To enhance the estimation accuracy of economic loss and casualty in seismic
risk assessment, a high-resolution building exposure model is necessary.
Previous studies in developing global and regional building exposure models
usually use coarse administrative-level (e.g. country or sub-country
level) census data as model inputs, which cannot fully reflect the spatial
heterogeneity of buildings in large countries like China. To develop a
high-resolution residential building stock model for mainland China, this
paper uses finer urbanity-level population and building-related statistics
extracted from the records in the tabulation of the 2010 population census of
the People's Republic of China (hereafter abbreviated as the
“2010 census”). In the 2010 census records, for each province, the
building-related statistics are categorized into three urbanity levels
(urban, township, and rural). To disaggregate these statistics into
high-resolution grid level, we need to determine the urbanity attributes of
grids within each province. For this purpose, the geo-coded population
density profile (with 1 km
The frequent occurrence of earthquakes and other natural hazards (typhoon,
flood, tsunami, etc.) can lead to tremendous and often crippling economic
losses. According to the estimation in Daniell et al. (2017), from
1900–2016, 2.3 million earthquake fatalities from 2233 fatal events occurred
worldwide. Economic losses (direct and indirect) associated with the
occurrence of over 9900 damaging earthquakes reached USD 3.41 trillion (in
2016 prices). For cases in China, the combination of high seismic activity,
population density, and building vulnerability caused even higher seismic
risk: earthquakes that occurred in China during the 110 years from 1900 to
2010 accounted for about 2.5 % of radiated energy globally, but the
earthquake fatality ratio is around
Modelling seismic loss to buildings requires quantifying their exposure in terms of floor area and monetary value (Paprotny et al., 2020). A series of micro-, meso-, and macro-scale approaches have been developed for this purpose. The scale of the method depends not only on the size of the study area but also on the goal of the investigation, the availability of necessary data, time, money, and human resources (Messner and Meyer, 2006). For example, micro-scale analyses calculate the asset value based on individual buildings, which requires detailed information on building characteristics (e.g. occupancy, age, structure type, building height, or the number of floors). However, since great efforts and considerable expenses are required to collect such information for each building, micro-scale methods are rarely applicable on a regional or (inter)national level (e.g. Figueiredo and Martina, 2016; Erdik, 2017). When further limited by the privacy protection issue, information on asset values of individual buildings is more difficult to obtain (Wünsch et al., 2009). In contrast, meso- and macro-scale methods that use aggregated exposure data on building characteristics procured from official statistics and organized in administrative units (e.g. country, province, prefecture, county or district, etc.) are more commonly used in modelling building values exposed to future earthquakes.
Since building-related statistics are usually aggregated at a coarse
administrative level, while seismic hazards are usually modelled with high
spatial resolution, there is a spatial mismatch between exposure data and
hazard mapping (e.g. Chen et al., 2004; Thieken et al., 2006). This
mismatch may delay and mislead the recuse decision-making after large
earthquakes. For example, after the occurrence of the
When using the dasymetric mapping method to spatialize the administrative-level building exposure data, the selection of appropriate ancillary information is thought to be the most difficult part (Wu et al., 2018) since such information should not only be geo-coded and readily available but also have a high correlation with the building exposure data to be disaggregated. A range of remote sensing data (e.g. nightlight data, road density, land use and land type, population spatial distribution datasets, etc.) have been employed as ancillary information in the literature. A detailed summary of these ancillary data are given in the “Data sources and methodology” section.
