In: Wagner W., Székely, B. (eds.): ISPRS ТС VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
3.6 Validation Method 
The automatically extracted tree tops of individual tree in each 
plot were validated with respect to field measured reference tree 
tops. A circular buffer of radius 3 m was created around each 
reference point, in Geographical Information System (GIS) 
environment. Only those extracted tree tops were considered 
which were close to a maximum of 5 m in 3-D Euclidean 
distance (ED) with that of the reference point. The extracted 
tree tops were intersected with that of the reference points of the 
identical plot within the buffered area. 3-D ED was calculated 
for each intersected point. Most suitable validation class as 
defined below was assigned for each intersected point. 
3.7 Validation Classes 
Following five validation classes were adopted for classifying 
the result and accuracy assessment. 
(i) Exact (E) - only one extracted tree top with respect to the 
nearby reference tree top. 3-D ED between the extracted tree 
top and reference tree top point is < 3 m. 
(ii) Nearly Exact (NE) - one extracted tree top with respect to a 
reference tree top nearly at the same height level. 3-D ED 
between the extracted tree top and reference tree top is 3-5 m. 
(iii) Split - more than one neighboured detected treetops up to 
the 3-D ED of 5 m from a neighbouring reference tree top. 
(iv) Missing - includes those reference tree top points for which 
there is no extracted tree top point in the neighbourhood up to 
the 3-D ED of 5 m. It also includes the reference points for 
which there is no detected tree tops within the buffer around 
each reference tree top point. 
(v) Extra - includes those extracted tree top points within the 
field boundary for which there is no reference tree top point up 
to the 3-D ED of 5 m. It also includes those extracted points 
within the field boundary for which there is no reference point 
within the buffer around each reference tree top point. 
4. RESULTS AND DISCUSSION 
Before running the modified £-means algorithm, normalized 
raw LIDAR points and local maxima points below 5 m height 
were filtered. This was done to avoid the effect of low ground 
vegetation and other smaller objects during the clustering 
process. After running the algorithm over normalized LIDAR 
points using local maxima as external seed points, the 3-D 
cluster points of the corresponding tree were extracted in all the 
study plots. Accuracy assessment of the five major validation 
classes of automatically detected tree tops with reference to the 
field measured tree tops has been presented (Table 3 and 4). 
Two validation classes, namely, (‘Exact’ and ‘Nearly Exact’) 
played a key role in determining the two kind of accuracy. 
Plot 
ID 
E 
NE 
S 
M 
Ex 
Xep 
Хегр 
FD 
(%) 
1 
17 
4 
8 
11 
15 
50 
29 
58 
2 
15 
10 
6 
24 
24 
70 
45 
64.3 
3 
17 
5 
3 
14 
8 
44 
22 
50 
4 
11 
3 
2 
8 
9 
30 
16 
53.3 
5 
10 
30 
7 
61 
6 
97 
57 
58.8 
6 
13 
21 
9 
13 
12 
67 
33 
49.3 
7 
2 
9 
3 
19 
21 
48 
37 
77.1 
Table 3. Distribution of validation classes and other attributes 
E = ‘Exact’ points, NE = ‘Nearly Exact’ points, S = ‘Split’ 
points, M = ‘Missing’ points, Ex = ‘Extra’ points, £ep = sum of 
extracted tree top points in the plot, £егр = sum of extracted 
error tree top points in the plot, FD = False detected points = 
Хегр* 1 00/£ ep . 
Plot ID 
E+NE 
Xep 
RP 
P асу (%) 
U acy(%) 
1 
21 
50 
32 
65.6 
42.0 
2 
25 
70 
49 
51.0 
35.7 
3 
22 
44 
36 
61.1 
50.0 
4 
14 
30 
22 
63.6 
46.7 
5 
40 
97 
101 
39.6 
41.2 
6 
34 
67 
47 
72.3 
50.7 
7 
11 
48 
30 
36.7 
22.9 
Table 4. Plot level accuracy 
E+NE = sum of exact and nearly exact points in the plot, Xep = 
sum of extracted tree top points in the plot, RP = total reference 
tree top points in the plot, P_acy (%) = producer’s accuracy = 
(E+NE)*100/RP and U_acy (%) = user’s accuracy = 
(E+NE)*100/Xep- 
It is visible from Table 4 that the producer’s and user’s 
accuracies in the broad-leaved deciduous dominated study plots 
(all plots except 3) are roughly varying between 37-72% and 
23-51%, respectively. In case of plot 3, which is dominated by 
evergreen coniferous trees, the producer’s and user’s accuracies 
are approximately 61% and 50%, respectively. There are 
highest accuracies obtained in plot 6. The high producer’s and 
user’s accuracies in case of plot 6 are detected due to fewer tree 
species which are present nearly the same height level. Due to 
this factor closely-matched seed points were generated that 
resulted in a comparatively more accurately positioned detected 
tree tops with respect to the referenced tree tops. The false 
detection is lowest (49%) in plot 6 and highest in plot 7 (Table 
3). In case of plots 2 and 7, the false detection is 64% and 77%, 
respectively, which is relatively higher than the remaining plots. 
It is noticeable that in both the plots, there are higher 
proportions of cherry trees. It is assumed that it was found 
difficult in automatic detection of small crowned and low height 
cherry trees. The accuracies are comparatively lower in case of 
study plots 5 and 7. In the former case (plot 5), it is mainly due 
to the mixed distribution multi-layered Oak and European beech 
with dense canopy. In the later case (plot 7), it is due to the 
presence of highly mixed tree species composition of varying 
age and high canopy density. In case of Oak dominated plot 5, 
the accuracies are not only lower but more or less in the same 
range. The average producer’s and user’s accuracies among all 
the study plots are 55.7% and 41.3%, respectively. 
Table 8 shows the percentage distribution of ‘Exact’ and 
‘Nearly Exact’ tree tops together and the corresponding 
reference tree tops for each species in all the 7 study plots.