ed above are
isting of the
reducing the
0 to 9.
d the classes
ssification is
iSpec called
fication of
algorithm is
that first
.omogeneous
X the initial
onal training
1e data were
new training
in Figure 11.
e used here
complex
orrection for
absolute
tion from
Aacintosh G3
significantly
building roof-
vation model
radient-based
lata show that
ent to sharply
undaries. À
t have region
ever, building
n scene are
's of condition
, the material
in streets and
lem, even for
)EM data can
daries, reduce
it for classifier
elevation of
> of a gradient
igher elevation
t operator was
o the result to
Edward M. Mikhail
The analyses can now be compared: Spectral analysis focuses on pixel-wise identification of the class rooftop. Note
that the task desires the identification of a specific usage (rooftop) in the scene, rather than the material classification
provided by spectral analysis. Thus, there is the possibility that spectrally similar materials will be identified with the
roof class, regardless of the manner of their usage in the scene. Gradient operator based analysis identifies the building
boundaries. In essence, the latter is a scheme to delineate building boundaries, while the other is a pixel classification
scheme. Figure 14 shows extracted rooftops.
The output in Figure 13 outlines buildings as objects with thick boundaries. It is possible to thin the delineated scene
objects by setting a high threshold on the output of the gradient operator. However this requires operand manipulation
on the part of the analyst, and is inefficient.
In general, spectral analysis is more robust over an extended scene. For instance, should the analyst note a different
type' of building rooftop in isolation, the set of scene-classes can be enlarged and training data included appropriately.
On the other hand, analysis of the DEM can be complicated by hilly terrain. In Figure 12, note the rise to the Capitol
Hill at the far right end of the DEM. It is evident that this particular section has to be processed in isolation.
In Figure. 14 we can observe considerable speckle misclassifications
in the output. In general there is some confusion in separating rooftop
- class data from spectrally similar classes asphalt and gravel path.
In highlighting the shortcomings of the respective analyses it has been
implicit that the problems associated with one technique can be
alleviated through the use of the other. For instance, the last point in
the discussion above leads to a significant conclusion. The emergence
of inter-class confusion in classification is not a result of" wrong"
data. The material used in construction of building rooftops is, quite
often, identical to that used in constructing roads, or laying paths.
However, the scene-classes are functionally distinct, and this
Figure 14. Extracted Rooftops from
Spectral Data
; :
are »* x MM 5
E. — Auc I aar "e ; =
A > Wu A = atte pos o e 3 ME ped
koh" NM. EL re
Gees: | JRF
ills og 18-24 . rms ko
distinction is strikingly apparent in the DEM. This conclusion is key
to the solution presented in the next section.
Procedure: Given the disparity in the two types of the data,
concurrent analysis is infeasible. Our analysis comprised maximum
likelihood classification, as discussed earlier, followed by a
thresholding operation on the elevation of all data elements identified as asphalt, gravel path or rooftop. The latter is
designed as a Boolean-type operation in which all data (identified as one of the three classes listed above) below a
certain elevation are said to be ground-level; the remaining filtered data are thus identified as building-rooftop.
Since there is a large amount of variation in scene elevation, the elevation threshold, discussed above, must be locally
determined. The following procedure was adopted towards this task.
Centroid Identification: The DEM was visually examined to identify zones or regions of relatively unchanging terrain.
Pixels representative of these zones were identified as zone centroids.
Zoning: The pixel grid was then segmented into zones identified by their respective centroid. The process involved
going through the grid and labeling each pixel according to the zone centroid closes to it. The metropolis distance
metric was used. The partitioned image
is shown in Figure 15. Zone centroids
have been highlighted as yellow dots in
the figure. Note that only pixels
identified as rooftop, asphalt or gravel
path are identified in the zoned output.
The remaining scene classes have been
absorbed into the black background.
Figure 15. Partitioned Scene with Centroid Identification
Threshold computation: For each zone, the median elevation for the pixels classified as rooftop, asphalt or gravel path
is computed. In zones with an insufficient count of rooftop pixels, it is clear that threshold will be biased towards data at
ground-elevations. The threshold for a given zone is thus chosen as the average of the median as calculated above, and
the elevation of the zone-centroid.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 599