Based on the aggregated building-related statistics and using the dasymetric mapping method, this paper develops a high-resolution residential building model (in terms of building floor area and replacement value) for seismic risk assessment in mainland China. This issue has been explored in many previous studies, and a series of global and regional building exposure models have been developed. One famous global model is the PAGER (Prompt Assessment of Global Earthquakes for Response) building inventory database, which is the first open, publicly available, transparently developed global model (Jaiswal et al., 2010). However, the PAGER inventory was developed to rapidly estimate human occupancies in different structure types for earthquake fatality assessment. It lacks information in actual building counts and does not use available information from a commercial database or remote sensing data and thus cannot be used for building asset evaluation immediately (Dell'Acqua et al., 2013). To overcome this difficulty, at least partially, the GED4GEM (the Global Exposure Database for the Global Earthquake Model) project develops a complementary approach that can provide a spatial inventory of exposed assets for catastrophe modelling and loss estimation worldwide (Gamba, 2014). The input datasets ingested into the GED4GEM are at multiple spatial scales, from coarse country-level statistics to finer compilations of each building in some sample regions. There are also other global models, such as the series of building stock models released by the Global Assessment Report (De Bono and Chatenous, 2015; De Bono and Mora, 2014; De Bono et al., 2013) of the United Nations International Strategy for Disaster Reduction (UNISDR) and the global exposure dataset created by Gunasekera et al. (2015). When focusing on the modelling of building stock in China, a common limitation shared by these global models is that the building-related statistics they disaggregate are only of country or sub-country level, although finer-level statistics are already available. Thus, a general assumption in the disaggregation process of these global models is that building stock value per capita within the country or sub-country is uniform. A similar assumption is also made in studies that develop building exposure models specifically for China (e.g. Yang and Kohler, 2008; Hu et al., 2010). For computational convenience, such an assumption is acceptable. However, for improving the seismic risk assessment accuracy in each specific country, more detailed aggregated data at a finer level, if available, should be fully employed in the development of their building exposure model.
By considering the depreciation of all physical fixed assets (including
residential and non-residential buildings, infrastructures, tools,
machinery, and equipment), Wu et al. (2014) estimated the wealth capital
stock (WKS) value for 344 prefectures in mainland China using the perpetual
inventory method (PIM). Later, Wu et al. (2018) decomposed the
prefecture-level WKS value into building assets, infrastructure assets, and
other assets with fixed percentage shares of 44 %, 19 %, and 37 % for
all 344 prefectures. And these three asset components were further
disaggregated into 800 m
Based on the county-level building-related statistics extracted from the
2010 census records, Xu et al. (2016b) developed the nation-wide dasymetric
foundation data (including population and buildings) for quick earthquake
disaster loss assessment and emergency response in China by using the
multivariate regression method (Xu et al., 2016a). The multivariate
regression method used in Xu et al. (2016a) was explained in more detail by
Chen et al. (2012) and Han et al. (2013), in which they developed the
population and building exposure models for areas in Yunnan Province. Fu et
al. (2014a) also used the multivariate regression method to produce the
1 km
To overcome the limitations in building exposure models developed for mainland China in previous studies, this paper aims to present an improved method for generating a high-resolution residential building stock model (in terms of building floor area and replacement value) for mainland China. The main improvements in this paper are that (1) compared with global building exposure models, we use finer urbanity-level (urban, township, and rural) building-related statistics extracted from the 2010 census records as model inputs; (2) compared with Wu et al. (2018), in which the building assets are decomposed from the composite WKS value with a fixed percentage share for all prefectures, we use statistics that are directly related to residential buildings for each urbanity level of each province; and (3) compared with Xu et al. (2016b), in which only land use data are employed in the multivariate method to derive the average building floor area density within each grid, we use the ancillary population density profile generated from the 2015 Global Human Settlement Layer (GHSL), which is considered to be the best available assessment of spatial extents of human settlements with unprecedented spatiotemporal coverage and detail (e.g. Freire et al., 2016).
The organization of the paper is as follows. Section 2 (“Data sources and methodology”) firstly describes the building-related statistics to be used as model inputs that were extracted from the 2010 census records (Sect. 2.1), the review and selection of ancillary data to disaggregate these statistics into grid level (Sect. 2.2), and the derivation of residential building floor area and replacement value in each grid based on these statistics and the ancillary data (Sect. 2.3 and 2.4). Then the major results are presented (Sect. 3.1), and comparisons with other independent data sources are conducted (Sect. 3.2). Limitations in this paper and further improvement directions are also discussed in Sect. 4. Conclusions are drawn in Sect. 5.
