Professional Documents
Culture Documents
Ebert 2008
Ebert 2008
DOI 10.1007/s11069-008-9264-0
ORIGINAL PAPER
Received: 16 November 2007 / Accepted: 6 June 2008 / Published online: 1 July 2008
Springer Science+Business Media B.V. 2008
A. Ebert (&)
UFZ, Helmholtz Centre for Environmental Research—UFZ, Permoserstr. 15, 04318 Leipzig, Germany
e-mail: [email protected]
N. Kerle A. Stein
International Institute of Geo-information Science and Earth Observation (ITC), Hengelosestraat 99,
P.O. Box 6, 7500 AA Enschede, The Netherlands
e-mail: [email protected]
A. Stein
e-mail: [email protected]
123
276 Nat Hazards (2009) 48:275–294
Abbreviations
DSM Digital surface model
nDSM Normalized digital surface model
DTM Digital terrain model
GIS Geoinformation System
OOA Object-oriented analysis
PV Physical vulnerability
SV Social vulnerability
SVI Social Vulnerability Index
TTA Test and training areas
1 Introduction
In recent decades the number of natural disasters has been increasing, affecting a growing
number of people by causing extensive loss of life and property damage. Every populated
place faces a certain risk to be affected by a disaster, the size of which depends on location-
specific (i) hazards present, (ii) vulnerability and (iii) the number of elements at risk
(Fig. 1). This relation can be expressed as:
Risk ¼ f ðhazard; vulnerability; elements at riskÞ ð1Þ
Disaster management tools are available to help minimize the risk and thereby the
impact of a hazardous event (Fig. 1) but require detailed knowledge about the risk a
particular area is facing.
Risk analysis encompasses the assessment of all factors shown in Eq. 1. The hazard, i.e.
the probability of a potentially damaging event with a certain magnitude to occur (Cardona
2003), can be expressed in absolute values, as can the number of elements at risk. Methods
to evaluate these physical concepts have been shown in previous studies (Bacon et al.
Fig. 1 Disaster management cycle with the position of social vulnerability assessment
123
Nat Hazards (2009) 48:275–294 277
1997; Giardino et al. 2005; van Westen et al. 2005). Vulnerability, on the other hand, is
more multi-faceted. In addition to the social and physical sides, also environmental and
economic aspects can be considered, resulting in a variety of research perspectives and
term definitions. Rashed and Weeks (2003a) see vulnerability as an ill-posed problem
defined by multiple solutions and uncertainty about the concepts, rules and principles
involved to reach these solutions. Physical vulnerability (PV) refers to the properties of
physical structures that determine their potential damage in case of a disaster (e.g. material
type and construction quality). Theoretical rules and considerations exist for its assessment,
although a comprehensive assessment is challenging as a detailed database is required
(O’Hare and Rivas 2005). Due to the sound conceptual basis, given adequate data avail-
ability, PV is relatively straightforwardly assessed and frequently used synonymously with
vulnerability in general. This results in implicit neglection of SV, which refers to the socio-
economic circumstances and individual characteristics that make people susceptible to the
impact of a hazardous event (Cutter et al. 2003). In this study the approach by Clark et al.
(1998) is applied, defining SV as ‘‘people’s differential incapacity to deal with hazards,
based on the position of the groups and individuals within both the physical and social
worlds’’, which has to be assessed with respect to the particular hazard or combination
thereof (e.g. floods and/or landslides; Coburn et al. 1994). Rashed and Weeks (2003b)
further discussed that vulnerability can have different sources. It can either be inherent, e.g.
due to the affiliation to a certain marginalized group (persistent vulnerability), or result
from a choice, e.g. people choosing to live in a hazard-prone environment (situational
vulnerability). This can also be seen as the difference between ‘who you are’ and ‘where
you are’, respectively (Rashed and Weeks 2003a).
Emdad Haque and Etkin (2007) point out the significance of societal dimensions in
hazard analysis, and thereby in risk analysis. At the moment, SV still lacks a broadly
accepted definition, though several examples for its assessment exist (Briguglio 2003; Haki
et al. 2004), based on indices derived from the analysis of census data (Cutter et al. 2003;
Dwyer et al. 2004) or from data collected using community-based methods (Flint and
Luloff 2005; Allen 2006). Census data, however, are collected for a different purpose and
consequently neglect important information about hazard perception and mitigation abil-
ities. Moreover, they are, if at all, only available on neighbourhood level or coarser, and
with a temporal resolution of no more than 5–10 years. Community-based surveys on the
other hand are very detailed but also time-consuming and lead to results that can be
subjective and are difficult to up-scale (Birkmann 2005; Villagrán de León 2006). Both
traditional approaches used for SV assessment thus can provide important information yet
have limited efficiency and/or transferability. Additionally, the generally low temporal
resolution of censuses and of community-based surveys that go beyond very local multi-
temporal studies are poorly suited to capture the dynamic character of SV in an operational
manner.
This work builds on recent studies that have identified physical expressions of SV.
For example, Wu et al. (2002) analysed housing structures and the built environment in a
GIS-based study of SV, similar to Clark et al. (1998) who also studied the link to land
use and transportation infrastructure and Rashed and Weeks (2003b) who considered the
physical and social conditions to be intricately linked, making the former indicative of
the latter.
The objective of this work was to test the utility of lidar, optical satellite and GIS data to
derive SV-relevant information by using physical proxy variables to describe not-directly
observable phenomena, with better time and cost efficiency and higher temporal resolution
compared to the traditional analysis methods. Proxies are measurable variables that can
123
278 Nat Hazards (2009) 48:275–294
provide insight into phenomena that cannot be directly observed or measured, but which
are conceptually linked.
