Environmental Earth Sciences

, Volume 65, Issue 1, pp 173–182

Social vulnerability assessment of natural hazards on county-scale using high spatial resolution satellite imagery: a case study in the Luogang district of Guangzhou, South China


  • J. Zeng
    • Key Laboratory of Marginal Sea GeologyGuangzhou Institute of Geochemistry, CAS
    • Key Laboratory of Marginal Sea GeologyGuangzhou Institute of Geochemistry, CAS
  • J. L. Zhang
    • Key Laboratory of Marginal Sea GeologyGuangzhou Institute of Geochemistry, CAS
  • T. P. Ouyang
    • Key Laboratory of Marginal Sea GeologyGuangzhou Institute of Geochemistry, CAS
  • S. F. Qiu
    • Key Laboratory of Marginal Sea GeologyGuangzhou Institute of Geochemistry, CAS
  • Y. Zou
    • Key Laboratory of Marginal Sea GeologyGuangzhou Institute of Geochemistry, CAS
  • T. Zeng
    • Key Laboratory of Marginal Sea GeologyGuangzhou Institute of Geochemistry, CAS
Original Article

DOI: 10.1007/s12665-011-1079-8

Cite this article as:
Zeng, J., Zhu, Z.Y., Zhang, J.L. et al. Environ Earth Sci (2012) 65: 173. doi:10.1007/s12665-011-1079-8


Social vulnerability assessment of natural hazards aims to identify vulnerable populations and provide decision makers with scientific basis for their disaster prevention and mitigation decisions. A new method based on remote sensing is presented here to establish a model of social vulnerability for county-scale regions that lack of relative data. To calculate population density, which is the most important indicator in social vulnerability assessment, first, a statistical model is established to estimate the population on village level. Then a new concept defined as “population density based on land use” is created to replace the arithmetic population density. The former has taken the dynamic human distribution related to land use into account; thus, it can map the population distribution more realistically. The other two indicators are age structure and distance to hospital. The application of this method to the Luogang District of Guangzhou, South China demonstrated its capability of providing high spatial resolution and reasonable social vulnerability for social vulnerability assessment of natural hazards.


Social vulnerabilitySPOTLand useStatistical model


Vulnerability assessment is a highly regarded issue in natural disaster research. Some pioneering investigations have been done in different countries on different types of hazards, such as earthquakes, hurricanes, tropical cyclones, volcanic eruptions, coastal hazards, and landslide, etc. (Dibben and Chester 1999; Dixon and Fitzsimons 2001; Armas 2008; Marfai and King 2008; Uzielli et al. 2008). Most of them concentrated on the physical vulnerability, but few of them on social vulnerability.

Although much has been reviewed on the definition of vulnerability and relative literatures from different perspectives of different disciplines (Alwang et al. 2001; Adger 2006), no agreement has been reached upon the definition as yet. In disaster management, vulnerability is defined as the potential loss due to potential damage in a specific area during a certain period. Social vulnerability, as a dimension of vulnerability, refers to the potential loss of life, and its measurement indicators include population density, age structure, risk perception capacity, access to resources, and income of a household, etc.

Presently, there are two dominant measurements of social vulnerability. One is the inventory method, as applied in the landslide risk assessment of Italy using a database of landslides with human consequence (Guzzetti 2000), and that in the Bajo Deba area (north Spain) using an inventory of direct loss during a period of nearly 50 years (Remondo et al. 2008). This method requires detailed historic hazard records, but few places except Hong Kong and Italy have such data sets (Dai et al. 2002; van Westen et al. 2006), making it difficult to put into practice and popularize. The other method, however, aims to identify the sensitive populations that may be less likely to respond to, cope with, and recover from a natural disaster (Cutter et al. 2003; Cutter and Finch 2008). Compared with the inventory method, this one is easier to obtain required data, more flexible and can cater for a wide range of situations. However, there are still some problems with it, such as data decay, the arbitrary nature of boundaries, problems of indicators weighting, etc. (King 2001). To solve these problems, some ways have been introduced recently. For example, Zhang et al. (2007) used physiological population density instead of arithmetic density to obtain more effective value of population distribution. Kienberger et al. (2009) presented a spatial vulnerability unit independent of traditional census unit and administrative boundary. Wei et al. (2004) built a data envelopment analysis (DEA)-based model for the regional vulnerability analysis of natural disasters and avoided the subjectivity from weighting the relative importance of inputs and outputs. In the present study, remote sensing technology is applied to solve the problems mentioned above.

