In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
249 
TREE SPECIES DETECTION USING FULL WAVEFORM LIDAR DATA IN A 
COMPLEX FOREST 
S. Gupta a ’ *, B. Koch a , H. Weinacker a 
3 Dept, of Remote Sensing and Landscape Information Systems (FeLis), Faculty of Forestry, Albert-Ludwigs University, 
Tennenbacher str. 4, 79106 Freiburg, i.Br., Germany - (sandeep.gupta, barbara.koch, holger.weinacker)@felis.uni- 
freiburg.de 
Commission VII, TC VII Symposium 
KEY WORDS: LIDAR, Extraction, Reconstruction, Algorithms, Pattern, Three-dimensional 
ABSTRACT: 
The three-dimensional single tree extraction by applying pattern recognition based modified clustering approach on full waveform 
normalized raw LIDAR data has been presented in this research work. The LIDAR data of medium density (16 points m' 1 2 ) was 
collected in August 2007 from the administrative forest district Hardt, Germany. The total study area is 1.75 ha and dominated by 
various deciduous tree species. The study plots selected contains multi-tier tree species of different age groups. Clusters of single tree 
extracted after running the algorithm were reconstructed using QHull algorithm. A validation procedure was devised and used for the 
accuracy assessment of the automatically detected tree species with respect to the forest inventoried data. The average producer’s and 
user’s accuracy for the total study area was around 56% and 41%, respectively. The results showed that the modified algorithm 
worked fairly well in the detection of evergreen conifers (79%) than the deciduous tree species (47%) beside the fact that conifers 
constitute roughly 18% of the total study area. The result showed that the algorithm for the upper tier trees species which are 
relatively mature and older worked better as compared to the tree species lying beneath the first-tier. The mixture of multi-storey tree 
species of varying age and height quintile with dense canopy cover was a limiting factor in the detection of single tree automatically 
in the presented work and shows the future scope of improvement in the algorithm applied. 
1. INTRODUCTION 
In a complex forest ecosystem, finding the distribution of 
different tree species of varying age and height quintile through 
traditional methods is a very thorny job. In the past one decade, 
the demand for high quality Light Detection And Ranging 
(LIDAR) data with more information has tremendously 
increased for various applications. The increasing demand of 
individual tree related information, as a basis to improve forest 
management performance, is the concerned factor for 
developing various methodologies for single tree extraction and 
related parameter estimation from airborne laser scanner (ALS) 
data. Clustering provide a good way of partitioning the whole 
normalized ALS dataset of the test area into an individual 
clusters. Because of the high point density full waveform 
LIDAR data provide a good platform to implement the 
clustering mechanisms via partitioning the data into individual 
clusters, each representing single tree. There are different 
clustering mechanisms, but the most popular ¿-means was 
chosen which is an iterative hill-climbing method and is a staple 
of clustering methods (Gupta et al., 2010). These has motivated 
to test the full waveform ALS data for the extraction of pattern 
of single tree crowns of different tree species in the selected 
plots of Hardt administrative forest district of Germany using 
modified clustering based approach and has been presented in 
the current work. 
2. EXISTING RELATED WORK 
Several studies has been carried out in the past on the 
application of airborne LIDAR data for vegetation related 
* Corresponding author. 
information retrieval using different methods (Hyyppa and 
Inkinen, 1999; Hyyppa et al., 2006; Ko et al., 2009; Nilsson, 
1996; Persson et al., 2002; Persson et al., 2006; Vauhkonen et 
al., 2009; Wang et al., 2008). Research work using clustering 
based approaches for 3-dimensional (3-D) single tree extraction 
using airborne LIDAR data has been carried out (Cici et al. 
2008; Doo-Ahn et al. 2008; Gupta et al., 2010; Morsdorf et al. 
2003; Morsdorf et al. 2004; Reitberger et al. 2008). Morsdorf 
et al. (2003) used first and last pulse data with an overall 
density of 30 points m' 2 and the ¿-means method to extract 
single tree in the Swiss National Park. In contrast to the 
modified algorithm used in the presented work, instead of 
scaling-down the height values, they scaled-up it by a factor of 
3. This has been done to accommodate the aspect ratio of pine 
tree crowns (ranged from 3 to 6). However, it was concluded on 
the basis of previous study (Gupta et al., 2010) that by scaling 
down the height value of the normalized raw LIDAR points and 
the external seed points (local maxima), squared error function 
is minimized which is the ultimate objective of the ¿-means 
method. The closer the points will be, more precise will be the 
cluster formation with regard to actual tree/tree crown and its 
shape. The algorithm used in the presented work differs from 
Morsdorf et al. (2003) in a way that unwanted local maxima 
points were deleted in the pre-processing step. Riano et al. 
(2004) estimated a derivative of foliage biomass, crown bulk 
density, using lidar metrics with ¿-means clustering at both plot 
and individual tree scales. However, individual tree level 
analyses were not completely successful in their work. In a 
study conducted by Ko et al. (2009) for deciduous-coniferous 
classification using single leaf-on high density full waveform