In dasymetric mapping, the use of finer-scale census data as input and the choice of appropriate ancillary remote sensing data to disaggregate the census data into a higher grid level are the two controlling factors for the quality of the building stock model. For China, after the 2010 sixth population census (namely the 2010 census), detailed statistical data related to residential building characteristics (e.g. building occupancy, structure type, height classes, etc.) are available for each province at the urbanity level (urban, township, rural). These urbanity-level building-related statistics are good data sources to develop the building exposure model for China. To disaggregate these statistics into grid level, the correlation between the ancillary remote sensing data and the building-related statistics needs to be established. Then, the building floor area and replacement value at the grid level can be estimated. Therefore, in this section we introduce the residential-building-related statistics as extracted from the 2010 census records, the review and selection of ancillary remote sensing data to disaggregate these statistics into grid level, and the method to derive the grid-level residential building floor area and replacement value based on these statistics and the ancillary remote sensing data.
The statistics to be used in this paper for building stock modelling are extracted from the tabulation of the 2010 population census of the People's Republic of China (namely the 2010 census), particularly for residential buildings. Like in most countries of the world, the nation-wide population and housing census in China is carried out in 10-year intervals. Detailed statistics for the year 2020 are not publicly accessible yet. Therefore, census data for the year 2010 are used to elaborate the modelling process. In the 2010 census, there are two types of tables: long table and short table. The long table includes summaries based on the surveys of 10 % of the total population in mainland China, while the short table summaries are based on the surveys of the whole population. Statistics on building characteristics (e.g. building occupancy type, height classes, structure type, etc.) are extracted from the long table of the 2010 census. Supplementary demographic statistics (e.g. the total population in each urbanity, the average number of people per family, and average floor area per person) are extracted from the short table of the 2010 census. A detailed introduction of corresponding sources of these data is given in Table 1.
Main data sources used in this paper. Access to these data is provided in the “Code and data availability” section; n/a stands for not applicable.
Note: the “2010 census” under “Data source” is the abbreviation of the “2010 Population Census of the People's Republic of China”; “Data location” refers to the serial number of the table in the original data source (see context in Sect. 2.1 for more details).
For each of the 31 provincial administrative units in mainland China (including five autonomous regions – Xinjiang, Tibet, Ningxia, Inner Mongolia, Guangxi – and four municipalities – Beijing, Shanghai, Tianjin, Chongqing, hereafter all referred to as provinces), statistics on building characteristics in the long table of the 2010 census are aggregated into three urbanity levels (urban, township, rural). The urbanity attribute is determined according to the administrative unit of the surveyed population. As listed in Table 2, these statistics are used as model inputs to develop the grid-level residential building model in terms of floor area and replacement value. Compared with country- and sub-country-level census data used in previous global or regional models, the further categorization of building-related statistics into urbanity level in the 2010 census helps differentiate the spatial heterogeneity of buildings within each province since the building-related statistics of the same urbanity level are from areas with similar development background but different administrative units. The spatial administrative boundaries used in this paper are from the National Geomatics Centre of China (see “Code and data availability” section for access).
In each urbanity, the population sum of the 2015 GHSL profile and the residential-building-related statistics extracted from the 2010 census records.
Continued.
Continued.
Note: the three urbanity attributes, namely urban, township, and rural, are represented by the numbers 1, 2, and 3 in the first column of this table; “Prov_id” refers to the ID number of each province; “Aver. pop. per family” refers to the average number of people per family; “Amp. factor” refers to the amplification factor used to amplify the building-related statistics from 2010 to 2015 (see Sect. 2.1 and 2.4.1 for more details).
