ation with
^d regions
a nDSM),
e Normal-
itroducing
1sed unsu-
mbination
m (in this
ent distri-
x between
satisfying.
he case of
ns - espe
emphasis
ormalized
oded area
indicator
TS Or Sin»
differenti-
ings from
ngs (B) or
an be de
Jochen Schiewe
Figure 4 demonstrates the result of this
thresholding process: More than 9996 of
buildings or wooded areas are falling
above this value, but the number of
commission errors (ie., other objects
being also higher than this threshold) is
rather large which is mainly due to errors
within the DSM (especially in shadow
regions, see also the DSM profile in
Result of height thresholding figure 3).
Image data es
Mm correctly elassified Using this result the second stage tries to
commission errors *1/1*
[777] omission errors separate buildings from wooded areas
(and potentially other objects) which fall
Figure 4. Classification using nDSM-altitude threshold. into these candidate regions by means of
various indicators that shall be shortly
discussed in the following.
The hypothesis that the density of slope gradients is larger for wooded areas compared to buildings and other objects
(see figure 5) can be formalized through the mean number of adjacent pixels with height gradients of more than 50?
(which is larger than the presumable steepest roof slopes) within a local window (e.g., with 10e10 m? size). Our statisti-
cal analysis showed a good selectivity between classes as well as a good coincidence of the histograms with a normal
distribution. Hence, the probabilities for the membership of a pixel to a certain object class can be determined by using
the quantity of the normal distribution, for instance for buildings (B):
P( B | gradient density) 2 dXno of high gradient. neighbours; Up; Op)
The mean ji as well as the standard deviation o; can be derived from training. After computing and comparing this
measure for all object classes (building, wooded area, others), the maximum value indicates the most probable member-
ship of the pixel to an object class. For our test site this method works very well with only few exceptions like for single
trees or small lines of trees.
threshold = 50°
Building Wooded area
Figure 5. Slope gradient profiles for buildings and wooded area ( gradients in degrees).
In principle, the slope aspect - for example proposed by Baltsavias et.al. (1995) - gives evidence of associated buildings
if significant peaks 90° apart from each other can be observed, while the histograms of wooded areas do not show clear
peaks. In the course of our tests this indicator is found to be not that useful because irregularly shaped buildings disturb
the expected tendency and the number of gradient orientation values is too small for a reliable statistical evaluation.
With regard to spectral indicators, the Normalized Difference Vegetation Index (NDVI) is known to be a very good
Measure to differentiate between buildings and wooded areas. Unfortunately the above described design of spectral
bands of the HRSC-A sensor (section 3.1) does not allow for a clear distinction of vegetated and non-vegetated areas.
Also computing pseudo indices using other band combinations does not lead to better results. In principle, the probabil-
ity of a pixel belonging to wooded areas could be described for instance with
NDVI if NDVI > 0.15
else
PW | NDVI) =
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 811
