increment data, it shows that only 0.3% of data amount was 
increased by using CIR images. The detailed results are shown 
in Table 2. 
  
Figure 5. Overview of multispectral point cloud data 
Unit: Points 
  
  
  
  
  
  
  
  
Image Category Visible Light CIR 
Photosynth Data 124,377 127,103 
Similarity Points -- 126,715 
Dissimilarity Points -- 388 
Data increment by CIR Point Cloud Data 
= 388 / (388 + 124,377) = 0.3 % 
  
  
  
Table 2. Multispectral point cloud increment information 
4.4 Classification 
Multispectral image produces more spectral information than 
visible light images. Therefore, by using the xyz coordinates 
and color information generated from Photosynth, it is benefit 
in classification. This subsection focused on using the height 
and color information as threshold to classify some basic 
ground features in the research area. 
The CIR similarity points were used as input data. First, by 
choosing one elevation as base height, the data is divided into 
upper base and lower base. Then, NDVI was computed through 
the greyscale value of NIR and red band contained in 
multispectral point cloud. In stage 1, the lower base was 
classified through NDVI threshold. It was selected by viewing 
the NDVI histogram, afterward, the grasses and cement plane 1 
were classified individually. On the other hand, upper base was 
divided continually through height into upper base 1 and upper 
base 2. Later on, the upper base 1 and base 2 were classified 
respectively by NDVI threshold. Then, one can get trees 1 and 
buildings from upper base 1, in addition, trees 2 and cement 
plane 2 from upper base 2. Finally, by merging trees 1 and trees 
2, the category of trees can be provided. Figure 6 shows the 
classification flow chart. 
Through these classification thresholds, the CIR point cloud 
data can be classified into five categories, buildings, cement 
plane 1, cement plane 2, trees and grasses. Afterward, the 
threshold classification results were assessed by manually 
classification results. The illustrations of classification of both 
methods are shown in Figure 7, and the computed classification 
results are listed in Table 3. As the result shows, threshold 
classification has omission error around 36% in trees and 
grasses; and commission error in buildings and cement planes. 
Most of the building points and cement plane 2 actually include 
tree points, as shown in Figure 7. But one interesting point is, 
when classifying through NDVI, the grasses growing on the 
cement plane or bare soil within grasses can be detected. 
CIR Point Cloud 
Data 
Y 
Z > -Base Height Z« Base Height 
p X. Base Height ^— — — —3, 
me M nk re 
Base Down | | Base Up | 
>>> 
' Y Y 
I 1 
! >= <-0.2 2-21 «Zl 1 
| — NDVI Et Zi 1 
1 1 
1 i 
1 i 
| Stage 1 : 
1 | | 1 
Vip M ta ey Yl 
! Cement ! 
i I 
V i 
L- 
Plane 1 
    
Figure 6. Classification flow chart 
  
* 
  
  
  
  
(a) Manually Classified (b) Threshold Classified 
  
  
Figure 7. Different classification results 
Unit: Points 
  
  
  
  
  
  
  
Categories | Manually oes Commission | Omission 
Buildings 9,548 15,538 vA -- 
Comet | vow | ne | Ziad 
CEN 5,335 12,941 an d 
Trees 10,911 6,986 -- mn 
Grasses 45,070 30,640 -- TS 
  
  
  
  
  
  
138 
Table 3. Classification assessment 
5. SUMMARY 
According to the results, overall accuracy, classification results 
and the suggestions are described in following sections 
respectively. 
5.1 € 
The C 
were 
light | 
check 
4.420 
light 
diffict 
There 
the da 
was 1 
unable 
82 ( 
Altho 
image 
classi: 
value 
NDVI 
featur 
the R 
point 
shows 
but w 
see gr 
thresh 
advan 
5.3 € 
(1) TI 
in cor 
to the 
cloud 
coord 
(2) 1 
thresl 
adopt 
same 
cause 
(3) il 
three- 
Also, 
furthe 
Chen 
Anal 
Conf 
205, : 
Lowe 
Invar 
Conf 
Jense 
Eartl 
Inc., |