Before disaggregating the urbanity-level building-related statistics into
1 km
All primary remote sensing data have their pros and cons when used for dasymetric disaggregation. For example, studies using LULC data (e.g. Globcover, GLC2000, MODIS, GlobeLand30) assume that the population within each land-use type is uniformly distributed, which is a better assumption compared with believing in an evenly distributed population within an administrative unit. But this assumption is not consistent with the real situation (Thieken et al., 2006), specifically in suburban and rural areas, where the dispersion of population is greater than in urban areas (Bhaduri et al., 2007). Therefore, LULC data are inadequate to fully reflect the spatial heterogeneity within each land use or land cover class. In contrast, nighttime light data, acquired by the US Air Force Defense Meteorological Satellite Program (DMSP) Operational Linescan System (OLS) (Elvidge et al., 2007) and provided by the National Oceanic and Atmospheric Administration (NOAA) every year, are considered the most suitable ancillary information for indicating both the distribution and the density of human settlements and economic activities (Wu et al., 2018). Nighttime light data have been widely used to produce grid-based global population and GDP datasets (e.g. Ghosh et al., 2010; Chen and Nordhaus, 2011; Ma et al., 2012). However, the drawbacks of nighttime light intensity data are also obvious. Limited by the operating conditions of DMSP satellites, the range of nighttime light density is within a narrow interval of 0–63, thus leading to the pixel oversaturation in urban centres (Elvidge et al., 2007). For areas other than city centres (e.g. mountainous rural area), the coverage of nighttime light data is incomplete as it cannot correctly reflect the distribution of nonluminous objects (e.g. road transportation facilities, electricity infrastructure). Compared with the LULC and nighttime light data, road distribution data are more frequently used for assessing infrastructure assets since power lines, energy pipelines, water supply, and sewage pipelines are generally buried along the roads (Wu et al., 2018). Currently, road density data can be converted from road networks like OpenStreetMap, which is an openly available but crowdsourced online database (Zhang et al., 2015). As these data are not systematically compiled, there is still room for improvements (Wu et al., 2018).
Given the limitation of all primary remote sensing data, a series of
secondary ancillary datasets are developed based on the combined use of
these primary datasets. For example, the famous LandScan population density
profile was produced by apportioning the best available census counts into
cells based on probability coefficients, which were derived from road
proximity, slope, land cover, and nighttime lights (Dobson et al., 2000).
Based on these primary and secondary ancillary datasets, a series of studies
have been conducted to disaggregate administrative-level building census
data into geo-coded grids. For example, Silva et al. (2015) disaggregated
the building stock at the parish level for mainland Portugal based on the
population density profile at
In this paper, we follow the assumption of Thieken et al. (2006) that the
distribution of residential asset values can be directly reflected by
population distribution. Now the remaining question is how to select appropriate
ancillary population spatial distribution data to disaggregate
building-related statistics in the 2010 census records. The candidate
population datasets include Gridded Population of the World (GPW; Balk and
Yetman, 2004), Global Rural–Urban Mapping Project (GRUMP) population (see
“Code and data availability” section), LandScan (Bhaduri et al., 2007), WorldPop
(Linard et al., 2012) or AsiaPop (Gaughan et al., 2013), PopGrid China (Fu
et al., 2014b), Global Human Settlement Layer (GHSL) population grids
(Freire et al., 2016; Pesaresi et al., 2013), etc. GPW is a product of simple
areal weighting interpolation, and GRUMP is derived through simple dasymetric
modelling, while LandScan is structurally a multidimensional dasymetric model
(Bhaduri et al., 2007). According to Gunasekera et al. (2015), the LandScan
gridded population dataset was identified as the best-suited dataset for
exposure disaggregation, while other gridded population datasets such as GPW
and GRUMP were too coarse in resolution and accuracy. According to Wu et al. (2018), LandScan, AsiaPop, and PopGrid China are the most promising
population density datasets for asset value disaggregation in China since
they all contain high-resolution attributes. However, some population data
of China are missing from the current AsiaPop. And compared with LandScan,
the spatial coverage of PopGrid China is limited, which is due to an
assumption in its development method, namely the multivariate regression
method (Fu et al., 2014a). It was assumed that the spatial distribution of
population is limited to the six land use types recognized from the Landsat
TM images, namely cultivated land, forest land, grass land, rural
residential land, urban residential land, and industrial and transportation
land. However, in reality, the population is distributed more widely beyond