Remote sensing, as a powerful tool to obtain data, is frequently used in risk assessment of natural disasters. But its application was limited in the hazard dimension (Mantovani et al. 1996; Alonso and Roq 1997; Metternicht et al. 2005; Pandey et al. 2008). Usually, remote sensing is used in disaster monitoring and identifying and used as the data source of indicator “land use” in hazard assessment of natural hazards, but rarely in vulnerability assessment, particularly in social vulnerability. In fact, it has great potential in solving social problems (Liverman et al. 1998). Although social economic data have been obtained by remote sensing (Liverman et al. 1998; Sutton et al. 2001; Li and Weng 2005; Wu et al. 2005), they have not been used in further social vulnerability assessment, however.

In the present study, an innovative method based on remote sensing is introduced in the Luogang District of Guangzhou, South China, which is a young county-level administrative region established in 2005, and lacks social-economic data for social vulnerability assessment. The data problem is solved by using remote sensing to provide accessory information to improve the incomplete population census data and help to understand the characteristics of population distribution. The selected indicators include population density on land use, age structure, and distance to hospital. Here, population density on land use is defined as the estimated population active on a specific unit of land use area during a given period, indicating the spatial–temporal distribution of population. Age structure is defined as the ratio of each age group in a population, and distance to hospital means the distance between a certain point and its nearest hospital. By comparison with the traditional method, the result obtained by this method is conceptually rational and more effective.


Study area

Luogang is situated in the middle of Guangzhou, between 23°03′52″N and 23°04′54″N, 113°31′09″E and 113°32′30″E (Fig. 1), covering a total area of 393.22 km2, and controlled by a south subtropical monsoon climate of high temperature and plenty rainfall which often contributes to geological hazards, such as landslide, debris flow, and rock fall. Besides, a lot of anthropogenic steep slopes occur as a result of the fast urbanization and intensive human activities in this area.
Fig. 1

A sketch map showing the location of the study area

Luogang District consists of six sub-districts, including Xiagang, Dongqu, Yonghe, Lianhe, Luogang, and Jiulong (with two small parts, Jiufo and Zhenlong). The economy of the former four sub-districts depends largely on industry, while the latter two mainly on agriculture; thus, big difference of economic conditions occurs between them.

The indicators

Of all the impact factors of social vulnerability, population density is the most important because densely populated areas are more vulnerable to hazards than sparsely populated ones. Arithmetic population density usually served as an indicator by averaging the total population in an administrative area or in the census area in previous studies (Cutter et al. 2000; Anderson-Berry 2003; Liu and Lei 2003; Simpson and Human 2008). But most human activities concentrate in or around the building area; therefore, the spatial and temporal differences of population density should not be neglected. To solve this problem, a new concept defined as “population density based on land use” is presented to indicate the population active on a specific unit of the land use area during a given period.

Age structure is also a key factor. Old people and children are more vulnerable than others because elderly people may have mobility constraints or mobility concerns, and little kids must move out the harm’s way with the help of their parents, thus increasing the burden of care and lack of resilience.

Another indicator is the distance to get help. In this study, the distance to hospital is used. The shorter the distance, the fewer casualties there will be.

Some other indicators, such as gender, race and ethnicity, level of education, may also be very important, but Luogang is a young and data-poor county; thus, we have to make use of the collected data as full as possible and obtain useful information by virtue of remote sensing.

