Spatial heterogeneity and drivers of social vulnerability in chinese floodplains: a PCA-MGWR approach

Abstract

The spatial heterogeneity of social vulnerability in China’s floodplains remains underexplored, particularly regarding how rapid urbanization and unique demographic shifts-specifically the “floating population” and an aging society-reshape human-water relationships. This study quantifies flood exposure across 361 Chinese cities by integrating the 2020 Seventh National Population Census with high-resolution (250 m) floodplain maps via a pixel-based spatial aggregation approach. This methodology captures both absolute population scales and relative structural dependencies within hazard zones, effectively eliminating estimation bias from administrative-level averages. We propose a coupled Principal Component Analysis and Multiscale Geographically Weighted Regression (PCA-MGWR) framework to mitigate multicollinearity among socioeconomic variables while capturing spatially varying scale effects. The analysis identifies four distinct vulnerability dimensions: floating population (PC-FP), livelihood instability (PC-LI), low education (PC-EL), and lack of action capabilities (PC-AC). The PCA-MGWR model (R² = 0.713) significantly outperforms traditional OLS and GWR models in explanatory power. Crucially, results reveal a “Paradox of Wealth Exposure” along the developed Yangtze River Economic Belt, where affluent regions exhibit higher flood exposure, challenging the conventional poverty-driven vulnerability narrative. Furthermore, the driving mechanisms exhibit significant spatial non-stationarity: (1) In eastern megacities, the floating population shows a statistical “risk aversion” effect driven by housing market sorting, yet remains vulnerable due to “institutional segregation” in emergency management; (2) In old industrial bases and eastern plains, an “in-situ aging” phenomenon has created a “spatial lock-in,” trapping elderly populations in high-risk zones. These findings suggest that “mobility” and “aging” have emerged as critical drivers of vulnerability in China alongside economic factors. Consequently, we advocate for spatially differentiated governance strategies, shifting from “one-size-fits-all” defenses to adaptive policies that address specific regional socio-hydrological traps.

Introduction

Due to the combined influence of global climate changes and rapid urbanization, flooding is projected to impact approximately two billion people globally, positioning it as one of the most catastrophic natural disasters¹. Additionally, according to the “UN World Water Development Report 2020”, the effects of flooding are distributed unequally. This disparity compounds existing social inequities, driving Disadvantaged Populations (DPs) deeper into the vulnerability trap. Recognizing these challenges, the Sustainable Development Goals of the United Nations², particularly Goal 13 (climate action) and Goal 11 (sustainable cities and communities), highlight the urgency of reducing climate hazards and promoting Environmental Justice.

As the country with the largest population, China has a long cultural history that, to a great extent, represents a history of coexistence between humans and flooding. However, since the 21st century, China’s unprecedented urbanization has led to the concentration of population and assets along riverbanks and low-lying areas. This creates a notable situation of “competition for land between humans and water”, with significant populations settling in floodplains. Data from “China Report of Flood and Drought Disasters 2023” indicates that 52.7893 million people were affected by flooding in 2023 alone, causing direct asset damage amounting to about USD 2.8 billion. This situation gives rise to a complicated scientific and ethical question: within the special context of China’s social and economic development, who actually lives in areas affected by flooding? Do these vulnerable populations reside in the dangerous areas involuntarily due to poverty, or do wealthy individuals live there actively, driven by the pursuit of economic gains along the rivers?

While existing work about flooding has largely centered on hydrological physical process simulation and assessments of economic losses^3–5, the literature of hazard-bearing bodies remains insufficient, particularly regarding social fragility arising from socio-economic disparities. Existing studies have extensively utilized the Social Vulnerability Index (SoVI) to quantify community resilience, with a focus on static demographic indicators⁶. However, as environmental hazards become more frequent, scholars have begun to emphasize the dynamic nature of vulnerability, transitioning from simple exposure assessments to complex social-ecological-technological systems (SETS) frameworks⁷. Numerous empirical research has been conducted on social inequity in floodplains in the western world, such as African and Latin American communities^8,9. However, these findings cannot be directly applied in China. Flood types in China are complex, encompassing fluvial floods, flash floods (mountain torrents), and urban pluvial floods (waterlogging). These are profoundly influenced by the household registration system, rapid population mobility, and regional development imbalances¹⁰. To date, studies on China’s entire territory have focused primarily on physical exposure¹¹. Recent scholarship has started to integrate China’s unique socio-institutional factors—such as the “Hukou” system and the massive “floating population”—into the vulnerability discourse¹². Despite these advances, the literature still lacks a nation-wide, high-resolution mapping that captures the non-stationary relationship between these unique Chinese drivers and flood exposure. However, there is insufficient mapping of social characteristics of floodplain populations (such as population mobility, age structure, and livelihood patterns). In particular, there is a lack of in-depth investigation into how these factors drive spatially differentiated exposure. While emerging studies have begun to utilize street-level data and local climate zones to analyze the complex driving mechanisms of vulnerability in Chinese megacities¹³, such high-resolution integrated assessments remain scarce in the context of nation-wide floodplains.