Physical expressions of social vulnerability (SV), such as settlement type or vegetation
density, can be best expressed in terms of spectral, geometric and topological character-
istics of land cover elements forming complex urban units. Therefore, the proxies in our
study were addressed in a framework of contextual object-oriented analysis (OOA), which
to our knowledge has not been employed before in SV research. Additionally, the focus is
laid on Honduras, a less developed country where spatial and physical factors of SV are
even more pronounced (Rashed and Weeks 2003b), where current census data as used by
Cutter et al. (2003), Clark et al. (1998) and Wu et al. (2002) are often not available, and
where disaster casualty figures show vulnerability to be the highest.
2 Study area
The test area for this study contained 87 neighbourhoods (in total ca. 3 9 3 km) in Teg-
ucigalpa, the capital city of Honduras. Tegucigalpa is located at about 14 N and 87 W in
the highlands of central Honduras at an elevation of approximately 1,000 m asl (Fig. 2).
With 53.0% of the population living below the national poverty line (UNDP 2005),
Honduras is a country with medium development (UNDP 2005) and limited industrial and
commercial infrastructure. Tegucigalpa today has more than one million inhabitants and is
growing steadily (on average about 2.8% p.a. from 1988 to 2001) in an unplanned manner,
also into hazard-prone terrain, such as steep slopes surrounding the city and along the
rivers, often the only available and affordable spaces for building construction (Angel et al.
2004). The proximity to the labour market and urban facilities is for most migrants more
important than living in a safe environment, a known worldwide phenomenon (O’Hare and
Rivas 2005).
Ecological and land use changes, e.g. by deforestation of slopes, straightening of rivers
and loss of natural flooding areas, also result from city growth, which in turn increase the
123
Nat Hazards (2009) 48:275–294 279
city’s vulnerability (Davidson 2006). This became apparent during Hurricane Mitch that
struck Tegucigalpa in 1998 and triggered numerous landslides on deforested slopes and
severe flooding, which caused thousands of fatalities and destroyed large parts of the city’s
infrastructure, that were located in hazard-prone areas (CINDI 1998).
The utility of geoinformatics tools for all aspects of disaster management has been amply
illustrated in the literature. In particular remote sensing technologies have been used for (i)
predisaster applications, such as scenario modelling (Iverson et al. 1998; Tralli et al.
2005), (ii) forecasting of events (e.g. of volcanic eruptions or windstorms), (iii) early
warning and event monitoring (Ramsey and Flynn 2004) and (iv) damage assessment and
monitoring of land use changes after a disaster (Kerle 2002; Arciniegas et al. 2007).
Progress is also being made in the use of remote sensing data in quantitative risk
researches, e.g. for hazard analysis (Lee et al. 2004), assessment of urban vulnerability
including SV (Rashed and Weeks 2003a), or vulnerability of buildings (Müller et al.
2006). So far, most studies are limited to the description of one or few components of risk
(Eq. 1) or provide a comprehensive assessment of only the factors that can be directly
measured (van Westen et al. 2005).
The main limitations in the application of geoinformatics in comprehensive risk man-
agement are (i) the high data demand and cost, (ii) the need for an integrated analysis of
multi-type/format data, (iii) the need for frequent risk assessments and database updating
due to rapid urbanization and (iv) that information also about concepts that are difficult to
map directly, such as SV, has to be included. However, the field is also benefiting from
developments in other remote sensing areas, such as automatic mapping and classification
of buildings with InSAR (Balz and Haala 2003; Stilla et al. 2003), laser data (Dash et al.
2004), or Ikonos imagery (Fraser et al. 2002). Rashed and Weeks (2003b) focused the
assessment of vulnerability of urban places (urban vulnerability) using pixel-based spatial
metrics and a spectral unmixing approach. Our study focuses on the assessment of SV
based on multi-source remote sensing and GIS data, using contextual information and
multi-scale interpretation implemented in OOA. Contextual information provides knowl-
edge about distances (e.g. to hazard zones) and the local environment (e.g. slope) of a
building, while multi-scale analysis considers different spatial dimensions, ranging from
sub-building features to entire city districts. The resulting delineation and application of
proxy variables, as explained in Sect. 4, support the identification of parameters that are
non-physical and not directly visible, hence which cannot be assessed from satellite data
using alternative methodologies, such as pixel-based analysis.
4 Methodology
Image classification has traditionally been done using pixel-based analysis, where each pixel
is classified based on its spectral characteristics and without contextual information. More
recently developed OOA methods aim at imitating human cognition and begin with seg-
mentation of image data on different spatial scales that depend on the desired level of
generalization, the spatial resolution of the image, as well as the inherent scale of the objects.
123
280 Nat Hazards (2009) 48:275–294
For example, chimneys can be preserved as individual objects in one level, while in another
level entire buildings constitute the smallest spatial unit. The result of the segmentation
process is an image hierarchy that contains different levels with objects of different sizes,
where each object knows all properties of its sub-, super- and neighbour-objects (Baatz et al.