Image interpretation

Image interpretation is the technology to obtain information that is indicated by the difference between various classes of features from remote sensing image. Usually, the more detailed a classification of features is, the more information it can provide. According to Chinese national standard land use classification, the conditions of Luogang, and the requirement of analyses, the classification of this study is established as shown in Table 1.
Table 1

The classification of land use

Class 1

Class 2

Class 3



Industry land


Office land


Commercial residential land

High-rise building (floors ≥ 10)


Mid-rise building (5 ≤ floors < 10)


Low-rise building (1 ≤ floors < 5)







Critical facilities




Water body


Rivers, ponds, reservoirs



The land used for cultivation of dry crops, aquatic crops and fruit trees



The land used for cultivation of trees, bamboos, shrubs, including the slash, except the green in residential area and along the road



The land with growth of weeds

Urban green


The land used for the green in the residential area and along the road

Open land


Young people, children and patients who are more vulnerable are often active in schools or hospitals; hence, schools and hospitals were classified as separate classes to emphasize their vulnerability. The road network is abstracted as line features

The satellite imagery used in this study is captured on 1st January 2009. It is new enough to reflect the current socioeconomic characteristics of Luogang District. To obtain higher spatial resolution and keep more spectral information, a fused image is produced by fusing band 1, 2, 3 of the multispectral SPOT imagery of 10 m spatial resolution with the panchromatic SPOT of 2.5 m spatial resolution. According to the classification in Table 1, the fused image is classified by the Nearest Neighbor classifier, a supervised classification module in eCognition 7.0.

The classification accuracy is good. However, there are still a little mixed classification and misclassification. To obtain better interpretation result, a lot of field investigations have been carried out, and the classification result was converted into vector format through a man–machine conversation, in which the misclassification and mixed classification were corrected manually.

It should be noted that the height of building is one of the difficulties in the interpretation work. Due to lack of another SPOT imagery, stereo image pair is difficult to implement. Although the shadow of imagery can be used as a sign to identify the height of building, the accuracy is not good enough. Therefore, the height of building in the whole Luogang District mainly derives from the above field investigations, during which the dominant number of stories was used as representative of the investigated unit, the subset of a village.

The final interpretation results are shown in Tables 2 and 3.
Table 2

Area (km2) of different land use type


Total area




Water body

Building up

Open land

















































Table 3

Area (km2) of different land use type


Low rise

Mid rise

High rise

Office land

Industry land

Critical facility



Urban green





























































Data processing

Input variables

Besides image interpretation, some basic data are also collected through field investigations and by consultation with public institutions, including police station, health bureau, agriculture bureau, and other municipal services. All the data are listed as follows:
  1. 1.

    Household population on village level;

  2. 2.

    Transient population on sub-district level;

  3. 3.

    Household population aged over 60 and below 7 on sub-district level;

  4. 4.

    Hospital locations;

  5. 5.

    Per capital area of farmland on sub-district level.


The above-mentioned interpretation results and data are used as input variables to calculate the selected indicators.

Population density based on land use

  1. 1.

    More detailed population data

More and more migrant workers have come to Luogang as a result of the quick development of industry, making up the majority of transient population. In some sub-districts, the transient population is even larger than the household population; therefore, population density estimation without taking the transient population into account will give a false result. However, the known transient population is on sub-district level, thus rougher than the known household population which is on village level. To normalize the data level, it is necessary to estimate the transient population on village level.
Population estimation methods can be grouped into two categories: areal interpolation and statistical modeling, and they are different in their goals and required information. The areal interpolation method uses census population data as input and applies interpolation or disaggregation techniques to obtain a refined population surface (Goodchild and Lam 1980; Lam 1983; Wu et al. 2005), whereas the statistical modeling method is based on the correlations between the census data and socio-economic variables. The statistical modeling method is suitable for estimating the population of an area difficult to enumerate (Wu et al. 2005), and it is suitable for the transient population which is also difficult to enumerate. Moreover, the transient population is dominated by adult workers in manufacturing industry (Yao et al. 2008), the distribution of which is subject to land use. Hence, the statistical modeling method based on the correlation between transient population and land use is applied in this study. The census data of transient population for the six sub-districts and the image interpretation results participate in the model training process. It should be noted that, however, Jiulong is treated as two parts; hence, there are seven datasets that are analyzed by Statistical Product and Service Solutions (SPSS), producing the linear regression equation (Eq. 1). The total population consists of the household population and the transient population; hence, Eq. 2 is established.
$$ P_{\text{t}} = 0.125h + 0.009m + 0.000l + 0.004f + 0.08o $$
$$ P_{\text{v}} = P_{\text{t}} + P_{\text{h}} $$
Pv is the total population on village level; Pt is the transient population on village level; Ph is the household population on village level; h, m, l are the high-rise, mid-rise, and low-rise buildings, respectively, in the commercial-residential area; f is the area of industrial land, o is the area of office land. The correlation coefficient of Eq. 1 is 0.999, and its sig value is 0.003. In the T test of linear regression equation, sig value is an indicator of significance level. When sig ≦ 0.05, the regression equation is acceptable. Therefore, Eq. 1 is acceptable.

