ISPRS Commission III, Vol.34, Part 3A ,Photogrammetric Computer Vision'*, Graz, 2002 
  
2), or if one crown presents more than one significant local 
maximum (case 3). In the first two cases one blob covers more 
than one crown, in the third case one crown presents more than 
one blob. Using significance, the position of potential crowns 
and a-priori information on potential crown diameters, these 
situations can be interpreted by means of various rules within 
auto-scale selection. The result of auto-scale selection are la- 
beled and all maxima belonging to one crown are labeled with a 
unique value. 
Crown segmentation 
Inverting the smoothed and thresholded tree height image 
The tree height model was inverted in a next step in order to 
adapt the given information to the needs and properties of clas- 
sical watershed algorithms. In short, watershed algorithms find 
local minima in a grayscale image and try to assign each pixel 
of the image to a local minimum. The problem to solve here 
was exactly vice versa. Each tree was represented by a single 
local maximum, and the pixels around the maxima should be 
assigned to the most probable maximum (treetop). A slightly 
adapted watershed algorithm could now be applied to the in- 
verted data in order to delineate single trees. 
Watershed segmentation 
There are several formal definitions and implementations of the 
watershed problem, which basically follow the same idea. More 
detailed information on the algorithm and formal definition can 
be found at Soille (1999) and Roerdink & Meijster (1999). 
Every grayscale image can be interpreted as a three-dimensional 
landscape, where each pixel's activation represents the altitude 
of the corresponding point in the landscape. Assuming a rain- 
drop is falling down somewhere onto the landscape according to 
the law of gravitation, it will flow down along the steepest slope 
path until it reaches a minimum. The starting points of this 
procedure are the labels produced by auto-scale selection. The 
Whole set of points whose steepest slope path ends in the same 
minimum, were assigned to this minimum. All pixels belonging 
to the same minimum were labeled with a unique value. In a 
next step each pixel in the image was assigned to a specific 
minimum. This decision could not be made for pixels with two 
or more steepest slope paths reaching different minima. These 
pixels were called watershed pixels and were labeled as such 
with a predefined value, different from all other labels. The 
watershed segmentation produced a partitioned image contain- 
ing a set of watershed pixels and segments which were con- 
nected regions labeled with a unique id. 
The presented watershed algorithm will run into trouble if the 
input image contains non-minimum gray-level plateaus. The 
difficulties occur when the steepest slope path reaches such a 
plateau. In this situation it is not clear in which direction the 
steepest slope path should be continued. Therefore the input 
image had to be turned into an image which does not contain 
non-minimum plateaus, without changing the steepest slope 
direction of non-plateau pixels. This was achieved by applying a 
lower completion algorithm, as is presented in Roerdink & 
Meijster (1999). This algorithm transforms non minimum pla- 
teaus into steep hills by iteratively increasing the plateau values 
from the borderlines to the center of the plateaus. This method 
ensures that after the transformation each pixel — with the ex- 
ception of minimum pixels - has a lower neighbor. Images 
fulfilling this precondition can then be deterministically seg- 
mented by the watershed algorithm presented. 
A - 306 
Verification 
Hohentauern test site 
Figures 5a and 5b represent segmentation results from the 
Hohentauern test site. For verification purposes 197 single trees 
scattered across the test site (8 reference areas with 15 — 30 trees 
measured within the area) were surveyed by ground measure- 
ments. The exact positions of the single trees were determined 
using a differential GPS system and accurate ground measure- 
ments (within 10 cm accuracy in x, y). The crowns of the refer- 
ence trees were delineated in the field; Bhd (breast height di- 
ameter) tree height and tree species were recorded. All the 
reference areas are homogeneous with regard to crown closure. 
The verification of the segmentation result was performed by 
comparing the segmented crowns with the crowns measured in 
the field. Table 3 depicts the overall accuracy of the segmenta- 
tion method within the Alpine test site. 
  
  
  
  
  
number of verification trees 197 
number of correctly segmented crowns 98 
number of crowns merged ! 89 
number of crowns split 9 
number of crowns not found 1 
  
  
  
  
! 64 of these crowns due to missing maxima, 20 due to wrong blob assessment, 5 
due to smoothing effects 
Table 3: Overall accuracy of segmentation method within the 
Alpine test site 
Two main errors occurred for crowns which had not been seg- 
mented correctly can be summarized as follows: 
- Within one tree crown two or more local maxima were 
detected in the tree height image (9 out of 197 cases). 
- Two or more crowns merge. This is mainly caused by 
missing local maxima in the tree height image, e.g. small 
trees in dense stands (68 out of 197 cases). 
  
  
  
  
{ 
  
Figure 5a: Area 3, crown closure 62 %, section of 50x50 m: left 
— segmentation result (outer circles) and reference trees (inner 
circles), right — tree height image and segmentation result 
Burgau test site 
At the Burgau test site 85 trees (spruce, pine and oak) were 
measured in the field. 62 % of these trees were segmented 
correctly. The reasons for segmentation errors are described 
above.