2004). All of those, together with the objects’ geometric and spectral characteristics, can
then be incorporated in the interpretation of the resulting image segments at only one or at
several segmentation levels (Blaschke and Strobl 2001; Baatz et al. 2004). Small spectral
inhomogeneities that normally lead to wrongly classified pixels are averaged out during the
segmentation, allowing increased classification accuracies. However, previous studies also
showed that segmentation alone does not improve the classification result, but rather that a
proper integration of semantic information is needed in the post-segmentation analysis (Gao
et al. 2007). In the context of risk, this is critical as the position and spatial arrangement of
image objects in their natural and man-made environment strongly determine the SV (Clark
et al. 1998; Wu et al. 2002; Rashed and Weeks 2003b).
Object properties in this study comprise spectral and textural information, as well as
shape characteristics, object size and distance to all other image objects, which were
analysed in Definiens. Beneficial is that both image data and auxiliary raster and vector
data types, such as elevation models, thematic layers or other GIS data, can be included in
image segmentation and classification. This allows class descriptions that support a
detailed land cover and land use assessment by providing a rich contextual analysis
framework. Specific classes, such as buildings within a hazard zone or buildings on steep
slopes, can be identified directly (Sect. 4.6). Also GIS data can be directly used as a basis
for class descriptions, as was done here with polygon outlines of a cemetery and rivers. The
combination of spectral features with shape characteristics and auxiliary data yields classes
that are suitable for a comprehensive assessment of living environments, and which in turn
are indicators for the assessment of SV. These indicators are then expressed using object-
based proxy variables that have previously been used in PV assessment, e.g. building
heights for flood vulnerability, but to our knowledge not in a SV context. Figure 3 presents
the methodology of this study, while Table 1 gives an overview of the data used.
Unlike the DTM, the normalized DSM (nDSM, Sect. 4.2) that contains absolute heights
of all objects on the surface was only available for a small part of the study area. Thus
123
Nat Hazards (2009) 48:275–294 281
the image analysis was performed both with and without the nDSM (Project 1 and
Project 2, respectively), with image segmentation in both cases also incorporating the
pansharpened Quickbird image and the digitized outlines of the cemetery and the rivers.
Different segmentation parameters were applied to create three image levels for the
classification of the Quickbird image only and four for the classification of the Quickbird
image in combination with the nDSM (Table 2). The additional level here was used to
distinguish flat and built-up areas based on data from the nDSM. The segmentation
parameters were found based on trial-and-error and evaluated by visual analysis. Even
though Project 2 was just a subset of the first, new segmentation parameters had to be
found to ensure suitable segment borders also reflecting the additional nDSM data, such
as building outlines. The lack of a methodology for finding objective parameters and
evaluation tools other than visual analysis is a known limitation of Definiens, or indeed
segmentation-based analysis in general, although several groups are working on ways to
overcome this limitation (e.g. Espindola et al. 2006; Zhang and Maxwell 2006).
123
282 Nat Hazards (2009) 48:275–294
Table 1 Raster and vector data used for this study in Tegucigalpa
Type Source Date Format Resolution
Table 2 Segmentation parameters and extracted land use/land cover classes for the analysis of the pan-
sharpened Quickbird image only (upper part) and in combination with the nDSM (lower part)
IL Classes SP Col. Comp.
Twelve main land use/land cover classes were initially classified from the data set (Fig. 3).
For the assessment of SV, the most relevant classes are built-up areas (buildings) and roads
as the principal indicator for habitation, and vegetation for a characterization of the
neighbourhood (Ridd 1995; Rashed and Weeks 2003b). Class descriptions were done using
representative samples from the study area (test and training areas, TTA) and by creating
123
Nat Hazards (2009) 48:275–294 283
knowledge-based fuzzy sets. Fuzzy sets refer to the object properties (features), such as
spectral values, shape characteristics, relations to other image objects, or information from
thematic layers (such as elevation data). They determine a range of values for a certain
feature that is characteristic for the respective class and are expressed mathematically in
membership functions (Baatz et al. 2004). A main aspect of Project 1 was the distinction
of different roof materials as a principal descriptor of housing types, based on spectral
information, while Project 2 yielded a classification of buildings based on their height.
Seven different roof types were distinguished in the study area, although additional ground
truth information or better multi-/hyperspectral data would be necessary to determine the
actual material of the roof types. Since neither was available for this study, it was only
tested if roof material in principle can serve as a proxy for SV (see Sect. 4.6). Different
vegetation types were distinguished using NDVI values, texture measures and visual
interpretation (Myint 2007). Finally, two main road qualities (paved and barren), critical
indicators of urban living conditions, were identified based on spectral properties. The
overall classification accuracy, calculated using independent TTAs, was 84.3% for the
image without the nDSM, and 89.1% for the image with the nDSM. Thus, elevation
information is a valuable contribution for the extraction of building footprints and for
certain class descriptions, e.g. for the separation between paved roads and spectrally
similar roof types.
Image texture describes the distribution of grey values in an image and thus characterizes
the homogeneity of settlements, and can be calculated in different ways, e.g. using mea-
sures such as homogeneity, variance and skewness (Tuceryan and Jain 1998; Herold et al.
2003). The homogeneity of the Quickbird image was calculated using a Grey Level Co-
occurrence Matrix (GLCM) in Definiens, which is based on the frequency of specific pixel
combinations in the image (Baatz et al. 2004). Variance and skewness were calculated in
Erdas Imagine for the Resource-Sat image (see Table 3).
Proxy variables can be statistical census data, grey values of single pixels (e.g. Lawrence
et al. 2002), or data related to diverse object characteristics. They are frequently used in
disaster research where the parameter of interest cannot be directly assessed. Cutter et al.