There are two methodological challenges to revealing the complex “human-land relationship”. First, China’s socio-hydrological mechanism exhibits significant spatial heterogeneity, ranging from dry river valleys in the northwestern interior to the extensive delta regions of the southeastern coast. Traditional global regression models overlook spatial heterogeneity, rendering them prone to ecological fallacy. Although GWR has been widely used to address spatial heterogeneity, its fixed bandwidth cannot capture the varying spatial influence ranges of different variables¹⁴. To overcome these limitations, the Multi-scale Geographically Weighted Regression (MGWR) has emerged as a superior tool for analyzing spatial processes operating at different scales, yet its susceptibility to multicollinearity when dealing with high-dimensional census data remains a significant hurdle¹⁵. Second, variations of socioeconomic factors, including income, education and employment rate, exhibit multicollinearity. This directly leads to biased coefficient estimations in regression models, thereby misleading policy formulation.

To bridge these gaps in theory and method, this study develops a coupled framework integrating principal component analysis and multiscale geographically weighted regression (PCA-MGWR). Utilizing the data from the 7th National Population Census of China (2020) and spatial data on flood hazards, this study covers 361 Chinese cities, aiming to address three core questions: (1) To investigate the spatial distribution of population exposure to flooding in China’s floodplains. (2) To identify the key dimensions of social fragility with Chinese characteristics. (3) To reveal how the impact mechanisms of these dimensions on flood exposure vary across different geographical divisions.

The contribution of this study lies in transcending the sole perspective of natural disasters, attempting to elucidate the spatial coupling mechanisms between environmental hazards and the social population structure. PCA was employed to eliminate multicollinearity and extract the fragility factors, such as population mobility and livelihood capabilities. By combining PCA and MGWR, this study captures the spatial heterogeneity of these processes. The study aims to reveal the complex relationship of “risk-development” in the Chinese context, providing scientific support for formulating inclusive flood control policies based on regional differences.

Data and methods

Study area

Given the vast territory and complex landscape of China, to facilitate analysis, the country is divided into seven regions according to physical geographical features and administrative divisions, including the Northeast (NE), Northern China (NC), Eastern China (EC), Central China (CC), Southern China (SC), Southwest (SW), and Northwest (NW). The blue areas along the Yangtze, the Yellow, the Pearl and Huai Rivers in Fig. 1 represent the floodplains, which serve as the primary hazard-affected bodies. These areas are not only the core regions for agriculture and economy, but also the most densely populated. Consequently, these regions are prone to significant casualties and economic losses from floods.

Fig. 1.

Study area.

From a spatial perspective (Fig. 1), the distribution of floodplains presents clear geographical features: in the SW region, it is mainly concentrated in plateau lakes and upstream basins of rivers; in the NC region, it is primarily located in downstream areas; while in the plains of middle and lower reaches of Yangtze River, it covers major urban agglomerations. Moreover, China is one of the countries most highly exposed to flood hazards. From 1990 to 2010, the four deadliest flood events globally all occurred in China. With increasing extreme precipitation due to climate change, and the growing concentration of economy and population in floodplains, unprecedented challenges from social vulnerability have emerged.

Data sources and processing

To enhance data transparency and ensure the reproducibility of this research, the primary datasets utilized in this study—including their sources, resolutions, and accessibility—are summarized in Table 1.

Table 1.

Summary of primary datasets used in this study.

Data Category	Data Category	Resolution/Scale	Source/Provider	URL
Floodplain Data	GFPLAIN250m	250 m	NASA SRTM & Hydro-geomorphological Algorithm	https://github.com/fnardi/GFPLAIN
Population Grid	WorldPop (WOPR)	100 m	WorldPop Open Population Repository	https://hub.worldpop.org/geodata/listing?id=135
Census Data	7th National Population Census (2020)	City level	National Bureau of Statistics of China	https://www.stats.gov.cn/sj/pcsj/rkpc/d7c/
Socioeconomic Indicators	China City Statistical Yearbook (2023)	City level	National Bureau of Statistics of China	https://www.stats.gov.cn/sj/ndsj/2020/indexch.htm
Administrative Boundaries	Vector boundaries (2021)	1:1,000,000	National Platform for Common GeoSpatial Information Services (NPCGIS)	https://mulu.tianditu.gov.cn/catalog?catalogDetail=5,MTAwMDAwMA==,1