Figure 2 shows the relationship between the estimated population and its true value. Obviously, the points are distributed along the line XY = 0 (X = Y), with a maximum relative error of 7% and an average relative error of 0.33%. Given the general tolerance of 10%, the error between the estimated population and its true value is tolerable.

  1. 2.

    Population density based on land use in the Luogang District

Population density has two aspects: On one hand, it varies with space and time; on the other hand, it is relatively static in a given area during a certain period. In traditional arithmetic density, only its static aspect is involved. Therefore, it is necessary to introduce a new population density in which both aspects are involved.
Generally, the larger the population of a place where a natural hazard happens, the more casualties there will be; hence, the utilization ratio of a place, or the intensity of human activities at a place, is positively correlated with social vulnerability. However, if two places have the same utilization ratio, the larger the place is, the fewer casualties there will be. Therefore, Eq. 3 can be used to calculate the population density at a place during a certain period.
$$ D = {\frac{{\sum\nolimits_{i = 1}^{p} {T_{i} } }}{S}} $$

Here, Ti is the temporal probability of person i at the place. S is the total area of the place. p is the total population.

Fig. 2

Relationship between the estimated population and its true value

However, the study area is a county-scale region with many places rather than only one, and the activity which should be taken into account is not that of one person but of many people in a community. To simplify the problem, all places and people must be categorized into different types so as to find out the internal relationship between land use type and human activities.

Usually, the time each person spends on each type of land use is regular. For example, workers often work at the industry land in the daytime from Monday to Friday, whereas farmers work on the farmland. Or if you are a civil servant, you will be on the office land at the same time; otherwise, the time each person spends on the road, at home, or on other land use types also have a statistical law. It is easy to understand that human activities mostly depend on their occupation. Assume that human activities follow the principle of proximity; they are usually in the village, and the activity of a vocation is simultaneous. Then an equation can be established as below:
$$ D_{\text{vl}} = {\frac{{\sum\nolimits_{i = 1}^{c} {P_{i} T_{i} } }}{{S_{\text{vl}} }}} $$
Here, Pi is the population of vocation i; Ti is the temporal probability of vocation i for a land use type; then \( \sum\nolimits_{i = 1}^{c} {P_{i} T_{i} } \) is just the estimated population active on the land use type. Svl is the total area of a certain land use type. Dvl is the population density on a certain land use type, which can be conceptually defined as the estimated population active on a specific unit of the land use area during a given period.
Generally, the temporal probability of each population is as follows:
  1. 1.
    1. (a)
      Time spent in warehouses, workshops, and factories:
      $$ {\frac{{ 8\,{\text{h work/day }} \times {\text{ 5 days/week }}}}{{ 2 4 { } \times {\text{ 7 days/week}}}}} \, = { 0} . 2 4 $$
    2. (b)
      Time spent in houses and apartments:
      $$ {\frac{{ 1 4\,{\text{h work/day }} \times {\text{ 7 days/week}}}}{{ 2 4 { } \times {\text{ 7 days/week}}}}} \, = { 0} . 5 8 $$

  2. 2.
    1. (a)
      Time spent in farmland
      $$ {\frac{{ 5\,{\text{h work/day }} \times {\text{ 7 days/week }}}}{{ 2 4 { } \times {\text{ 7 days/week}}}}} \, = { 0} . 2 0 $$
    2. (b)
      Time spent in houses and apartment
      $$ {\frac{{ 1 6\,{\text{h work/day }} \times {\text{ 7 days/week }}}}{{ 2 4 { } \times {\text{ 7 days/week}}}}} \, = { 0} . 6 7 $$

      The more detailed classification of the population, the more accurate the result will be. As mentioned above, however, Luogang is a data-poor county; thus, the population is classified into only two classes, agricultural population and non-agricultural population. Table 4 shows the temporal probability matrix in this case.