(2003) identified a total of 17 measures, such as age, gender or socio-economic status to
characterize SV. Similarly, Rashed and Weeks (2003b) used vegetation ratios as a proxy
for wealth of a neighbourhood, while Wu et al. (2002) considered land use to be a general
proxy of a place’s exposure to flood hazard. Given our focus on image-derived informa-
tion, the following indicators were found to be relevant for SV assessment and are expected
to have physical expressions:
• Socio-economic status
• Commercial and industrial development
• Service infrastructure/lifelines, and distance to those.
Table 3 shows how the original indicators for SV assessment were translated into
proxy variables that can be primarily delineated from remote sensing data. It was
found that the socio-economic status of a household in Tegucigalpa can best be
123
284 Nat Hazards (2009) 48:275–294
Table 3 Criteria for the assessment of SV with spatial expression and explanation of the delineated proxy
variable
Original indicator Parent proxy Supporting proxy
described via the settlement type where it is located, and by its topographic location.
Areas with only little vegetation and poor road quality, little economic and infra-
structural development and/or situated in hazard-prone areas such as steep slopes were
considered to have the lowest coping capacities and thus to be most socially vulner-
able. Tegucigalpa, similar to other Central American capital cities, also has more
luxurious neighbourhoods located in areas with steeper slopes, for reasons more related
to the view, seclusion or better environmental characteristics such as air quality. This
generally results in higher situational and lower persistent vulnerability. Such areas can
still be identified based on spatial metrics related to building patterns and surfaced
road and vegetation ratios.
To describe the settlement type, the proportion of built-up and vegetated area was
calculated for each neighbourhood of the city (see outlines in Fig. 4). Proportion measures
are spatial metrics originating from ecology that characterize and compare settlement
areas, allowing spatial variations and temporal changes in the urban morphology to be
quantified (Herold et al. 2003; Rashed and Weeks 2003b; Herold and Clarke 2007). This
study was devised as a monotemporal assessment; thus, the calculation of spatial metrics
focused on proportion measures to characterize neighbourhoods. For example, roads were
123
Nat Hazards (2009) 48:275–294 285
Fig. 4 Scores from the social vulnerability index per neighborhood (bottom) with the pansharpened
Quickbird image as reference (top)
separated into paved and barren roads, yielding the fraction of paved roads within the total
road network in each neighbourhood. A high proportion of paved roads corresponds to a
high development of the neighbourhood.
123
286 Nat Hazards (2009) 48:275–294
Building types and state have been identified previously as physical expressions of SV
(Cutter et al. 2003; Rashed and Weeks 2003b). Hence in imagery, largely restricted to a
vertical perspective, the roof construction material may be a useful proxy for the socio-
economic status of the building occupants. Due to the lack of ground data on actual roof
materials, it was tested whether different materials that can be distinguished in the satellite
data have a statistical potential to explain SV variation. Seven different roof types were
classified from the Quickbird image, and the proportion of each roof type calculated
compared to the total area covered by buildings.
The available service infrastructures (commercial centres, transport infrastructure,
institutional or governmental buildings, gas stations, universities) were digitized from
existing city maps (Table 1) and quantified for each neighbourhood as another indicator for
development. In case such infrastructure information is not available for a given study area,
visual image analysis or tools such as Google Earth could be used to identify large
commercial areas or transport hubs such as bus or train stations.
To describe the topographic location, the slope position was calculated from the DTM,
to which each image object was then associated. Using membership functions, only the
built-up areas were masked and combined with the slope information. Twelve slope classes
in 5 intervals, from 0 to 60, were delineated. As another supporting proxy variable, all
buildings in a flood or landslide hazard zone (delineation based on existing USGS hazard
maps prepared in 2002) were masked out, and the percentage compared to all buildings in
the neighbourhood calculated.
Building heights were delineated from the lidar data set that covers only a part of the
study area (Table 2). Using this supporting proxy variable, the commercial development
was characterized. In Tegucigalpa, high buildings ([7 stories) can to a large extent be
associated with commercial or industrial development, which in turn is regarded as an
indicator for low SV values.
Lastly, three possible distances to service infrastructure and lifelines (\100, 100–250,
[250 m) were defined using membership functions, relating the position of each building
to those features.
A total of 47 supporting proxy variables (Table 3) were delineated from the digital data
sets to describe the three original indicators for SV (first column in Table 3). While
delineating proxy variables, information about the environment was derived from different
data sources as shown in Table 1. By using membership functions and fuzzy rule sets, each
classified image object was characterized according to its (i) slope, (ii) distance to hazard
zone and (iii) distance to service infrastructure. In addition, spatial metrics (Herold et al.
2003; Rashed and Weeks 2003b) were calculated from the classification-based land use/
land cover information to describe the settlement characteristics for each administrative
neighbourhood.
After the extraction of the 47 proxies, their relevance as predictors of SV was tested. In a
previous study, SV was calculated based on spatial metrics and compared with values
derived from multi-criteria GIS analysis (Rashed and Weeks 2003b). Here, using a
traditional method for comparison, a reference SV map on a neighbourhood level in
Tegucigalpa based on the 2000 census data and the method used by Cutter et al. (2003)
123
Nat Hazards (2009) 48:275–294 287
was calculated. Seven main variables were available: (i) gender, (ii) literacy, (iii) roof
material, (iv) wall material, (v) water availability, (vi) waste disposal and (vii) building
type.
These variables differ from those employed in image analysis. This complicates a direct
comparison of the results but also reflects the reality of the different methods with which
SV is being assessed. Rashed and Weeks (2003b) faced a similar problem when they used
an index of wealth (based on census data) and an index of vulnerability (based on HAZUS
seismic risk simulations) to assess the explanatory potential of image-based SV proxies.