these land use types. Thus, the LandScan dataset was used for the final
disaggregation of building assets in Gunasekera et al. (2015) and Wu et al. (2018). However, due to its commercial nature, the details to create the
LandScan population datasets are less transparent, although it is considered
to be one of the best global population density datasets (Sabesan et al.,
2007). In contrast, the population datasets developed by the GHSL project of
the Joint Research Center of the European Commission based on the global human
settlement areas extracted from multi-scale textures and morphological
features are transparent and freely available. The built-up area in GHSL was
built by combining the MODIS 500 urban land cover (MODIS500) and the
LandScan 2010 population layer and are among the best-known binary products
based on remote sensing (Ji et al., 2020). Preliminary tests confirm that
the quality of the information on built-up areas delivered by the GHSL is better
than other available global information layers extracted by automatic
processing of Earth observation data (Lu et al., 2013; Pesaresi et al.,
2016). Furthermore, different from LandScan, which aims at representing the
ambient population, namely the average population over a typical diurnal
cycle (Elvidge et al., 2007), GHSL population grids represent the
residential population in buildings (Corbane et al., 2017). The
building-related statistics in the 2010 census are also for residential
buildings. Therefore, the GHSL population grids are the best candidate
ancillary information for this paper to disaggregate the urbanity-level
building-related statistics extracted from the 2010 census records into grid
level. The high correlation (
County-level comparison of the population between the 2015 GHSL profile and the 2010 census records.
In the 2015 GHSL population density profile, the number of people in
each geo-coded grid is given (it is worth noting that this dataset has been
updated in 2019 during the preparation of this work). The original
resolution of the 2015 GHSL population density profile is
250 m
Aubrecht and Leon Torres (2015) identify the geospatial areas of mixed and residential grids within the urban extent of the city of Cuenca, Ecuador, by using the Impervious Surface Area (ISA) data as they show strong spatial correlations with the built-up areas. The assumption behind their method was that intense lighting is associated with a high likelihood of commercial and/or industrial presence (which is commonly clustered in certain parts of a city, such as central business districts and/or peripheral commercial zones, and such areas are defined as “mixed-use area”), and areas of low light intensity are more likely to be pure residence zones (defined as “residential-use area”). In Gunasekera et al. (2015), a similar procedure was used in developing the building stock model for the entire globe. The difference is that Gunasekera et al. (2015) sorted the grids according to the population density in the LandScan population dataset and assigned the grid with urban or rural attributes. For each country, the largest and most populated contiguous grids are classified as urban. This step was repeated iteratively until the urban population proportion for each country was reached.
In this paper, to assign the urbanity attributes (namely
urban, township, or rural) to geo-coded population grids in the 2015 GHSL
profile, for each province we follow the urban, township, or rural population
proportions (as listed in Table 3) derived from the population statistics in
the short table of the 2010 census. The assumption behind this urbanity
attribute assignment practice is that the larger the population density in a
grid, the higher its potential to be assigned as “urban”. An example
demonstrating the distribution of the 2015 GHSL population grids assigned
with urban, township, and rural attributes for Baoshan District of Shanghai is
shown in Fig. 2. For instance, in Shanghai, the urban, township, and rural
population proportion derived from the 2010 census records is 76.64 %,
12.66 %, and 10.7 %, respectively. Then, following Gunasekera et al. (2015), the grids (1 km
The population proportions and thresholds used for each province to assign the grids in the 2015 GHSL profile with urban, township, or rural attributes.
Note: for each province, “PT1(urban or township)” and
“PT2 (township or rural)” are the population thresholds to
assign the grids in the 2015 GHSL profile with urban, township, or rural
attributes. According to the population density
An example showing the assignment of urbanity attribute
in the 2015 GHSL population grids for Baoshan District in Shanghai. The
urban and township as well as township and rural population thresholds for Shanghai are
4936/km
The following section introduces the key steps in residential building stock modelling, including the disaggregation of urbanity-level statistics extracted from the 2010 census records into grid level, the reclassification of building subtypes with both structure type and storey class, and the derivation of residential building floor area and replacement value in each grid. The flowchart in Fig. 3 gives an overview of the whole modelling process.