Table 4

Temporal probability matrix in this case

Land use type




Commercial-residential area



Industry land/office land









Then the equation below is put forward to calculate the indicator “population density based on land use”.
$$ D_{\text{vl}} = {\frac{{(P_{\text{va}} T_{\text{la}} + P{}_{\text{vc}}T_{\text{lc}} )}}{{S_{\text{vl}} }}} $$
Here, Dvl is the population density on a certain land use type; Pva is the agricultural population of a village; Pvc is the non-agricultural population; Tla is the time probability that agricultural population spends on the land use type; Tlc is the time probability that non-agricultural population spends on the land use type; Svl is the total area of the land use type.
Since only the agricultural population data on sub-district level are available, the data have to be normalized into village level based on the interpretation results again.
$$ P_{\text{va}} = \left(P_{\text{da}} \times {\frac{{I_{\text{v}} }}{{I_{\text{d}} }}}\right) $$
$$ P_{\text{vc}} = P_{\text{v}} - P_{\text{va}} $$
It is assumed in the equations above that agricultural population correlates the area of farmland positively. Here, Pda is the agricultural population on sub-district level; Iv/Id is the ratio of the farmland area of a village in the total farmland area of the sub-district to which the village is affiliated; Pv is the total population of the village.
According to Eqs. 57, the population density on a certain land use type can be modeled as:
$$ D_{\text{vl}} = {\frac{{P_{\text{da}} \times {\frac{{I_{\text{v}} }}{{I_{\text{d}} }}} \times T_{\text{la}} + \left( {P_{\text{v}} - \left( {P_{\text{da}} \times {\frac{{I_{\text{v}} }}{{I_{\text{d}} }}}} \right)} \right) \times T_{\text{lc}} }}{{S_{\text{vl}} }}} $$
Now all the values of Pda, Iv, Id, Tla, Pv, Tlc, and Svl needed in Eq. 8 are available. After putting these data in their places of Eq. 8, a map of population density based on land use is obtained in Fig. 3.
Fig. 3

Population density based on land use

Age structure

In the present study, the percentage of elderly people and little kids is used as the proxy indicator of age structure. Under conditions that only the data of sub-district level are available, and that the household population, as a relatively closed system, has a static age structure under similar population policy and economic condition, the villages have the age structure of household population as that of the subdistrict which they are affiliated to (Eq. 9). However, the transient population that changes the age structure substantially is absolutely dominated by the middle- and young-aged (Yao et al. 2008) populations; thus, Eq. 10 is put forward as below to calculate the age structure on village level.
$$ A_{\text{dh}} = {\frac{{P_{\text{o}} + P_{\text{c}} }}{{P_{\text{dh}} }}} $$
$$ A_{\text{v}} = {\frac{{A_{\text{dh}} \times P_{\text{h}} }}{{P_{\text{v}} }}} $$
Here, Adh is the household age structure index on sub-district level; Po and Pc are the old population and young child population on sub-district level, respectively, Pdh is the household population on sub-district level; Ph is the household population on village level; Pv is the total population of the village. Av is the age structure index on village level.

The distance to hospital

The distance between the center of a feature and the center of the nearest hospital serves as a proxy indicator.
$$ L = (x - x_{\text{h}} )^{2} + (y - y_{\text{h}} )^{2} $$
Here, (x, y) is the coordinate of the feature’s center. \( \left( {x_{\text{h}} ,y_{\text{h}} } \right) \) is the coordinate of the nearest hospital’s center.

Dimensionless indicators

To make the population density dimensionally consistent with other indicators, the indicator values were divided into ten levels and then assigned a new value between 0 and 1 at an equal interval.

