In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009
vegetation detection). In addition, new synthetic bands were
generated from the selected channels: a) 3 images from
principal component analysis (PCI, PC2, PC3); b) one image
from NDVI computation using the NIR-R channels and c) one
saturation image (S) obtained by converting the NIR-R-G
channels in the IHS (Intensity, Hue, Saturation) colour space.
The separability of the target classes was analyzed through use
of plots by mean and standard deviation for each class and
channel and divergence matrix analysis of all possible
combinations of the three CIR channels and the additional
channels, mentioned above. The analysis showed that:
• G and PC2 have high correlation with other bands
• NIR-R-PC1 is the best combination based on the plot
analysis
• NIR band shows good separability based on the divergence
analysis
• PC1-NDVI-S combination shows best separability over
three-band combinations based on the divergence analysis.
Therefore, the combination NIR-R-PC1-NDVI -S was selected
for classification. The maximum likelihood classification
method was used. As expected from their low values in the
divergence matrix, grass and trees, airport buildings and
residential houses, airport corridors and bare ground, airport
buildings and bare ground could not be separated. Using the
height information from nDSM, airport ground and bare ground
and roads were fused into “ground” and airport buildings with
residential houses into “buildings”, while trees and grass, as
well as buildings and ground could be separated. The final
classification is shown in Figure 2. 84% of the building class is
correctly classified, while All of 109 buildings have been
detected but not fully, the omission error is 9% . Aircrafts and
vehicles are again mixed with buildings.
mr. ss ymwn
Figure 2. Building detection result from method 2. (Left: airport
buildings, Right: residential area).
4.3 Building detection using density of raw Lidar DTM and
NDVI (Method 3)
Buildings and other objects, like high or dense trees, vehicles,
aircrafts, etc. are characterized by null or very low density in
the DTM point cloud. Using the vegetation class from NDVI
channel as a mask, the areas covered by trees are eliminated,
while small objects (aircrafts, vehicles) are eliminated by
deleting them, if their area is smaller than 25m 2 . Thus, only
buildings remain (Figure 3). 85% of building class pixels are
correctly classified, while 108 of 109 buildings have been
detected but not fully extracted, the omission error is 8% .
Figure 3. Building detection result from method 3. (Left: airport
buildings. Right: residential area).
4.4 Building and tree detection from Lidar data (Method 4)
As mentioned above, in the raw DSM data the point density is
generally much higher at trees than at open terrain or buildings.
On the other hand, tree areas have low horizontal point density
in the raw DTM data. We start from regions that are voids or
have low density in the raw DTM (see Method 3). These
regions represent mainly buildings and trees and are used as
mask to select the raw DSM points for further analysis. In the
next step, we used a search window over the raw Lidar DSM
data with a size of 5 m x 5 m. Neighboring windows have an
overlap of 50%. The window size has a relation with the
number of points in the window and the number of the points in
the search window affects the quality of the detection result.
The method uses all points in the window and labels them as
tree if all parameters below have been met. The size of 25m 2
has been agreed to be enough to extract one single tree. A
bigger size may result in wrong detection especially in areas
where the buildings are neighboring with single trees. On the
other hand, the data has low point density: 1 pt / 2 m 2 , that
means about 13 pts / 25 nr. A smaller size will contain less
points and this may not be enough for the detection.
The points in each search window are projected onto the xz and
yz planes and divided for each projection in eight equal sub-
regions using x m j n , x m i d , x max , z mm z m j d ] z m j d 2 2 m j(j3 z max as
boundary values of sub-regions, with x mid = x min + 2.5m , x max
— x m id "F 2.5m, z m j d ]— z rn i n +(z max -z m j n )/4, z ivnd 2 ~z n - nn +2 (z max -
Zmin)/4, z mid3 =z min +3*(z max -z min )/4 and similarly for the yz
projection. The density in the eight sub-regions is computed.
The first step is the detection of trees and the second the
subtraction of tree points from all off-terrain points. The trees
have been extracted by four different parameters. The
parameters have been calculated using tree-masked areas of the
raw Lidar DSM data. The tree mask has been generated by
Method 2. Then, the calculated parameters (the average of all
search windows) have been applied to the raw Lidar DSM data
for detection of trees.
The first parameter (s) is similarity of surface normal vectors.
We assume that the tree points would not fit to a plane. With
selection of three random points in the search window, the
surface normal vectors have been calculated n (number of
points in search window) times. Then, all calculated vectors
have been compared among each other. In case of similar value
of compared vectors, the similarity value was increased by
adding 1. In the tree masked points, the parameter (s) has been
calculated as smaller than 2. The second parameter (vd) is the
number of the eight sub-regions which contain at least one
point. The trees have high Lidar point density vertically. Thus,
at trees more sub-regions contain Lidar points. Using the tree
mask, we have observed that at least 5 out of the 8 sub-regions
contain points. Thus, the parameter (vd) has been selected as