Environmental Engineering Reference
In-Depth Information
of the various steps starting from a sample of an Ikonos II image
of the town of Ica (Peru). The image was made available by
GeoEye to the press after the August 2007 earthquake event in
the area. The classification results in Fig. 3.4 are meaningless
from a quantitative point of view, because the image details were
degraded by GeoEye before making it public and are useful only
to understand the procedure and see how robust is. Indeed, one
might easily note that the recognition of the various elements of
the scene is improved by comparing Fig. 3.4(b) with Fig. 3.4(d).
●
Houses: houses are eroded and reconstructed as well. The SE
element is a 5
×
5 square, since we may assume that blobs
smaller then 3
×
3 m are not buildings. In fact, besides noisy
pixels, there are also cars on the streets. At the same time, blobs
bigger than 6000 pixels (2000 m
2
) are discarded, assuming
they are very large bright soil areas.
●
Shadows/roads: blob areas smaller than 15 pixels are con-
sidered noise. No assumptions about geometric features are
made.
All classes are finally merged into a unique image and eroded
by a 2
2 square element. This marker image is then used
to impose minima on the multispectral gradient image (using
a1
×
1 square SE) of the original scene to obtain the final
segmentation function. The watershed algorithm can now be
applied avoiding over-segmentation.Actually,noisecouldbe
still present, thus is preferable to remove isolated watershed
lines shorter than 40 pixels at the end of the segmentation
procedure.
Spectral and spatial features aremerged via themajority voting
step described in the previous section, and the final regularized
map is proposed in Fig. 3.5(b). A comparison of the original
and final maps shows that the ISODATA classifier has quite
good performances, even though no parameter tuning has been
done. The main drawbacks are, on the one hand, the very similar
spectral responses of buildings and bright soil, on the other hand,
the presence of a few roofs made of different materials, which
are not correctly classified as ''houses''. A very accurate spectral
classification was not the aim of this work, thus there are margins
for improvement. The final map is definitely more accurate
and precise than the ISODATA output. Besides detecting more
homogeneous and better defined areas, this process reduces some
misclassification errors, like the presence of huge bright soil areas,
and even removes ambiguities, for example when the same roof
has different cover materials.
Once the segmentation scheme has been applied to both pre-
and post-images, it was also tried to evaluate any difference in
building features. The basic assumption is that all undamaged
buildings must be found in both the images. In case of damage,
instead, post-event buildings must have a different shape and
presumably a smaller footprint than pre-event ones. These con-
siderations do not take into account all possible changes (e.g.,
the presence of new constructed buildings or bright debris areas
larger than original buildings), but they cover the most part of
cases. Damage mapping is thus shifted into the evaluation of
changes in the buildings' shape.
To this aim, common areas of the buildings are detected
by doing the difference between pre- and post-classified maps.
Then, specifically for building damage evaluation, and following
the abovementioned assumptions, a
building damage index
(BDI)
is introduced, defined as the ratio between the building area in
the difference image,
common building area
,anditsareainthe
pre-event image,
pre-building area
. By choosing an appropriate
threshold, it is possible to get an output grayscale map in which
buildings can be separated into two classes: slightly damaged
(including undamaged ones), and heavily damaged. According
to the intuitive idea that a building which is heavily damaged is
more than half destroyed, the above mentioned threshold can be
set to 0.6.
The validation of the results is made by a direct com-
parison with the visual inspection in Yamazaki, Yano and
Matsuoka (2005) and shown in Fig. 3.5(d). These validation
×
3.7
Building damage
assessment
A very interesting application of the procedure proposed in the
previous section is related to built-up object detection for build-
ing inventories, cadastral map updates and, for civil protection
purposes, damage assessment after amajor disaster. The examples
in the following paragraphs refer to earthquakes because of the
specific research interests of the authors, but the same approach
may be used for any other situation involving rapid change detec-
tion in urban areas. Moreover, two examples are shown because
the areas under test are different, both in the sense that the two
considered urban areas are in different part of the globe, but also
have very different structures and building typologies.
The first site is Bam (Iran) and the dataset is formed of a
couple of Quickbird images acquired on 30 September 2003 and
3 January 2006, respectively before and after a major earthquake
occurred in the area (26 December 2003). To exploit the high
spatial resolution of the panchromatic data and high spectral res-
olution of the multi-resolution data, two pan-sharpened images
have been generated and coregistered. For testing our algorithm,
only a limited 2000
×
2000 subsample has been used.
As stated in previous sections, the first step of the algo-
rithm is a classification of the raw image in order to extract
relevant spectral classes. In this example, the quick and simple
ISODATA algorithm was considered, looking for four different
land cover categories, labeled as ''houses,'' ''soil,'' ''trees,'' and
''shadow/roads''. The choice of this classifier has all those draw-
backs derived from using unsupervised algorithms, but basically
matches the requirement of the application. In particular, it
provides rapid output maps, even in case of unknown areas. A
sample of the resulting classification map is shown in Fig. 3.5(a),
with an intuitive color legend, with buildings highlighted in
white.
Following the analysis procedure, the four classes must be
separately processed by the morphological ''regularization'' step.
Strictly speaking, the only meaningful class for the application is
the one referring to built-up structures (''houses''). It is however
easy to understand that the improvement of the whole map yields
a segmentation function that enhances the watershed algorithm
performance.
Therefore, first of all
salt and pepper
classification noise
is reduced by filling holes in ''shadows/roads,'' ''trees,'' and
''houses'' classes. Then, specific steps for each class are applied:
Trees: noise blobs are removed after considering some geo-
metric information about trees. Basically, they have a circular
form and at least a 2 m diameter. Thus, erosion by a 1
●
×
1
disc and then reconstruction by dilation are performed.
Search WWH ::
Custom Search