The alternatives for each variable (e.g. male and female for gender) were ranked
according to their impact on SV based on previous research (Ranganath 2000; Cutter et al.
2003; Haki et al. 2004) and field observations (Angel et al. 2004). Pairwise comparison
(Saaty 1980) was used to assign weights to all alternatives based on their ranking. A Social
Vulnerability Index (SVI) adapted from Haki et al. (2004) was then applied to calculate the
SV per neighbourhood based on the census data:
X
m
SV ¼ vi qi ð2Þ
i¼1
where SV is the vulnerability score for each neighbourhood, vi is the weight derived from
the pairwise comparison for each variable (values ranging from 0 to 1), and qi is the
relative frequency of the variable per neighbourhood (values ranging from 0 to 1). Cal-
culated values for each neighbourhood are shown in Fig. 4 (bottom). A similar additive
calculation was also used by Cutter et al. (2003), whereas Wu et al. (2002) normalized SV
values to range between 0 and 1. This presents a conceptual challenge for quantitative risk
assessment. While physical or economic vulnerability characterize a clearly quantifiable
degree of loss, a SV concept that links to a person’s ability to anticipate, cope with, resist to
and recover from the impact of a hazardous event (Wisner et al. 2004) faces challenges
such as scaling. Indeed, Adger and Kelly (1999) and Smit and Pilifosova (2003) focus on
understanding functional relationships rather that calculating absolute SV values. For
Rashed and Weeks (2003a), an assessment of SV in absolute terms is not possible, as it is
continuously modified and varies over space and time. Hence, our principal aim is to
identify variables that help explain how SV is generated, and it is concurred with Clark
et al. (1998) for whom the main purpose of vulnerability maps is the ability to identify
threatened populations and areas on which to focus limited resources.
A two-step procedure to identify the significant variables to explain SV was used. First, a
stepwise regression model was used, based on ordinary least squares, to select those
variables out of the set of 47 proxy variables that significantly contribute to explaining the
vulnerability score SV from the SVI (Draper and Smith 1981; Rashed and Weeks 2003b;
Jain 2005). In a stepwise regression, the variables are entered and/or removed successively
and thus accepted for inclusion if a tolerance threshold pIN is exceeded for entrance and a
pEX is no longer exceeded for exclusion. After testing various values for pIN, it was decided
to use a pIN-value of 0.15. These values were large enough to allow also entering of less
strongly related explanatory variables into the regression model, hence remaining on the
safe side. A somewhat lower pEX value of 0.1 ensured that no endless iterations occurred.
The second step used spatial regression to estimate values of the coefficients on
the comprehensive set of explanatory variables thus obtained (Cressie 1991). The
123
288 Nat Hazards (2009) 48:275–294
6 Discussion
The frequent lack of suitable data, as well as conceptual difficulties, explain why many
studies do not incorporate SV assessment as a critical aspect of a comprehensive risk
assessment. Existing methods either use data with limited suitability and availability
(census data) or rely on detailed house-to-house surveys that are insightful but preclude
wider and more frequent use. The goal of this study was to test the utility of indicators of
123
Table 4 Results from the stepwise regression model for the explanation of the vulnerability scores VPC (pos., positive effect; neg., negative effect)
Step Variable entered Variable removed No. of Partial R2 Model R2 F-value Pr [ F (P-value) Estimated
variables parameter
123
290 Nat Hazards (2009) 48:275–294
SV from high resolution images and other spatial data using OOA. Proxy variables, i.e.
variables that translate and express the not-directly observable concept of SV, had to be
defined and identified, and their explanatory capacity had to be assessed. The 47 proxy
variables delineated were tested in a stepwise regression analysis and could explain almost
60% of the variation of a reference SVI calculated from census data with conventional
methods (Wu et al. 2002; Cutter et al. 2003).
The initial land cover/use classification of the pansharpened Quickbird image using
OOA is a critical basis for the delineation of spatial metrics and the advanced land use
classes that incorporate GIS data (e.g. hazard zones) and the DTM (Table 3). Based on this
information, contextual analysis allowed suitable proxy variables to be defined that
describe non-physical indicators of SV. In this study the high spatial resolution was more
valuable than high spectral resolution, as many relevant details, such as single houses,
could be identified. The incorporation of lidar data was very useful to extract building
heights. Since such data are rather cost-intense, they might not be always available.
Nevertheless, building heights can also be extracted from stereo-images or orthophotos
(Fraser et al. 2002). A higher spectral resolution might have allowed a more detailed
assessment of construction and roof materials. Indeed, Rashed and Weeks (2003b) based
their land cover assessment for SV analysis on Landsat TM data.
The proxy variables defined described the non-physical concept of SV to a substantial
degree, though it is clear that SV assessment also comprises a range of indicators that have
no physical expression and thus cannot be delineated from image data, e.g. gender, age,
knowledge about the hazard and individual disaster preparedness (Cutter et al. 2003).
Although no comprehensive assessment of SV is feasible with image data alone, it can help
overcome the low spatial detail of census-based SV assessment. In Cutter et al.’s approach,
the spatial resolution is limited to the 3,141 US counties (with approximately 100,000
people in one census tract). Rashed and Weeks (2003b), who also used US census data,
also concluded that both the social and the physical aspects of SV need to be understood
and that thus census data alone are not sufficient. Finally, it has to be stressed that also SV
is hazard-dependent. For example, while meteorological hazards such as heatwaves
challenge an individual’s health coping capacities (Nakai et al. 1999), floods pose a more
immediate physical threat to younger and shorter people. In the SV analysis, weights have
to be assigned that correspond to such individual characteristics, leading to particular
challenges when multi-hazard vulnerability is to be assessed.
