Jochen Schiewe
Len, |
730 a DSM m
cU T ^
ET rofile
nn e P I——]
= ^
= in X x
Tr À
I sou )
^,
tO aa in
V. d 3 m
690 Mun :
n B 10 15 eT 25
T J |
"CEU D
= 0 Approximated ex
* DTM profile
r 730 E.
710
aiii B3ü : T T T
Image data 0 5 10 15 20 25
Distance [m]
Figure 3. Example for DTM approximation through compressing opening.
33 Object extraction
Experiences from interactive stereoplotting have shown the importance of using elevation data in combination with
spectral information for visually detecting and describing objects. The general idea is that buildings and wooded regions
can be distinguished from roads and flat vegetated areas through a significant height difference (derived from a nDSM),
The discrimination between classes of similar heights should be possible using spectral information (e.g., the Normal
ized Difference Vegetation Index, NDVI).
3.3.1 Discussion of previous work. Haala and Brenner (1999) propose a multispectral classification by introducing
normalized height information as an additional channel. The disadvantage from our point of view is that the used unsv-
pervised classification makes the process less transparency and not generally applicable. Hug (1997) uses a combination
of laser altitude and reflectance data, thus reducing the spectral information content to a very narrow spectrum (in this
case with A = 810 nm). While other parameters like “elevation texture”, gradient variation or directional gradient distr
bution are of minor importance, the laser signal reflectance is found to be most useful for discriminating between
buildings and forest regions. But in fact, the reported classification error rates (up to 23%) are by far not satisfying.
Baltsavias et.al. (1995) propose a separation of buildings from trees by looking at the slope aspects which in the case of
a single building show significant peaks at 90? apart from each other. - In practise, still a couple of problems - espe:
cially with the separation of buildings and wooded areas - occur with the described methods.
3.3.2 Our approach. Considering the advantages and drawbacks of existing approaches we will place the emphasis
on the detection of buildings and wooded areas using a combination of various indicators derived from the normalized
Digital Surface Model (nDSM) and from multispectral imagery. The decision for an object class (building, wooded area
or others) is based on probabilities for the membership of a pixel to a certain object class with respect to every indicator
which are then combined with a maximum a-posteriori estimation rather than relying only on single parameters or sit
ple thresholding operations. In order to decrease the search space we rely on a two-stage concept that firstly differenti
ates between buildings or wooded regions against other objects, while the second stage tries to separate buildings from
wooded areas.
In the first stage a "safe" nDSM-altitude threshold of 2.0 m is applied in order to differentiate between buildings (B) tI
wooded areas (W) against other objects. The probability of a pixel belonging to one of the desired classes can be d
fined for example by using the linear relationship
nDSM altitude — 2.0
P(BUOWInDSM _ altitude) ={ MAX (nDSM _ altitude) — 2.0
0 else
if nDSM _ altitude > 2.0
810 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
The h
(see fi
(whic
cal an
distrib
the qu
The m
measu
ship of
trees o
In prin
if signi
peaks.
the exp
With r
measur
bands
Also c
ity of a
