shold Classified — | 
results 
Unit: Points 
  
  
  
  
  
  
  
  
  
  
nission | Omission 
990 
5594) m 
661 
74%) m 
588 
64%) I 
E 3,925 
(35.9794) 
- 18,718 
(37.92%) 
nent 
lassification results 
ollowing sections 
5.1 Overall Accuracy 
The CIR point cloud data generated by multispectral images 
were transformed into the local coordinate system of visible 
light point cloud, for data increment analyzing. The maximum 
check point positioning error is 6.499 unit, and the RMSE is + 
4.420 unit. It can be considered that although CIR and visible 
light point cloud data covers the whole research area, but it is 
difficult to ensure the control points are the same feature pairs. 
Therefore, the RMSE seems not so well. Due to high RMSE, 
the data increment analyzed by point to point distance threshold 
was influenced. Most of the probable increment data were 
unable to detect, only 0.3% of data increment was calculated. 
5.2 Classification Results 
Although it is difficult to tell the benefit by adding NIR band in 
image matching, but it is sure to improve the ability in 
classification. By using the xyz coordinate and RGB greyscale 
value provided in Photosynth, it is possible to compute the 
NDVI, which can be considered as a criterion to classify ground 
features. By combining the z coordinate from point cloud, and 
the RGB greyscale value from multispectral images, the CIR 
point cloud data is classified into five categories. The result 
shows approximately 36% of commission and omission error, 
but when focusing on the category of cement plane 1, one can 
see grasses being classified out of cement by using the NDVI as 
threshold. This makes the multispectral point cloud more 
advance for classification than LiDAR point cloud. 
5.3 Suggestions 
(1) The three-dimensional affine coordinate transformation used 
in combining two data still remains some systematic errors, due 
to the manually selection of effective control points. Other point 
cloud registration method should be considered for better 
coordinate transformation results. 
(2) Besides of using NDVI and height as classification 
threshold, other index value or band greyscale value may be 
adopted. Also, one should avoid illumination difference in the 
same view angle of visible light and NIR images. Since this may 
cause wrong NDVI while computing. 
(3) The classified point cloud data can be used in visualizing a 
three-dimensional idea of traditional two-dimensional results. 
Also, DEM generation and building tree models are available in 
further studies. 
6. REFFERENCE 
Chen, Szu-Han and Jin-Tsong Hwang, 2010. Precision 
Analyses of Photosynth point cloud data. In: The 29" 
Conference on Surveying and Geomatics, Taipei, CD-ROM, S- 
205, 8 pages. 
Lowe, David G., 1999. Object Recognition from Local Scale- 
Invariant Features. In: Proceeding of the International 
Conference on Computer Vision, pp. 1150-1157. 
Jensen, John R., 2007. Remote Sensing of the Environment: An 
Earth Resource Perspective, 2"* Edition. Pearson Education, 
Inc., pp. 443-506. 
139 
Wolf, Paul R. and Bon A. Dewitt, 2004. Elements of 
Photogrammetry: with Applications in GIS. 3" Edition. 
McGraw-Hill, pp. 518-550.