The method presented here can also help overcome high cost and limited spatial cov-
erage of ground surveys. Figure 5 illustrates the cost-benefit concept of an integrated
approach. The use of satellite data (solid lines) is efficient compared to house-to-house
surveys, comparatively easy to repeat and relatively low in cost per mapping unit, but by
itself not detailed enough for a comprehensive assessment of SV. Most detailed but also
most expensive and time-consuming are house-to-house surveys, while census data are
most efficient but least detailed (both in dashed line). In general, the higher the level of
detail, the higher the costs and the higher the time input required for data processing. The
gap between the dashed and the solid line shows the trade-off between costs, efficiency and
the level of detail. The level of detail is a relative measure compared to the alternative
method. The width of the cost-benefit area is mainly dependent on the accepted trade-off at
the considered scale. In general, the higher the level of detail required, the higher the costs
and the lower the efficiency.
Additionally, image data are available more frequently than census or ground infor-
mation and the methodology is generic, thus can be applied in other areas if adapted to the
data availability and scale of the area of interest, while also allowing the dynamics of SV to
123
Nat Hazards (2009) 48:275–294 291
Fig. 5 Cost-benefit area of an integrated approach for the assessment of social vulnerability implementing
both traditional field surveys and remote sensing and GIS technology
be captured. For the definition of proxy variables, local knowledge about the study area is
inevitable.
Although the proxies are generic to a large extent, they have to be adapted and eval-
uated to suit the specific situation. Most important is that the key factors that drive SV in
the specific region are recognized and translated into appropriate variables. However,
exactly such adaptation is also needed for alternative SV assessment methods. Census-
based approaches have to be modified depending on the variables available, while com-
munity-based work, be it based on focus groups or questionnaires, has to adapt to the local
setting, as well as hazards present.
It is argued that the optimal approach for SV assessment is a combination of both
traditional methods (analysis of census data and community-based surveys) and new geo-
informatics-based methodologies (OOA of high-resolution satellite and other spatial data).
The potential of OOA for the delineation of proxy variables has also yet to be fully
realized. The combination of different data sources and object information offers a variety
of possibilities to describe single segments or a group of adjacent segments (e.g. neigh-
bourhoods). The better the understanding of SV, the better the resulting proxy variables.
That means that the results and assessment methods based on remote sensing data can be
improved if more extensive input is given from social scientists that work on a definition
and indicators of SV. The increasingly rich conceptual basis for vulnerability assessment
emerging from previously disparate research fields (Eakin and Luers 2006) is particular
encouraging. Similarly, the better the relation between SV indicators and their physical
expression in urban morphology is understood, the more suitable and better adapted to non-
visible concepts the description of proxy variables can be.
Even if only data with medium spatial resolution are available, some relevant indicators,
such as settlement extent and texture measures for the characterization of the settlement,
can be delineated. Recent studies on satellite data interchangeability for urban risk
assessment showed that also medium resolution data, e.g. from Landsat TM, can be used to
delineate land use classes such as built-up areas and vegetation to a sufficient degree
(Rashed and Weeks 2003b; Shamaoma et al. 2006). Those were also the classes that
proved to be important for this study, and thus while high spatial resolution improves the
accuracy, medium resolution data can also be applied. However, this emphasizes also that
more than one method exists to calculate SV.
123
292 Nat Hazards (2009) 48:275–294
7 Conclusions
In this article it could be shown that OOA comprising the evaluation of contextual
information and the analysis of image segments on several scales is a valuable tool for the
delineation of a variety of proxy variables that can describe non-physical indicators of SV
that are not directly mappable.
These proxy variables were derived from air- and spaceborne imagery as well as GIS
data and successfully used to describe those indicators of SV that relate to the socio-
economic status of a household and to settlement characteristics of a neighbourhood. It can
be argued that, when used in conjunction with limited ground-based data and improved
spatial extrapolation techniques, it constitutes an optimized and more economical
approach. Thus, the methodology presented here can be considered as a significant con-
tribution to disaster management, especially for rapidly changing but data scarce areas.
Acknowledgements We would like to thank Micah Perlin from Princeton University for providing GIS-
Data, Guillermo Suazo from PMDN for providing the DTM and census data and Kunda MM Rao from
NRSA for making available the Indian ResourceSat P-6 image.
References
Adger WN, Kelly PM (1999) Social vulnerability to climate change and the architecture of entitlements.