Flowchart of the residential building stock modelling process adopted in this paper (see context in Sect. 2.4 for more details).
Like in many other countries, the population and housing census data in
mainland China are particularly surveyed for residential buildings.
Therefore, the building stock model developed in this paper is for
residential building stock. As listed in Table 2, building-related
statistics extracted from the 2010 census records include the number of
families living in buildings grouped either by the number of storeys (i.e. 1, 2–3, 4–6, 7–9,
The geo-coded population grids in the 2015 GHSL profile with assigned
urbanity attributes (Sect. 2.3) and the number of people living in
buildings grouped by storey number or structure type derived for each
urbanity of each province seem to allow the direct disaggregation of the
2010 census statistics into the 2015 GHSL grids. However, the GHSL
population is for the year 2015, while the derived population living in
different structure type or storey class from the building-related
statistics is for the year 2010. The increase in population and buildings from
2010 to 2015 must be considered. Here we assume that the increase in
population living in buildings grouped by storey class or structure type
from 2010 to 2015 is equal to the increase in population from the
2010 census records to the 2015 GHSL profile. Therefore, for each urbanity
of each province, the derived number of people living in building types
grouped by storey class or structure type (after performing Step 1-1 and 1-2
in Fig. 3) will be further amplified to the year 2015 by multiplying by the
population amplification factor (namely factor
Thus, for each urbanity of each province, the number of people living
in buildings grouped by storey class or structure type in 2015 is derived by
multiplying the original number of families living in different building
types (based on surveys of 10 % of the whole population) in Table 2 by the
factors
As explained in Sect. 2.4.1, after multiplying the original number of
families living in different building types extracted from the 2010 census
records (Table 2, based on surveys of 10 % of the whole population) by the factors
In the following description, we first introduce the reclassification of building subtypes with both storey class and structure type attributes. Then we estimate the population living in each of the 17 building subtypes. Based on the statistics of average floor area per capita in each urbanity level extracted from the 2010 census records (as listed in Table 2), the total floor area of each of the 17 building subtypes in each grid can be derived. Finally, for each building subtype, their replacement value emerges from a multiplication of the floor area with the unit construction price.
By combining the five storey classes (1, 2–3, 4–6, 7–9,
Average unit construction price (per m
After performing the calculations in Step 1 of Fig. 3, the grid-level
populations living in buildings grouped either by the number of storeys (1,
2–3, 4–6, 7–9,
The strategy we employ here to derive the population living in each of the
17 building subtypes of each grid in a series of distribution steps based on
a prioritized ranking of building types and storey classes. For example, we
first assign storey class 1 buildings to the brick–wood structure type and
distribute the storey class
Based on the distribution processes in Appendix A, we derive the number of
people living in each of the 17 building subtypes in each grid. To
derive the residential floor area of each building subtype, the average
residential floor area per capita is needed, which is given in the short
table of 2010 census (namely factor
With the residential building floor area for each building subtype in each
grid being derived in Step 3, to get the corresponding replacement value,
the unit construction prices of the 17 building subtypes need to be
estimated (namely factor
The grid-level residential building floor area and replacement value (unit:
CNY, in 2015 prices) are aggregated into urbanity level
(urban, township, rural) for each province, as listed in Table 5. The total
modelled residential building floor area for mainland China in 2015 reaches
42.31 billion m
The modelled floor area and replacement value of residential buildings in the urban, township, and rural urbanities of the 31 provinces in mainland China.
Note: (a) in this paper, for each of the 17 building subtypes in each grid, the same unit construction price is used to derive the replacement value in different urbanities and provinces, and (b) the modelled floor area and replacement value are for residential buildings (see context in Sect. 3.1.1 for more details).
