In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009
55
orthophotos were generated from the R, G, B and NIR channels
(a true colour orthophoto was also generated) and the NDVI
was calculated using the following formula:
NDVI = (NIR-R) / (NIR+R)
NDVI values range from -1 to +1 which suggests that if the
pixel value is close to -1 it does not belong to healthy
vegetation or vice versa. As a result, NDVI data could assist in
separating vegetation from buildings in a DSM.
Figure 3 shows the effect of sun position on selecting an NDVI
threshold to separate buildings from vegetation. In the shadow
area NDVI values are larger than the portion of the building
directly facing the sun. A larger threshold value of 0.3 was
selected to differentiate between buildings and vegetation.
Because of this large threshold value some vegetation also
appears with the buildings.
NDVI Values
Distance (m)
Figure 3: NDVI Threshold for Buildings
Within the LiDAR group of tasks, the first step is the
generation of a DSM and DTM from the LiDAR data
(TerraSolid software was used). In order to get the absolute
height of the objects the DTM was subtracted from the DSM to
give the NDSM. A further refinement of the NDSM can then be
achieved by making use of multiple LiDAR echo data. These
occur from building edges and trees. Figure 4 (Clode et al.,
2005) shows how the laser beam interacts with building edges
and trees.
Firstly, the filtered multiple echoes (figure 5) were converted
into an image. Gaps between pixels of less than 3 metres were
filled and a binary image was generated. Selecting a value for
gap filling depends on the density of the original point data. If
the density is high a small value can serve the purpose but it
should not be too high that it causes individual trees close to
each other to merge.
The separation of multiple echo data (Figure 5) from the
NDSM, by multiplication by the binary data, results in data
only from those objects that record a single reflection. These
include buildings and other solid objects but also vegetation
that returned single echoes.
\\ y* ;
»*-
„
/ f-n-"
‘*¿0$ : " u
. fa. ;r
; *
J v .. ' l * • *
" ( - {
'V
. . .*
r -*i” ' ~ •
*
■"
* *
o %
. H
* «.
, ' 4' ' :
.1?
V *
''' ' ¿V. J
!
<3/ > r >£*-
' * r,*> V
* • Vr •>,.
•
» « •'i
' K. *
+ <*■
. . t
■* . ■— .•*
* * • >
&
Figure 5: Filtered Multiple Echoes
The next step was to apply a height threshold of 2.5m to the
NDSM to eliminate objects such as hedges, cars etc and the
resultant NDSM containing buildings, vegetation and other tall
objects was converted to a binary image. All pixels having a
value lower than or equal to 2.5 m were assigned a zero value
and the remainder a value of one (Figure 6).
A morphological operation such as closing and opening was
used for filling small gaps in the binary image. Care should be
taken as too many repetitions can result in rounding of the
sharp building edges and loss of important detail.
This binary image contains pixels that belong to buildings and
remaining trees and needs further classification. This was
achieved by introducing the NDVI image described as part of
the image group of tasks.
The NDSM and the NDVI images were combined and the
maximum likelihood classification method was used for the