Mitigation Adapt Strat Glob Change 4(3–4):253–266
Angel S, Bartley K, Derr M, Malur A, Mejia J, Nuka P, Perlin M, Sahai S, Torrens M, Vargas M (2004)
Rapid urbanization in Tegucigalpa, Honduras. Preparing for the doubling of the city’s population in the
next twenty-five years. Report from Woodrow Wilson School of Public and International Affairs,
Princeton University, 74 pp
Allen KM (2006) Community-based disaster preparedness and climate adaptation: local capacity-building in
the Philippines. Disasters 30:81–101. doi:10.1111/j.1467-9523.2006.00308.x
Arciniegas G, Bijker W, Kerle N, Tolpekin V (2007) Coherence- and amplitude-based analysis of seism-
ogenic damage in Bam, Iran, using Envisat ASAR data. IEEE Trans Geosci Remote Sens 45:1571–
1581. doi:10.1109/TGRS.2006.883149
Baatz M, Benz U, Dehghani S, Heynen M, Höltje A, Hofmann P, Lingenfelder I, Mimler M, Sohlbach M,
Weber M, Willhauck G (2004) eCognition user guide 4. Munich
Bacon C, Mastin L, Scott K, Nathenson M (1997) Volcano and earthquake hazards in the Crater Lake
region, Oregon. U.S. Geological Survey Open-File Report 97-487
Balz T, Haala N (2003) SAR-based 3D-reconstruction of complex urban environments. Int Arch Photo-
gramm Remote Sens 34:181–185
Birkmann J (2005) Measuring vulnerability. Report on the 1st meeting of the expert working group
‘‘Measuring Vulnerability’’ of the United Nations University Institute for Environment and Human
Security (UNU-EHS), Bonn
Blaschke T, Strobl J (2001) What’s wrong with pixels? Some recent developments interfacing remote
sensing and GIS. GIS 6:12–17
Briguglio L (2003) The vulnerability index and small island developing states: a review of conceptual and
methodological issues. Department of Economics, University of Malta, Msida
Cardona OD (2003) Indicators for disaster risk management. First expert meeting on disaster risk
conceptualization and indicator modelling. Manizales, March 2003
Clark GE, Moser S, Ratick S, Dow K, Meyer WB, Emani S, Jin W, Kasperson JX, Kasperson RE, Schwarz
HE (1998) Assessing the vulnerability of coastal communities to extreme storms: the case of Revere,
MA, USA. Mitigation Adapt Strat Glob Change 3(1):59–82
CINDI (1998) Central America: Hurricane Mitch OFDA-12: 12-Nov-98. USAID/OFDA Hurricane Mitch
Fact Sheet 12
Coburn AW, Spence RJS, Pomonis A (1994) A vulnerability and risk assessment, 2nd edn. United Nations
Development Programme (UNPD)
Cressie NAC (1991) Statistics for spatial data. Wiley, New York
123
Nat Hazards (2009) 48:275–294 293
Cutter SL, Boruff BJ, Shirley WL (2003) Social vulnerability to environmental hazards. Soc Sci Q 82:242–
260. doi:10.1111/1540-6237.8402002
Dash J, Steinle E, Singh RP, Baehr HP (2004) Automatic building extraction from laser scanning data: an
input tool for disaster management. Adv Space Res 33:317–322. doi:10.1016/S0273-1177(03)00482-4
Davidson C (2006) Was 2005 the year of natural disasters? Bull World Health Organ 84:4–8
Draper NR, Smith H (1981) Applied regression analysis. In: Bradley RA, Hunter JS, Kendall DG, Watson
GS (eds) Wiley series in probability and mathematical statistics. Wiley, New York
Dwyer A, Zoppou C, Nielsen O, Day S, Roberts S (2004) Quantifying social vulnerability: a methodology
for identifying those at risk to natural hazards. Geoscience Australia
Eakin H, Luers AL (2006) Assessing the vulnerability of social-environmental systems. Annu Rev Environ
Resour 31:365–394. doi:10.1146/annurev.energy.30.050504.144352
Emdad Haque C, Etkin D (2007) People and community as constituent parts of hazards: the significance of
societal dimensions in hazards analysis. Nat Hazards 41:271–282. doi:10.1007/s11069-006-9035-8
Espindola G, Camara G, Reis I, Bins L, Monteiro A (2006) Parameter selection for region-growing image
segmentation algorithms using spatial autocorrelation. Int J Remote Sens 27:3035–3040. doi:
10.1080/01431160600617194
Flint CG, Luloff AE (2005) Natural resource-based communities, risk, and disaster: an intersection of
theories. Soc Nat Resour 18:399–412. doi:10.1080/08941920590924747
Fraser C, Baltsavias E, Gruen A (2002) Processing of Ikonos images for submetre 3D positioning and
building extraction. ISPRS J Photogramm Remote Sens 56:177–194. doi:10.1016/S0924-2716(02)
00045-X
Gao Y, Kerle N, Mas JF, Navarrete A, Niemeyer I (2007) Optimised image segmentation and its effect on
classification accuracy, Spatial Data Quality 2007, Enschede, 4 pp
Giardino M, Mondino EB, Perotti L (2005) Geomatics application for geological and geomorphological
mapping and landslide hazard evaluation: case studies in the Italian NW-Alps. In: 31st International
symposium on remote sensing of environment, Saint Petersburg
Haki Z, Akyuerek Z, Duezguen S (2004) Assessment of social vulnerability using Geographic Information
Systems: Pendik, Istanbul case study. In: 7th AGILE conference on geographic information science,
Heraklion
Herold M, Clarke K (2007) Remote sensing derived mapping to support urban growth theory. In: Weng Q,
Quattrochi DA (eds) Urban remote sensing. Taylor & Francis, 500 pp
Herold M, Liu X, Clarke KC (2003) Spatial metrics and image texture for mapping urban land use.
