^ T
zT
x OF *ÍDM O
2
Za $1DM, 45
7006
2 ii *iDM 90
ZONE
i 8IDM, 135
00:
— s [DM meat
ü
#IDM 6
SIM af
*IDM 9
*IDOM 138
5 XM, menn.
FS sta 9:0 E» wiz ai 24 ; 925 S26 927
Destroved buildings
Figure 4. Randomly selected objects present the variation of
textural features for intact (top) and destroyed (bottom)
buildings.
Thus, using maximum and minimum of the angular textural
features, we can exploit the presence of the textural orientation
for building condition identification. The use of the average
values of the angular features reduces this phenomenon.
The selection of features is based on the analysis of the
frequency distribution histograms (Figure 5). Consequently, the
maximum value of the angular features for homogeneity
(IDM max) is taken for the classification together with the
DPC.
pppoe
C BE 189
as eo
DP vaine
Figure 5. Frequency distribution histograms of the selected
features. Red values correspond to destroyed buildings,
blue values to intact buildings.
3.4 Classification
After selection of the most relevant features the building
condition is calculated as a binary classification. There are
many classification techniques implemented in the Open Source
data mining software ‘Orange’. However, we are of the opinion
that an optimally selected feature set is an indispensable
condition for a successful classification. Unlike supervised
techniques, the utilization of unsupervised classification does
not require a suitable training data set which is difficult to adapt
for all types of constructions. Reasoning from this assumption
we use an unsupervised k-means clustering with the Euclidean
distance in our study. The main rational is the simplicity of its
implementation and its performance and applicability even on
large data sets.
For the clustering, the centres of cluster are initially chosen
arbitrarily. Then every point is assigned to the cluster according
to a similarity measurement, for example, distance. The centres
are recomputed as centres of mass of their assigned points. The
algorithm comes to an end, when there are no changes by the
next iteration or when the number of changes is below a given
threshold.
4. RESULTS AND DISCUSSION
The developed change detection algorithm is evaluated using
remotely sensed image data acquired after the Yushu earthquake
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
on April 2010 and a related vector map containing 610 objects
corresponding to the pre-event urban situation. After the feature
extraction process and the analysis of the calculated features the
building conditions are calculated by an unsupervised k-means
clustering. The following confusion matrix shows the result of
the experiment.
Prediction
CC
Cl | 175 93 268
C2 | 40 302 | 342
215 | 395 | 610
Correct
class
Table 2. Confusion matrix. C1 — intact buildings,
C2 —destroyed buildings.
The classification accuracy of 78% demonstrates the ability of
this method to detect destroyed buildings. The independent
features that depict the different types of the information are
prerequisite to the sufficient result of the analysis. Furthermore,
the application of the maximum value of the angular textural
features instead of the generally used average values leads to a
significant improvement. Considering the intact and damaged
building separately, we observe that average values decreases
feature performance and consequently the change detection
accuracy.
Detailed analysis of the results shows that the main
classification errors are due to the image quality and
rectification errors; for example, shadow from a building
covering another one, not well defined destruction that cannot
be recognized as a heterogeneity, and incorrect position of
vector objects.
The object-oriented GIS technology makes it possible to
concentrate on the investigation of specified objects, thereby
reducing the false alarms which are due to natural changes in
the environment and not to buildings destruction. A subset of
the generated damage map is shown in Figure 6.
Figure 6. Change detection result by the proposed method. Red
polygons indicate destroyed building, blue polygons intact
buildings.
5. SOFTWARE ENVIRONMENT
The methodology is implemented using Open Source software
components. The processing of vector and raster data sets
including vector data selection and conversion, and data
visualization is performed in the GIS GRASS (Geographic
Resources Analysis Support System) environment (Neteler &
Mitasova, 2004). It is currently the most popular system among
