and had little success in detecting the presence of the understory 
vegetation strata. Korpela et al. (2012) used airborne LIDAR 
data to study the understory trees by designing a conceptual 
compensation model for the transmission losses of laser pulses 
through overstory canopies. However, it was still an area-based 
detection and assessment method of the understory. Ferraz et al. 
(2012) has applied mean shift clustering to airborne LiDAR 
data of a multi-layered forest to extract single trees, assuming 
the spatial pattern of forest and boundaries of forest stratums is 
known in advance. They achieved a detection rate of 12.8% for 
the suppressed trees. 
Full waveform LiDAR systems can overcome drawbacks of 
conventional laser scanners by detecting significantly more 
reflections in the understory forest strata, and providing the 
intensity and width of pulses as reflectional parameters. The 
objective of this paper is (i) to develop an enhanced approach 
that detects single trees for multilayered forests with an 
integrated 3D segmentation, (ii) to enable the new approach to 
utilizes the geometric and reflectional features derived for local 
dense modes object level) of point clouds (iii) to show how the 
detection and location of single trees across datasets of different 
properties are achieved using the developed approach. 
The paper is divided into five sections. Section 2 focuses on the 
detection of single trees by combining normalized cuts with 
mean shift clustering. Section 3 shows the results which have 
been obtained from full waveform data acquired in the Bavarian 
Forest National Park. Finally, the results are discussed with 
conclusions in sections 4 and 5. 
2. METHODOLOGY 
2.1 Decomposition of full waveform data 
As usual, a single waveform is decomposed by fitting a series 
of Gaussian pulses to the waveform which contains Nr 
reflections (Figure 2). 
  
"i Y % 
Figure 2 3D points and attributes derived from a waveform 
The vector X" =(x,y,,z, W,,1,) is provided for each reflection i 
with (x,,y,,z;) as the 3D coordinates of the reflection. 
Additionally, the points X; are given the width W, - 2:c, and 
the intensity 7, 2 2-7 -0,- A, of the return pulse with o; as the 
standard deviation and A; as the amplitude of the reflection ; 
(Reitberger et al, 2009; Jutzi and Stilla, 2005). Note that 
basically each reflection can be detected by the waveform 
decomposition. 
The sensor data are calibrated by referencing WW; and /; to the 
pulse width J/^ and the intensity /° of the emitted Gaussian 
pulse and correcting the intensity with respect to the travel 
length s; of the laser beam and a nominal distance sy. 
Ww? -w,[w* (1) 
I; 20, sD[Q* s) Q) 
The correction assumes a target size larger or equal to the 
footprint (Wagner et al., 2006). The points from a waveform are 
subdivided into four point classes depending on the order of 
reflections within a waveform. 
2.2 Singe tree detection 
2.2.1 Local maximal filtering 
The coarse detection of single trees is achieved by searching 
local maximal in CHM, which is derived by subdividing the 
ROI into a grid having a cell spacing of cp and Nc cells. Within 
each grid cell, the highest 3D point is extracted and adapted 
with respect to the ground level. The ground level is estimated 
from a given DTM. In the next step, all the highest 3D points 
XT=(x,,V, 24") =1..,Nç) of all Nc cells are robustly 
interpolated in a grid that has Ny and Ny grid lines and a grid 
width g,. Both steps are carried out simultaneously in a least 
squares adjustment. The result is a smoothed equally spaced 
CHM. The local maximums derived on the CHM act as 
potential positions where single (overstory) trees could be 
located and can be used as prior knowledge in controlling 3D 
segmentation. The results could be improved by an additional 
stem detection method to further detect sub-dominated trees 
which are not represented by local maximums, when sufficient 
stem points are available. 
2.3.» Mean shift clustering 
Mean shift (MS) is a versatile tool for feature-space clustering. 
MS has been successfully applied to image segmentation tasks 
by exploiting the spectral-spatial feature space (Comaniciu and 
Meer, 2002). As the feature-based analysis depends on the 
quality of selected features, the derivation of feature set play a 
fundamental role in design of a segmentation algorithm. Since 
we want to avoid the bias caused by deriving geometric features 
such as height textures, planarity and curvature caused by 
neighborhood selection, the 3D geographic space of forest 
stands spanned by X; (x; , y; ,z; ) coordinates of point clouds is 
chosen to explicitly represent the feature space. ALS point 
clouds convey a multimodal distribution in which each given 
mode defined as a local maximum in density correspond to a 
crown apex or a part of crown (Ferraz et al., 2012). MS vector 
is defined as 
Sw XE (le SM np) 
Mi gl 3) = mes X 3) 
XL e(l -x)/Al) 
where x is the center of the kernel (window), and / is a 
bandwidth parameter for the kernel Given the function 
g(x) = —k'(x) for profile, the kernel G(x) is defined as 
G(x) = g(x] - 
  
  
Figure 3 Cylindrical shaped kernel for density estimation with 
horizontal Gaussian profile 
350 
In th 
cylin 
clou 
tree 
clust 
point 
of or 
assoc 
805; 
800: 
795+ 
790 
785 
780 
  
inten: 
provi 
defin: 
reflec 
e Sw : 
entire 
limite 
distin 
2.2.4 
Withi 
norm: 
cloud 
(Figui 
the ui 
the C| 
laser 1 
width 
Addit 
used 
applie