For better visualization of the modelled floor area at grid level and to help potential readers to conduct direct comparison with other reports or modelling results, we plot the residential building floor area distribution map and the 2015 GHSL population of Shanghai as an example. As can be seen from Fig. 4, grids with a high density of floor area typically cluster in the downtown area (including eight administrative districts, namely Yangpu, Hongkou, Zhabei, Putuo, Changning, Xuhui, Jing'an, and Huangpu) and the Pudong District. This corresponds to the fact that these districts are the most developed in Shanghai.
An example illustrating the building stock model of
Shanghai:
As of now, we have developed a high-resolution (1 km
Due to the lack of officially published datasets on the value of fixed capital stock in China (Wu et al., 2018), previous studies (e.g. Holz, 2006; Wang and Szirmai, 2012) mainly employed the perpetual inventory method (PIM), in which economic indicators (e.g. gross fixed capital formation, total investment in fixed assets, etc.) are used. The resolutions of these estimations were almost exclusively limited at the national or provincial level (Wu et al., 2014). This coarse spatial resolution forms a major obstacle in applying the model in disaster loss estimation, where high-resolution hazard data are used. To overcome this gap, Wu et al. (2014) estimated the net capital stock values from 1978 to 2012 for 344 prefectures in mainland China by using the PIM. In their Appendix Table A1, the net capital stock values calculated in 2012 prices for 344 prefectures were provided, with the depreciation of all exposed assets (i.e. residential and non-residential building structures, tools, machinery, equipment, and infrastructure) being considered.
To compare with the net capital stock value in Wu et al. (2014), the grid-level residential building replacement value modelled in this paper (namely
the gross value of residential building stock) was aggregated into
the prefecture level. Pearson's correlation coefficient (
Prefecture-level comparison of the modelled residential building replacement value in this paper (unit: billions of CNY in 2015 prices) with the net capital stock value estimated in Wu et al. (2014) by using the perpetual inventory method (unit: billions of CNY in 2012 prices). Note: the net capital stock value estimated in Wu et al. (2014) includes the depreciated value of all exposed elements, namely the residential buildings, non-residential buildings, infrastructures, and equipment (see context in Sect. 3.2.1 for more details).
Compared with previous studies related to building stock modelling in China, we have used finer urbanity-level building-related statistics as input to generate the grid-level residential building stock model. In each urbanity, the building-related statistics extracted from the 2010 census records are from areas with a similar development background, but they belong to different administrative units (i.e. prefectures and counties). Also, within the same prefecture or county, the geo-coded grids are of different urbanity attributes. Therefore, the reliability of our model can be better proved if the modelled results correlate well with actual records at the county or prefecture level. After a thorough search, we find that county-level records of residential building floor area are also available for 28 provinces in mainland China, except for Hunan, Liaoning, and Sichuan provinces, for which only prefecture-level records of residential building floor area can be found from the 2010 census records. Then, to compare our modelled floor area with the 2010 census records at the county or prefecture level, the modelled grid-level residential building floor area was first aggregated into counties or districts for the 28 provinces as well as prefectures for Hunan, Liaoning, and Sichuan, respectively. The final comparison between our estimated residential building floor area with that recorded in the 2010 census is plotted in Fig. 7.
Comparison of the sum of the annual fixed asset investment (unit: billions of CNY) on residential buildings with investment on all types of buildings during 2004–2014 in each of the 31 provinces in mainland China. Detailed investment statistics are available from the Supplement.
As can be seen from Fig. 7, there is a high correlation (
The regression parameters and correlation coefficients for population and floor area in each province.
