oe fd NEU MES TUNES IR. YUM A NP, id 
MS 
  
  
   
Aerial Imagery-based (Figure 2(a)): 
Three bands (R, G, B): To remove noises in the RGB image, 
convolution operation must be operated. In this paper, we use 
median convolution, a technique aiming at reducing image 
noise without removing significant parts of the image content, 
typically edges, lines or other details that are important for the 
interpretation of the image (Perona and Malik, 1990). After 
mean convolution, bands red (R), green (G) and blue (B) are 
used as three individual spectral features. 
Grey-level Co-occurrence Matrix (GLCM); GLCM proposed by 
Julesz (1962) can be used to calculate several statistical 
measures, such as contrast (Cont.), dissimilarity (Diss.), 
homogeneity (Homo.), entropy (Ent.), mean (Mean), variance 
(Var), second-moment(S-M) and correlation (Corr) for 
representing specific textural characteristics of the processed 
image. 
Lidar Data 
Although a 2D lidar range image is used in the presented land- 
use classification scheme, lidar height-based features are 
calculated by 3D original point clouds in a given spherical 
neighbourhood. Mainly determined by the point density, the 
radius of the given sphere is required to guarantee at least 6 
points to get involved in processing lidar features. As a result, 
height-based features can be computed. 
Height-based features (Figure 2(c)) 
o Height difference (Height-Diff): The distance is between 
the current point and the lowest point in a cyclone with 
radium of about 30m. 
o Normalized height ( nDSM=DSM-DTM): This feature will 
help distinguish elevated objects from the ground or near- 
ground objects (Haala and Walter, 1999). 
Local hei ht variation (Local-Hei-Var, 
    
    
the absolut 
    
    
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
© 
ue rm Toa a. 
distance between the maximum and minimum height 
values in 3*3 pixels or 3*3 m): This feature will assist in 
discriminating ground and non-ground objects. 
o Height difference between echoes (FL-Diff- First echo - 
last echo): This feature will help distinguish high-rise 
penetrable vegetation. 
o  Normalized Difference (FL-NDiff, a lidar-based vegetation 
(Hrirst echo ^ Hrast echo) Similar to 
index): It is calculated by TER d 
NDVI (Normalized Difference Vegetation Index) in 
multispectral image classification, FL-NDiff will highlight 
vegetation (Arefi et al., 2003). 
o Deviation angle of plane normal vector from the vertical 
direction (P-Deviation-Ang): This feature will assist in 
discriminating the ground with small values of deviation 
angles. 
o Distance from the current point to the local estimated plane 
(P-Normalized-Var): This feature reflects the local height 
variation that can be used for the discrimination of the 
ground and non-ground objects. 
o Eigen-based features (Anisotropy, Linearity, Planarity, 
Sphericity): The eigenvalue related features are defined as 
the spatial features of each point by calculating a variance- 
covariance matrix of its neighbours. It is another auxiliary 
indicator for distinguishing planes, edges, corners and 
volumes (Chehata et al., 2009). 
Intensity-based features (Figure 2(b)): 
o Intensity image: Analogue to a grey image, GLCM related 
measures are calculated. 
o Lidar-TVI (Transformed vegetation index): It is calculated 
Lidar Intensity -RED ; 
by DIT 4 0.5 based on Red band of aerial 
Lidar Intensity+RED 
imagery and intensity values of lidar data. 
  
  
    
GLCM-SM 
  
  
  
  
  
> 45 
2 ia 
GLCM-Ent 
  
     
  
  
  
@ 
     
  
  
  
»