ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision“, Graz, 2002 
  
Statistical analyses have shown that the laser-derived predicted 
top heights describe 72 % of the variability of timber volume in 
forest inventory data (R square = 0.72). The corresponding 
standard error obtained amounts to 76 m? / ha. Typically, stand- 
wise inventories of the most important forest parameters are 
carried out with an error of about 15 % (in this case 55 m? / ha 
for timber volume). This error affects the obtained R square and 
standard error and thus the outcome of the previous analysis. 
Assuming that these two errors (predicted timber volume and 
timber volume derived from forest inventories) are independent 
of each other, the corrected laser-based error can be estimated 
by taking the root of the difference of the squared errors, de- 
noted by s, 
sz 767557 
which is even slightly better than the error of the conventional 
forest inventory (s = 52.5 m? / ha). 
Conclusion 
In summary this study demonstrates the feasibility of using laser 
scanner data for the stand-wise assessment of timber volume for 
forest inventories. 
4.2 Method and Results of the Tree-wise approach 
Different image processing steps are necessary in order to de- 
rive the tree wise parameters listed in Chapter 2. The methods 
described in the following section are suitable especially for 
conical-shaped coniferous trees, as it was developed for the 
Alpine Hohentauern test site. 
4.2.1 Processing line for segmentation 
In the following the whole processing chain is explained, start- 
ing with the tree height image and ending with a segmented 
image, where each homogenous region represents a single tree. 
Figure 4 depicts the working steps for segmentation (Ziegler et 
al. 2000). The basis for these processing steps are first pulse 
data which are resampled to a grid of 25 cm resolution and the 
forest floor model described in Chapter 2. 
Inverting R 
  
  
   
    
     
      
     
  
Threshold Watershed 
segmentaion 
Smoothing 
  
  
  
Maxima Auto-scale 
detector selection 
  
  
Tree Crown 
height area 
image segments 
  
  
  
  
  
  
  
  
  
Figure 4: Working steps for single tree segmentation 
Threshold filter 
Most forests are a conglomerate of large trees, young trees, 
bushes, grasses and other undergrowth. But the main interest of 
forest economists is focused on large trees representing the 
major part of timber volume, while young trees and under- 
growth are only of minor interest for timber volume assessment. 
Therefore, all pixel values representing heights less than a 
certain threshold (6 m in this case) were not taken into account 
for further processing. Another reason for neglecting these 
small trees was that they mostly stand close together and thus 
cannot easily be segmented. Although the spatial resolution of 
the TopoSys laser scanner is higher than provided by other 
systems, it is not high enough to capture every single crown. 
Smoothing filter 
In order to obtain optimal results for “tree top detection” 
(maxima detector) and “crown diameter” (watershed segmenta- 
tion)” each single tree should be represented by a blob with one 
single intensity maximum at the position of the tree top. Opti- 
mally, the gray values should decrease with the distance from 
the top. Looking at geometrical models of some tree species 
derived by ground measurements, one can see that most trees, 
especially conical-shaped coniferous trees, come close to this 
ideal shape. Nevertheless, as laser scanning data represent the 
“real situation”, one single tree crown can be represented by 
two or more intensity maxima and, thus, single trees do not 
appear as compact blobs. This smoothing step was introduced to 
eliminate smaller peaks and to approximate the height image to 
the ideal model. Because of its unique properties a Gaussian 
Kernel was chosen for the smoothing task. One of the advan- 
tages of this filter, which was very important for this applica- 
tion, is that Gaussian smoothing does not produce new extrema 
when increasing the smoothing scale. In spite of the good per- 
formance of the filter applied, single branches sometimes cause 
local maxima even after smoothing. This circumstance must be 
taken into account in the subsequent processing steps. Further 
information about continuous discrete Gaussian smoothing 
kernels, their properties and the setting of kernel parameters can 
be found at Lindeberg (1993 and 1994). 
Maxima detector and auto-scale selection 
After smoothing, coniferous trees mostly show a characteristic 
tip at the top of the tree. This is true if the local gray value 
maxima within the data image can be interpreted as tree tops. 
Local maxima must be interpreted carefully (see auto-scale 
selection), since they are caused not only by tree tops but also 
by the single branches remaining after filtering. 
Maxima are localized by means of a maxima detector in a first 
step. As tree crowns can present more than one local maximum 
and each maximum would produce one segment in the subse- 
quent watershed segmentation process, the number of tree tops 
would be overestimated and large crowns would be split into 
several segments. Hence, each detected maximum must be 
reviewed. Assuming that single crowns appear as blobs with a 
gray value maximum in the center and decreasing gray values 
towards the edge, diameter and significance can be assessed for 
blobs by means of a scale-space approach (Lindeberg 1994, 
Roerdink & Meijster 1999). This algorithm searches within a 
predicted range of blob diameters of typically 1m to 10m for 
maximum responses. The resulting blob diameter (ideally pre- 
senting a single crown) is then determined by the maximum 
significance of all potential crowns. 
The detection of blob diameters allows single crowns to be 
identified in most cases. Errors are caused, if crown density is 
very high (case 1), if larger crowns shadow smaller trees (case 
A - 305