Note: “Pop_a” and “Pop_b” are the linear regression parameters between the 2015 GHSL population
and the 2010 census-recorded population; “FloorArea_a” and “FloorArea_b” are the linear regression
parameters between the modelled residential building floor area in this paper
and that extracted from the 2010 census records;
“Pop_R
County- and prefecture-level comparison of the modelled
residential building floor area (km
From Table 6 we can see that the correlation between the 2015 GHSL
population and the 2010 census-recorded population and the correlation
between the modelled floor area and the 2010 census-recorded floor area are
generally very high for a majority of provinces (with
Since the residential building model developed in this paper is targeted for
seismic risk analysis, we now use the modelled replacement value to estimate
the seismic loss to residential buildings in Sichuan Province caused by the
Wenchuan
Macro-seismic intensity map of the 2008 Wenchuan
Our estimated seismic loss of residential buildings in Sichuan Province due
to the Wenchuan
Distribution of seismic loss ratio (the ratio between
repair cost and replacement cost) of residential buildings in affected
districts and counties of Sichuan Province due to the 2008 Wenchuan
According to studies on assessing the resolution of exposure data required
for different types of natural hazards (e.g. Chen et al., 2004; Thieken et
al., 2006; Bal et al., 2010; Figueiredo and Martina, 2016; Röthlisberger
et al., 2018; Dabbeek et al., 2021), the 1 km
In the future, with the increasing availability of open-source datasets that
track individual building features in detail, the current limitations in
this paper can possibly be overcome. Attempts have been made to combine
publicly available building vector data (which contain the spatial
location, footprint, and height of each building) and census records to
improve the exposure estimation (e.g. Figueiredo and Martina, 2016; Wu et
al., 2019; Paprotny et al., 2020). Algorithms to extract building footprints
and height from aerial imagery and using computer vision techniques have
been used by commercial companies like Google and Microsoft (Parikh, 2012;
Bing Maps Team, 2014). More recently, by using an unmanned aerial vehicle
and a convolutional neural network, Xiong et al. (2020) introduced an
automated building seismic damage assessment method in which not only the
3D building structure can be constructed, but also the building damage state
can be predicted automatically with an accuracy of 89 %. In addition, Li
et al. (2020) developed the first continental-scale dataset on 3D building
structure (including building footprint, height, and volume) at
1 km
In this paper, a 1 km
In Appendix A, to derive the population living in each of the 17 building subtypes of each grid, the distribution strategy mentioned in Sect. 2.4.2 is explained in detail. In addition, a MATLAB script is provided to help understand this strategy.
For each grid, to derive the population living in each of the 17 building
subtypes (their abbreviations are given in Table 4), namely the 17
to-be-solved variables on the left side of the equation set in Sect. 2.4.2,
a series of distribution steps based on a prioritized ranking of building
types and storey classes are used in this paper. A MATLAB script and an
input file illustrating the distribution processes are also available from
the Supplement online. With the help of the MATLAB script, it will be easier
to understand the distribution steps as follows.
For the brick–wood structure type, in each grid if If the remaining population living in brick and wood buildings
For steel–RC structure type, in each grid if
Following the above step (3), if After determining the population living in seven building subtypes (
The access to data used or mentioned in this paper is as follows: (1)
2010 China Sixth Population Census tabulation
(
DX conducted the data collection and preparation, analyses of results, and model validation and prepared the draft manuscript. JED guided the data collection and preparation process, developed the modelling methodology, and performed the calculation of and co-analysed the results. HHT and FW supervised the project and provided advice and feedback in the process. All authors contributed to the revision of the manuscript.
The authors declare that they have no conflict of interest.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
The authors thank the editor Sven Fuchs for actively monitoring the whole review process. We also appreciate the efforts and time spent by the two anonymous reviewers for this work and the two reviewers from a previous round of submission. Their suggestions have greatly improved the quality of this work. We also want to thank the careful review of the language copy-editor and the typesetter of the journal NHESS. Their hard work has greatly improved the presenting quality of this work.
This research was jointly supported by the China Scholarship Council (CSC), the Karlsruhe House of Young Scientists (KHYS) from the Karlsruhe Institute of Technology (KIT), the China Postdoctoral Science Foundation (grant no. 2021M691408), and the National Natural Science Foundation of China (grant no. 41922024). The article processing charges for this open-access publication were covered by the Karlsruhe Institute of Technology (KIT).
This paper was edited by Sven Fuchs and reviewed by two anonymous referees.