud over a DTM; Right: normalized point 
cloud over a zero height level surface 
The procedure is based on the analysis of the object height of 
the captured points. Thus, the influence of terrain must be 
eliminated. A raster DTM (Digital Terrain Model) is used for 
the normalization of raw point heights. As shown in Figure l.(b) 
Left, raw points are projected above the DTM, the height 
difference between a raw point and its correspondent terrain is