The social vulnerability model and result

Anther important issue is how to transfer individual evaluation values into a composite model that is both scientific and rational. Different composite models represent different ideas and principles of evaluation. At present, there are various evaluation methods, of which the simplest method—exponentially weighted method—has the same efficiency and accuracy as those of complicated ones. Moreover, the present study focuses on the application of remote sensing. Consequently, the exponentially weighted model is selected (Eq. 12) here to perform the population vulnerability evaluation.
$$ F = \sum\limits_{i = 1}^{n} {f_{i} \times w_{i} } $$
F is the social vulnerability index. fi is the value of indicator i. n is the number of indicators. wi is the weight of fi that is created by expert evaluation. In the case of the Luogang district, the weight of population density on land use, the age structure, and the distance to hospital is 0.8, 0.1, and 0.1, respectively.
Four steps are required to evaluate the social vulnerability:
  • First, map the three indicators of social vulnerability assessment, one map for each indicator. After overlapping the three maps (indicators), a map in which each feature has its own value of properties will be produced;

  • Second, add the property item “social vulnerability index” into the above map to save the social vulnerability index;

  • Third, use the “field calculator” module of ArcGIS to perform the evaluation in which the properties are put into their places of Eq. 12;

  • Finally, use hypsometric method to show the spatial difference of social vulnerability visually.

The social vulnerability for the seven parts of the Luogang district evaluated by the traditional method which treats the subdistricts as homogenous can be categorized as seven levels at most. For a convenient comparison between the result of present study and that of the traditional method, the result of present study is also categorized into seven levels with the method of natural breaks (jenks).
Obviously, traditional method (Fig. 5) can only tell which administrative region, or subdistrict, is more vulnerable. The difference between different zones in a subdistrict is obscure, with very little information for decision makers. By comparison, the result of present study (Fig. 4) is much more detailed and clearer.
Fig. 4

Social vulnerability evaluated by present method

Equation 13 is used to estimate the average vulnerability index of zones where people are intensively active in every district since those zones are target regions for disaster prevention and mitigation.
$$ {\text{vul}} = {\frac{{\sum\nolimits_{i = 1}^{n} {A_{i} {\text{vul}}_{i} } }}{{\sum\nolimits_{i = 1}^{n} {A_{i} } }}} $$
vul is the average vulnerability index. Ai is the area of feature i in the classes shown in Table 4. vuli is the social vulnerability index of the feature, n is the total number of those features. The general situation of average social vulnerability in the Luogang district is shown in Fig. 6.
Compared with the result derived from traditional method (Fig. 5), this result (Fig. 6) is relatively satisfactory.
Fig. 5

Social vulnerability evaluated by traditional method

Fig. 6

Average vulnerability level of sub-districts


Our new method based on remote sensing has established a model of social vulnerability for the Luogang District of Guangzhou. The method succeeded in solving the data problem that constrained the social vulnerability evaluation with conventional methods in areas that lack of relative data.

The land use data which were obtained from the interpretation of remote sensing imagery played an important role in this study. With the land use data, the statistical data of socio-economy were normalized to the same level. Considering the relationship between dynamic human distribution and land use, a new concept, population density based on land use, was created to map population distribution more realistically. Besides population density, another two indicators, age structure and distance to hospital, were also considered in this study. The assessment model employed here is an exponentially weighted method. Because of the optimized input data, the output result of this study has more detailed information than estimated with traditional method; thus, it has higher reference value for policy makers.

It can be argued that the application of limited census data with remote sensing is an optimized approach to evaluate the social vulnerability of a place, particularly in those areas with rapid changes but poor data. However, problems concerning objective indicator enrichment and weighting still need further attempts.


This is contribution No. IS-1334 from GIGCAS. This study is financially supported by the project of Finance department in Guangzhou (grant No. GZCY2008FG10008C-2), the science and technology project of Guangzhou (grant No. 2008J1-C051), and the key project of Guangdong province (grant No. 2008A030203003). Thanks are also given to Wu Zhifeng, Zhou Xinjing, Zeng Wei, Gong Jingping, Lv Huiping, Liu Weiping and Lu Wei for their help.

Copyright information

© Springer-Verlag 2011