The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
urban areas. In the field of SV research, on one level this opens
up new possibilities. Even though SV tends to relate to personal
and individual characteristics, physical indicators and spatial
and contextual relationships that relate to the “physical and
social world” of the individual (Clark et al., 1998) have
previously been identified (Wu et al., 2002; Rashed & Weeks,
2003) and may well be observable in RS data. In Figure 2,
indicators for SV assessment on two scales, the
individual/household and the neighbourhood scale, are
presented with their best-suited assessment methods.
Besides a generally accepted definition a more efficient and
comprehensive approach for the assessment of SV is needed.
Data availability and low efficiency have been pointed out as
two of the most important constraints in current SV assessment
(Birkmann, 2005).
4. METHODOLOGY: OBJECT-ORIENTED ANALYSIS
OF REMOTE SENSING DATA
So far, studies on vulnerability assessment using satellite data
have mainly been focusing on the assessment of physical
vulnerability (Müller et al., 2006). But by applying object-
oriented image analysis, a wide range of additional, contextual
information that is relevant for SV assessment can be extracted
from remote sensing images. Very high resolution satellite data
(0.6 m ground resolution from Quickbird satellite) have been
used in combination with various digital data sets to extract
important semantic information, e.g. the position of a building
in relation to the hazard zone. This, for example, is a critical
feature as the position of each image object in its natural and
man-made environment partly determines its SV (Rashed &
Weeks, 2003; Clark et al., 1998).
Given that SV is more of a non-tangible concept it has to be
evaluated which criteria for its assessment potentially have a
spatial expression in an image (Table 2). The selection of
indicators for this study was based on the list of generally
accepted criteria according to Cutter et al. (2003, Ebert et al.,
2007). Selected according to assessment capabilities with
remote sensing data and relevance for this study, four main
indicators (Table 2) remained. Meaningful proxy variables
(parent proxy, second column Table 2) have been found to
describe the content of the indicators. These in turn had to be
translated into proxy variables that can be directly delineated
from remote sensing data and other available information (GIS
data, city maps) using object-oriented analysis (supporting
proxy, third column Table 2).
Abundance of
lifelines
Number of
lifelines
Number of lifelines
(1)
Distance to
lifelines
Distance to
lifelines
Distance measures
13}
Table 2. Physical proxies identified for the original SV
indicators. Numbers in parentheses indicate the actual number
of proxy variables for each parent proxy.
A land use/land cover classification was performed in two
modes: with and without the usage of an additional normalized
Digital Surface Model (nDSM) that was available for a part of
the study area. In both modes, spectral as well as spatial and
contextual information have been used for the class description.
The classification accuracy for the classification with and
without the nDSM was 89.1 % and 84.3 %, respectively. Based
on the classification results, and by using additional GIS data
and city maps, the supporting proxy variables were delineated.
The proportion of built-up and vegetated area was calculated for
each neighbourhood, using spatial metrics (Herold et al., 2003).
Similarly, the percentage of paved roads compared to the total
road network per neighborhood was calculated. Seven different
roof types and their neighbourhood proportion were also
classified from the Quickbird image, which can give
information about the socioeconomic status of the residents.
However, field data would be necessary to associate actual roof
materials with the roof types classified, as Quickbird’s spectral
characteristics are insufficient for such determination. The
available service infrastructure was digitized from city maps
and quantified for each neighbourhood. Image texture measures
were used to describe the distribution of grey values in an image
and thus to characterize the heterogenic structures of a city
(Tuceryan & Jain, 1998). To describe the topographic location,
the slope position was calculated from a digital terrain model
(DTM, 1.5 m resolution), to which each image object was then
associated. Twelve slope classes in 5° intervals were delineated.
All buildings in a hazard zone (flood and landslide) were further
masked out, and the percentage compared to all buildings in the
neighbourhood calculated. Building heights were further
delineated from the lidar data set for a smaller part of the study
area, and used to characterize commercial development. Finally,
three possible distances to lifelines (less than 100 m, 100 to 250
m, more than 250 m) were defined that refer to the position of
each segment in the image. A total of 47 supporting proxy
variables were thus delineated (Table 2).
Original
indicator
Socio-economic
status
Commercial and
industrial
development
Parent proxy
Supporting proxy
Settlement
type
Topographic
location
Commercial
development
Proportion of built-
up and vegetated
area (4 proxies);
Road conditions (1);
Roof type (7);
Texture (10)
Slope position (12);
Proportion of
buildings in hazard
zone (2)
Building heights (7)
5. RESULTS
The goal of the evaluation was to investigate if the proxy
variables can be used to extract information from the satellite
data that are relevant for the assessment of SV. We thus tested
for a relation between the remote sensing derivatives (proxy
variables) and a reference SV index (SVI, modified from Haki
et al., 2004) based on census data from 2000 census in
Tegucigalpa. The set consist of seven variables (gender, literacy,
wall material, roof material, water availability, waste disposal,
building use) to calculate the SV for each neighbourhood. The
resulting values for each neighbourhood are shown in Figure 3.
We used a stepwise regression model to determine which of the
47 proxy variables (explanatory variables) can best explain the
variation of the vulnerability score V (dependent variable) (Jain,
2005). The main observations were the following: