Assessing Chinese flood protection and its social divergence

China is one of the most flood-prone countries, and development within floodplains is intensive. However, flood protection levels (FPL) across the country are unknown, hampering the present assertive efforts on flood risk management. Based on the flood-protection prescriptions contained in the national flood policies, this paper develops a FPL dataset of China and investigates how China should be protected accordingly and the divergent protections between demographic groups. The 10 dataset agrees with local flood protection plans in 34 of archived 51 counties, validating the policy-based FPLs as a reliable proxy for actual FPLs. The FPLs are much higher than that in the previous global dataset, suggesting Chinese flood risk may have been overestimated. High FPLs (≥50-year return period) are seen in 282 or only 12.6% of the evaluated counties, but with a majority (55.1%) of the total exposed population. However, the low-FPL counties (<50-year return period) host a disproportionate share (52.3%) of the exposed vulnerable population (children and elders), higher than their share (44.9%) of 15 the exposed population. These results imply that, to reduce social vulnerability and decrease potential casualties, investment into flood risk management should also consider the demographic characteristics of the exposed population.


Introduction
Flood protection level (FPL) is the degree to which a flood-prone location is protected against flooding (Scussolini et al., 2016).
It is a key determinant of flood risk, making its quantification a prerequisite to reliable risk assessment (Ward et al., 2013).

20
With the emergence of large-scale flood models, the necessity to quantify FPLs has increased in recent years. For example, expert estimates to improve coastal flood risk assessment. Scussolini et al. (2016) developed FLOPROS, a global database of FPLs based on information included in protection design documents and in protection policy documents, supplemented with 25 FPL estimates based on flood risk modeling.
Improved FPLs reduce the frequency of floods in flood-prone areas and decrease flood risk (Ward et al., 2013). From a cost-benefit view, high FPLs are more economically attractive in areas with high density of population and economy (Ward et al., 2017). However, high FPLs can have a 'levee effect': creating a sense of security and lowering risk awareness, which boosts floodplain development and population growth and can in turn cause catastrophic consequences once a low-probability 30 flood happens (Di Baldassarre et al., 2015;Haer et al., 2020). On the other hand, low FPLs generally mean limited human and financial resources and therefore imply a lower capacity of flood risk reduction (Cheng et al., 2018;Cross, 2001;Han et al., 2020). Moreover, the low FPLs may coincide with a concentration of vulnerable people, e.g., the elders and children, increasing the severity of the human consequences of floods (i.e., more likely fatalities), and more in general exacerbating the local social vulnerability (Birkmann et al., 2016;Gu et al., 2018). Therefore, FPL study is also a key to understand the integrated socio-35 hydrological system.
China is one of the countries that experience the most serious floods and the fastest urbanization. Each year between 1990 and 2017, floods in China affected 149 million people, led to 2165 deaths, and caused an economic damage of US$ 34 billion (Du et al., 2019). Moreover, flood risk changes rapidly due to socioeconomic dynamics (Du et al., 2018) and, in the longerterm, due to climate change (Alfieri et al., 2017;Winsemius et al., 2018). For instance, Du et al. (2018) found that urban lands 40 in the floodplain increased by 26,430 km 2 , i.e., 542%, from 1992 to 2015, a process which is still in full swing and thus likely to exacerbate flood risk in the future. Moreover, the urbanization process witnesses an enormous migration from countryside to cities (Liu and Li, 2017;Li et al., 2018), which selectively leaves the vulnerable population behind, and may increase social vulnerability in the countryside (Cheng et al., 2018). year return period in 29 (or 85.3%) of the 34 Chinese provinces including the capital Beijing, which is probably incompatible 50 with the massive Chinese investment in improving FPLs in the past decades particularly for metropolises (Du et al., 2019). Therefore, it is reasonable to presume that the FPLs of China are significantly underestimated in the FLOPROS, especially for urban areas. With such sparse data, the picture of FPL on the national scale is still unclear, representing a critical knowledge gap in the context of rapid urbanization. This also limits the understanding of the relationship between population exposure, vulnerability, and flood protection. 55 Therefore, the paper here develops and validates the first FPL dataset for China, based on the current Chinese policy on FPLs. On this basis, the following questions are addressed. 1) What level of protection against river floods does Chinese policy imply across the country? 2) Does the FPL policy take into account relevant demographics of the exposed population, such as elders and children who are known to be most vulnerable to floods?

China's flood protection policy and the study framework
FPL data are typically difficult to access at a large scale. Scussolini et al. (2016) proposed that FPL can be assessed based on protection plan documents, policies, or assumed based on hydrodynamic and flood risk simulations and on wealth distribution.
Flood protection policy documents generally contain information on how a region should be protected from floods and provide an opportunity to establish a large-scale FPL dataset (Mokrech et al., 2015;Jonkman, 2013). Presently, the key policy document This is conceptually akin to the policy layer of the FLOPROS dataset, but the framework yields information at the much finer spatial scale of the county. In the study framework, the FPL of an urban county (in Chinese: shi or qu) is evaluated by population exposure and GDP-weighted population exposure; the FPL of a rural county (xian) is evaluated by the population exposure and arable-land exposure (Table 1), as prescribed in the Standard for flood control. Additionally, local flood 75 protection plan documents are collected to verify the policy-based FPL dataset. The spatial pattern of the FPLs is identified using spatial statistic techniques, and the FPL of the exposed (vulnerable) population are evaluated.

Data
Six datasets are employed. First, an administrative boundary is adopted from He et al. (2016), which considered administrative 80 boundary adjustments from 1990 to 2010. Second, a fluvial flood depth map with a 100-year return period is provided by the CIMA foundation (Rudari et al., 2015), which is accessible from the Global Risk Data Platform (http://preview.grid.unep.ch/). It has a spatial resolution of 1 km and has been used for analyzing China's urban land expansion (Du et al., 2018) and Environmental Sciences, Chinese Academy of Sciences, which is provided on the Resource and Environment Data Cloud Platform (http://www.resdc.cn/). It has a resolution of 1 km and is used to extract arable lands in floodplains. Besides, the county-level gross domestic product (GDP) in 2015 is used to calculate the GDP-weighted population exposure.

Assessment of flood protection level
Three exposure indicators are employed to assess the FPL of a certain flood-prone county: population exposure (PopE), GDP-95 weighted PopE, and arable-land exposure (ArableE). For county i, the PopE is the population in the floodplain and is calculated by overlaying the flood depth and the population density maps using a geographical information system (Fang et al., 2018).
Floodplain is defined as the maximum extent (i.e., where flood depth > 0 cm) of the 100-year flood map, which is consistent with previous flood exposure analyses (Du et al., 2018;Fang et al., 2018;Jongman et al., 2012). Then, for an urban county i, the PopE is transformed into the GDP-weighted PopE using the relative factor of the county's GDP per capita to the national 100 average GDP per capita, following Eq. (1), where Gi refers to the GDP per capita in county i; and Ga refers to the national average of GDP per capita in China.

Verification of the flood protection levels
Local documents of flood protection plans are collected for 51 counties to verify the policy-based FPL results. From those documents, the planned FPLs are derived, which are akin to the design layer of the FLOPROS (Scussolini et al., 2016).
Assuming that the planned flood protection standards reflect the reality of flood protection implemented in practice, the agreement between the policy layer and the design layer is checked for each 51 counties and an overall accuracy is further calculated.

Pattern clustering of flood protection level
The LISA (local indicator of spatial association) or local Moran's I (Anselin, 1995) is used to identify spatial relationships 120 between a county's FPL and its neighboring FPLs. The local Moran's I statistic is calculated as follows (Chakravorty et al., 2003), where Ii is the local Moran's I in county i; xi and xj refer to the FPL of county i and its neighboring county j, respectively; ̅ is the mean FPL across all counties; Wi,j is a n-by-n weight matrix defining the spatial contiguity between county i and any 125 county j, where Wi,j =1 if county i and county j share a border and otherwise Wi,j =0; s 2 is the variance of FPLs across all floodprone counties.
A positive value for the local Moran's I statistic indicates that FPL in a county is similar to those in its neighboring counties, while a negative I value indicates dissimilar values (Zhu et al., 2018;Frigerio et al., 2018;Shen et al., 2019). The local Moran's I is calculated by applying the Queen Contiguity matrix in software GeoDa (version 1.12), which is available from 130 http://geodacenter.github.io. The significance is evaluated at an alpha level of 0.05. Four different LISA clustering patterns of FPLs are identified. 1) High-High: both the county and its neighbors have high FPL. 2) High-Low: the FPL is high in a county while low in its neighboring counties. 3) Low-High: FLP is low in a county while high in its neighboring counties. 4) Low-Low: both a county and its neighbors have low FPL.

135
PopE refers to the population exposure in a certain county, which is calculated as the population in the floodplain by overlaying the flood depth and the population density maps (Sect. 2.3). Exposed vulnerable population comprises the exposed children and elders because children and elders are generally considered more vulnerable to flooding, due to limited mobility and https://doi.org/10.5194/nhess-2020-264 Preprint. Discussion started: 24 August 2020 c Author(s) 2020. CC BY 4.0 License.
physical resistance (Gu et al.;Salvati et al., 2018). Assuming that the proportion of exposed children to the total population is spatially homogeneous within a county, the exposed children are calculated in each county using Eq. (4) where PopE refers to the population exposure in the county; children and total population are the number of children and total population in a county, respectively. Similarly, the exposed elders can be calculated. The exposed vulnerable population is the sum of exposed children and exposed elders.
Equation (5) where PopE2015 and PopE1990 refer to the population exposure in 2015 and 1990, respectively. Similarly, the growth rates of exposed children, elders, and vulnerable population are calculated.

150
The policy-based FPL dataset matches to a good degree the information from protection plan documents. In 34 (

Spatial pattern of flood protection level
According to the prescriptions of the policy Standard for flood control, a majority (87.4%, or 1955) of Chinese counties have https://doi.org/10.5194/nhess-2020-264 Preprint. Discussion started: 24 August 2020 c Author(s) 2020. CC BY 4.0 License.
<50-year FPLs that are defined hereafter as relatively low FPLs (Fig. 2), while only 282 counties (12.6%) have high FPLs 160 (≥50 years). A considerable proportion (33.1%, or 741) of the evaluated Chinese counties are protected with a ≥30-year FPL, which is much higher than that in the global FLOPROS database (Scussolini et al., 2016), in which only 5 (14.7%) out of 34 provinces have ≥30-year FPLs. Therefore, Chinese FPLs are significantly underestimated in previous studies. [Insert Fig. 2] The FPLs show significant divergence between eastern and western China (Fig. 2, Fig. 3 "High-high" FPL clusters include 112 counties. They are mainly located in the three primary urban agglomerations of the Beijing-Tianjin-Hebei, the Yangtze River Delta, and the Pearl River Delta (Fig. 4). The three primary urban agglomerations are home to most of the counties with the highest FPLs of ≥200 years (Fig. 3). Besides, the "high-high" FPL clusters are also located in the middle Yangtze River reaches. The "low-high" FPL clusters include a total of 66 counties, surrounding the "high-175 high" FPL clusters. These counties within the "low-high" FPL clusters can be more vulnerable when they are needed to sacrifice to protect their surrounding large cities that are more expensive to be flooded (Wang et al., 2016). "Low-low" FPL clusters include 158 counties, which are mainly located in southwestern China and scattered along a belt from Hohhot to Kunming. Surrounding the "low-low" FPL clusters, 48 counties have relatively high FPLs and form "high-low" clusters.
contrast, 188.4 million (44.9%) of the flood-exposed people are in low-FPL counties, lower than that in the high-FPL counties.
The ratio of vulnerable population to the total exposed population is as high as 20.3% in low-FPL counties, while it is 15.1% 190 in high-FPL counties (Table 2). Both exposed children and elders are found disproportionally in the low-FPL counties.
Specifically, the children's share of the total exposed population is 10.3% in the low-FPL counties, higher than in the high-FPL counties (7.5%); similarly, the elders' share is 10.0% in the low-FPL counties, higher than in the high-FPL counties (7.6%).
Therefore, the low-FPL counties have a disproportionally high share of vulnerable population than the high-FPL counterparts, in terms of both exposed children and elders.

Changes in the exposed (vulnerable) population across protection levels
The total exposed population has grown by 60.3%, rapidly from 1990 to 2015 in counties that are presently protected by high FPLs, while it has remained relatively stable in the low-FPL counties (2.34%) (Fig. 5a). In 1990, the exposed population was primarily located in counties with 20-30 years FPLs (95.0 million, 28.9%), while in 2015 it is primarily in counties with ≥200-

200
year FPLs (97.8 million, 23.3%). [Insert Fig. 5] The exposed vulnerable population has decreased by 41.9%, from 126.0 million in 1990 to 73.3 million in 2015; and decreased more sharply (by 53.7%) in the low-FPL counties (Fig. 5b). The decrease of the exposed vulnerable population is mainly caused by a sharply declining exposed population of children. The exposed children, in total, has decreased by 65.6% 205 from 106.8 million in 1990 to 36.8 million in 2015. The exposed children's share to the total exposed population has declined rapidly across all FPLs, which decreases the ratio of vulnerable population to the total exposed from 38.4% in 1990 to 17.5% https://doi.org/10.5194/nhess-2020-264 Preprint. Discussion started: 24 August 2020 c Author(s) 2020. CC BY 4.0 License.
In contrast, the exposed elders has increased across all the six FPLs, with a total growth by 90.2% from 19.2 million to 36.5 million (Fig. 5b). This trend reflects China's aging population. Moreover, the elders' share of the total exposed population 210 has risen from 5.9% in 1990 to 8.7% in 2015 (Fig. 6). Particularly in the low-FPL counties, it has increased from 5.7% in 1990 to 10.0% in 2015 with a growth of 4.3%, much higher than that in the high-FPL counties (1.7%).
[Insert Fig. 6] 4. Discussion However, the high flood protection does not represent absolute safety. On the contrary, low-probability floods can still occur and flood protection structures may technically fail, causing residual flood risk (Haer et al., 2020). Particularly, levee breaches can cause a catastrophe for the areas with high density of population and assets (Jongman, 2018). In high-FPL counties, a sense of safety brought by the flood protection structures can reduce the perception of risk and cause "levee effect"

Residual flood risk is nonstationary and should be effectively managed
-boosting floodplain development and increasing flood exposure (Cheng and Li, 2015;Kates et al., 2006). Such a phenomenon is probably at play in China, as suggested by the faster increase of the exposed population in the high-FPL 235 counties than in the low-FPL counties. The rapid increase in the exposed population can exacerbate residual flood risk, rendering these high-FPL areas vulnerable to low-probability and high-impact floods (Koks et al., 2015;Di Baldassarre et al., 2013). The residual risk can be further aggravated by future climate change (Alfieri et al., 2017;Winsemius et al., 2018). Alfieri et al. (2017) indicated that the future annual expected economic losses in China may be the highest of all countries, rising by 1.5-fold to 3.4-fold and reaching 50 to 110 billion EUR/year, based on global warming scenarios of 1.5℃ to 4℃, respectively.

240
The residual flood risk can be higher if the real-world flood protection lags behind the policy requirement and design. In fact, the policy-based FPL dataset only reflects how a county should be protected according to the flood protection policy, which also stipulates that the flood protection should be updated along with population growth and economic development.
Unfortunately, a survey of 2013 found that 44% (284) of the 642 Chinese cities did not update their flood protection planning according to their socioeconomic growth (Cheng and Li, 2015). A neglect of the real-world flood protection lagging behind 245 the policy-based flood protection can distort the selection of adaptation measures.
Flood risk management will inevitably face an ongoing challenge from population growth when reducing the residual risk.
It is predicted that, in the next ten years, Chinese urban population will increase by 17% (United Nations, 2018), which will increase residual flood risk in the high-FPL counties because high-FPL counties are usually urban areas. The flood protection structures should be upgraded along with socioeconomic development and climate change to keep the residual flood risk to an 250 acceptable level (Kwadijk et al., 2010). Non-structural measures such as early warning systems, land use planning, building codes, and insurance/reinsurance can be a complement to the flood protection structures for effectively managing flood risk Jongman, 2018;Du et al., 2020).

Demographics should be included in the flood protection policy
Although the low-FPL counties see less exposed population, the majority (52.3%, 38 million) of exposed vulnerable population 255 are concentrated there. Particularly, the elders' share of the total exposed population was increasing rapidly in these low-FPL counties. These findings are consistent with other studies (Cheng et al., 2018). The low-FPL counties are often located in rural areas with economic downturn and insufficient job opportunities, which causes a large number of young adults to temporarily migrate to cities for work opportunities Meng, 2014). Therefore, it may be difficult for these low-FPL counties to respond to and recover from flooding due to economic backwardness and labor shortages.

260
Hence with more vulnerable people, a higher potential casualty or injury rate caused by floods is expected in these low-FPL counties. The elder Chinese are predicted to more than double from 128 million in 2015 to 348 million in 2050 (The World Bank, 2019), implying an increase of exposed elders. This will further increase social vulnerability and challenge flood risk management, particularly in the low-FPL counties. However, the policy Standard for flood control neglects the demographic characteristics of the exposed population. It is economically reasonable to employ a relative low FPL for areas 265 that have low density of population and economy. Such a strategy, however, may aggregate flood risk because the less protected areas coincide with high social vulnerability that is caused by a disproportional distribution of vulnerable people, particularly elders.
Therefore, local demographic characteristics should be considered for an economically and socially beneficial strategy of flood adaptation (Koks et al., 2015). The low-FPL areas can employ decentralized and soft adaptation measures, such as 270 elevation of buildings, wet flood-proofing and dry flood-proofing to reduce flood vulnerability (Aerts, 2018;Du et al., 2020), since structural flood protections are generally less cost-effective in areas with fewer exposed population (Ward et al., 2017;Jongman, 2018). Considering the relative concentration of exposed vulnerable population in the low-FPL counties, flood risk information, adaptation measures, and emergency plans should be made accessible and understandable to children and elders (De Boer et al., 2014). Communities should pay more attention to children and elders during early warning, evacuation 275 and resettlement; and a one-on-one assistance scheme can be developed at the community level to help the vulnerable people.
Emergency plan and flood adaptation design should consider the particular needs of children and elders, which can be promoted by their participation in the planning and designing processes (Liang et al., 2017).

Limitations and future perspectives
The newly developed FPL dataset reflects how China should be protected against river floods according to the flood protection 280 policy. It does not report actual FPLs although it agrees with local flood protection plans very well. Given the scarcity of the real-world flood protection data, the new dataset can be considered a valid proxy of actual FPLs, and can assist efforts to understand, evaluate, and manage flood risk. Moreover, the real-world FPLs are not fixed but are plausibly updated along with socioeconomic development and climate change. For these reasons, we invite relevant departments, communities, and users to use, verify, and improve the newly developed FPL result (which are accessible as Supplement to this paper). With a wide 285 participation of public stakeholders, the flood protection data can be much improved in the future.
Limitations also come with our data and methods. The exposed population is calculated based on a gridded population dataset from the WorldPop program, which is a disaggregation result of census population using auxiliary variables such as land use conditions and nightlight brightness. However, neither the disaggregation methods nor the auxiliary data are free of uncertainty and error (Smith et al., 2019). Moreover, due to a lack of gridded demographic data, the exposed vulnerable 290 population is calculated assuming that its proportion to the total population is spatially homogeneous within a county. But in fact, the demographic characteristics can be spatially heterogeneous (Han et al., 2007;Qiang, 2019). Nowadays, crowdsourcing population data are emerging, thanks to social media (Goodchild and Glennon, 2010;Smith et al., 2019) and mobile phone records (Wu et al., 2012). These new data can help to improve the exposed (vulnerable) population accuracy and in turn the FPL estimates. The produced FPL dataset shows that western China is dominated by low FPLs while high-FPL counties are concentrated in the east. There are 282 counties with a high FPL (≥50 years), which account for only 12.6% of the total flood-prone counties 305 but host 55.1% (231.1 million) of the total exposed population. In contrast, more exposed vulnerable population (52.3%, 38 million) are concentrated in the low-FPL counties. Moreover, exposed population grows rapidly (by 60.3%) in the high-FPL counties while the proportion of elders increases more rapidly in the low-FPL counties than in the high-FPL counties. These findings imply that the flood protection policy has a relatively efficient strategy to protect the majority of the exposed population within a minority of well-protected counties. However, the rapid growth of exposed population can increase residual 310 flood risk. Moreover, the disproportional concentration and rapid increase of exposed vulnerable population, particularly the elders, in the low-FPL counties can probably increase the places' vulnerability.

Conclusions
Therefore, diversified adaptation measures including both structural flood defenses and non-structural solutions should be employed to reduce flood risk in both the high-and low-FPL counties. Local demographic characteristics should be considered for an economically and socially beneficial strategy of flood adaptation. Particularly, the vulnerable population in the low-FPL 315 counties should receive dedicated attention. This study shows that combining FPL and demographic information is critical to understand and manage flood risk.
Data availability. The Chinese flood protection data are available as supplement. Supporting data are accessible through the associated references.   https://doi.org/10.5194/nhess-2020-264 Preprint. Discussion started: 24 August 2020 c Author(s) 2020. CC BY 4.0 License.

Figure 2
The county numbers of different flood-protection levels between western and eastern China. (The boundary between western and eastern China is shown in Fig. 3) https://doi.org/10.5194/nhess-2020-264 Preprint. Discussion started: 24 August 2020 c Author(s) 2020. CC BY 4.0 License.