Photogramm Eng Remote Sens 69:991–1001
Hirschmugl M, Gallaun H, Perko R, Schardt M (2005) ‘‘Pansharpening’’ - Methoden für digitale, sehr
hochauflösende Fernerkundungsdaten. In: 17. AGIT symposium, Salzburg
Iverson R, Schilling S, Vallance J (1998) Objective delineation of lahar-inundation hazard zones. Geol Soc
Am Bull 110:972–984. doi:10.1130/0016-7606(1998)110\0972:ODOLIH[2.3.CO;2
Jain S (2005) System evolution using high resolution satellite data for urban regimes. PhD thesis,
Department of Architecture and Planning, Indian Institute of Technology Roorkee
Kerle N (2002) Volume estimation of the 1998 flank collapse at Casita volcano, Nicaragua: a comparison of
photogrammetric and conventional techniques. Earth Surf Process Landf 27:759–772. doi:10.1002/
esp. 351
Lawrence W, Imhoff M, Kerle N, Stutzer D (2002) Quantifying urban land use and impact on soils in Egypt
using diurnal satellite imagery of the Earth surface. Int J Remote Sens 23:3921–3937. doi:10.1080/
01431160110115951
Lee S, Choi J, Min K (2004) Probabilistic landslide hazard mapping using GIS and remote sensing data at
Boun, Korea. Int J Remote Sens 25:2037–2052. doi:10.1080/01431160310001618734
Müller M, Segl K, Heiden U, Kaufmann H (2006) Potential of high-resolution satellite data in the context of
vulnerability of buildings. Nat Hazards 38:247–258. doi:10.1007/s11069-005-8637-x
Myint SW (2007) Urban mapping with geospatial algorithms. In: Weng Q, Quattrochi DA (eds) Urban
remote sensing. Taylor & Francis, 500 pp
Nakai S, Itoh T, Morimoto T (1999) Deaths from heat-stroke in Japan: 1968–1994. Int J Biometeorol
43:124–127. doi:10.1007/s004840050127
O’Hare G, Rivas S (2005) The landslide hazard and human vulnerability in La Paz City, Bolivia. Geogr J
171:239–258. doi:10.1111/j.1475-4959.2005.00163.x
Ramsey M, Flynn L (2004) Strategies, insights, and the recent advances in volcanic monitoring and mapping
with the data from NASA’s Earth Observing System. J Volcanol Geotherm Res 135:1–11. doi:
10.1016/j.jvolgeores.2003.12.015
Ranganath P (2000) Mitigation and the consequences of international aid in post disaster reconstruction.
McGill University and University of Colorado, Natural Hazards Research Working Paper 103
123
294 Nat Hazards (2009) 48:275–294
Rashed T, Weeks J (2003a) Assessing vulnerability to earthquake hazards through spatial multicriteria
analysis of urban areas. Int J Geogr Inf Sci 17:547–576. doi:10.1080/1365881031000114071
Rashed T, Weeks J (2003b) Exploring the spatial association between measures from satellite imagery and
patterns of urban vulnerability to earthquake hazards. Int Arch Photogramm Remote Sens Spat Inf Sci
XXXIV 7(W9):144–152
Ridd MK (1995) Exploring a V-I-S (vegetation-impervious surface-soil) model for urban ecosystems
analysis through remote sensing: comparative anatomy for cities. Int J Remote Sens 16:2165–2185.
doi:10.1080/01431169508954549
Saaty TL (1980) The analytic hierarchy process. McGraw-Hill International Book Company, New York
Shamaoma H, Kerle N, Alkema D (2006) Extraction of flood-modelling related base-data from multi-source
remote sensing imagery. In: ISPRS mid-term symposium 2006 remote sensing: from pixels to pro-
cesses, Enschede, 8–11 May 2006
Smit B, Pilifosova O (2003) From adaptation to adaptive capacity and vulnerability reduction. In: Huq S,
Smith J, Klein RTJ (eds) Enhancing the capacity of developing countries to adapt to climate change.
Imperial College Press, London
Stilla U, Sörgel U, Thönnessen U (2003) Potential and limits of InSAR data for building reconstruction in
built up-areas. ISPRS J Photogramm Remote Sens 58:113–123. doi:10.1016/S0924-2716(03)00021-2
Tralli DM, Blom RG, Zlotnicki V, Donnellan A, Evans DL (2005) Satellite remote sensing of an earthquake,
volcano, flood, landslide and coastal inundation hazards. ISPRS J Photogramm Remote Sens 59:185–
198. doi:10.1016/j.isprsjprs.2005.02.002
Tuceryan M, Jain A (1998) Texture analysis. In: Chen C, Pau L, Wang P (eds) The handbook of pattern
recognition and computer vision, 2nd edn. World Scientific Publishing Co., pp 207–248
UNDP (2005) Human development report. International cooperation at a crossroads. Aid, trade and security
in an unequal world, New York
van Westen CJ, Kumar Piya B, Guragain J (2005) Geoinformation for urban risk assessment in developing
countries; the SLARIM project. In: 1st International symposium on geo-information for disaster
management, Delft
Villagrán de León JC (2006) Vulnerability. A conceptual and methodological review. UNU-EHS, Bonn
Vosselman G, Gorte BGH, Sithole G, Rabbani T (2004) Recognising structure in laser scanner point clouds.
Int Arch Photogramm Remote Sens Spat Inf Sci 36(8/W2):33–38
Wisner B, Blaikie P, Cannon T, Davis I (2004) At risk. Natural hazards, people’s vulnerability and disasters.
Routledge, London
Wu SY, Yarnal B, Fisher A (2002) Vulnerability of coastal communities to sea-level rise: a case study of
Cape May County, New Jersey, USA. Climate Res 22:255–270
Zhang Y, Maxwell T (2006) A fuzzy logic approach to supervised segmentation for object-oriented clas-
sification. In: ASPRS 2006 annual conference, Reno
123