Floodplain data

The high-precision GFPLAIN250m dataset used in this study was derived from NASA SRTM elevation model and hydrological-geomorphological algorithm¹⁶. With a spatial resolution of 250 m, this dataset accurately identifies the floodplain boundaries worldwide. In contrast to hydrodynamic models that necessitate specific rainfall intensities as direct inputs, this hydro-geomorphic approach identifies the maximal potential hazard zones based on terrain and drainage characteristics. These zones represent the physical landscape most susceptible to inundation during extreme precipitation events, thereby serving as a robust and spatially consistent proxy for national-scale flood exposure assessments. Distribution of floodplains across 361 cities in mainland China was derived from the vector boundaries of administrative divisions published in 2024 by the National Geomatics Center of China (NGCC). Subsequently, the floodplain coverage ratio (representing relative exposure) was calculated by dividing the total area of the identified 250-meter binary floodplain mask within a city’s administrative boundary by the city’s total land area.

Although megacities like Shenzhen and Beijing have adopted avoidance strategies in their urban planning, Fig. 1 shows that avoiding floodplains imposes constraints on urban development. Differences in such physical exposure constitute the foundation of the spatial differentiation of social vulnerability.

Spatial downscaling of population data

To address the limitation of spatial resolution in conventional statistical data, this study utilizes a “top-down” spatial downscaling method. Data from the 7th National Population Census of China (2020) were used in this study, which serve as the latest and authoritative census data. The census data and 100-meter resolution gridded data from WorldPop Open Population Repository (WOPR) were calibrated and integrated to generate a high-resolution raster map of China’s population distribution.

The actual exposed population in each city was accurately calculated by spatial overlay of calibrated population raster and floodplain raster. Specifically, the absolute flood exposure (E_i) for each city i was quantified using a pixel-based aggregation method. We overlaid the 100-meter resolution calibrated population density grid with the binary floodplain mask (where 1 indicates floodplain and 0 indicates non-floodplain). The exposure level was determined by summing the population values of all pixels j situated within the floodplain boundaries of city i:

1

where P_ij denotes the population count in pixel j. This spatial intersection approach effectively eliminates the estimation bias inherent in administrative-level averages.

Social vulnerability variables

Based on the framework of the Social Vulnerability Index (SoVI) proposed by Cutter et al.,⁶ and combined with China’s unique social policies, such as the household registration system and population mobility, nine key indicators were selected to reflect the multi-dimensional nature of disaster susceptibility. The selection criteria prioritized indicators that represent the three core stages of a flood disaster: preparedness, emergency evacuation, and post-disaster recovery. To ensure spatial consistency across all variables, the distribution of specific age groups (i.e., children under 5 and the elderly over 65) was generated using a spatial weight-based downscaling method. Specifically, we multiplied the city-level proportions of these demographic groups, as reported in the statistical yearbook and census, by the 100-meter resolution total population raster. This approach allows for the estimation of the spatial density of vulnerable age groups while preserving the high-resolution physical distribution patterns of the general population. The selected variables and their abbreviations are summarized in Table 2, together with their definitions.

Table 2.

Social vulnerability variables.

Variable	Abbreviation	Description
Medical insurance	MI	Proportion of population covered by medical insurance, reflecting access to healthcare resources during and after flood events.
Unemployment	UE	Proportion of unemployed population, indicating economic instability and limited recovery capacity.
Living alone	LA	Proportion of individuals living alone, representing limited household-level social support.
Unmarried women	UW	Proportion of unmarried women, reflecting potential social and economic vulnerability in disaster contexts.
Rent population	RP	Proportion of population living in rented housing, associated with residential instability and weaker community attachment.
Educational level	EL	Proportion of population with low educational attainment, indicating constraints in risk awareness and access to early warning information.
Action capabilities	AC	Ratio of elderly and young populations to the working-age population, representing mobility and evacuation constraints.
Disposable income	DI	Average disposable income of residents, reflecting economic capacity for disaster preparedness and recovery.
Salary level	SL	Average wage level of employed population, indicating income stability and financial resilience.

Economic dimension: urban-rural resident’s disposable income (DI), wages of on-the-job staff and workers (SL), and unemployment rate (UE). Economic capital directly determines the capacity for disaster preparedness and post-disaster recovery.

Demographic structure dimension: ratio of the elderly to the young population (AC), proportion of people living alone (LA), and the proportion of unmarried women (UW). These individuals usually face limitations in mobility and cognitive abilities during disaster evacuations¹⁷.

Social adaptation dimension: proportion of rental population (RP), proportion of illiterate people (EL), and medical insurance coverage rate (MI). Renters and less-educated groups tend to be deficient in both community attachment and channels for obtaining early warning information¹⁸.

Specifically, the Economic dimension (DI, SL, UE) reflects the financial ‘buffer’ required for post-disaster reconstruction; the Demographic dimension (AC, LA, UW) addresses physical and cognitive constraints during rapid flood onsets where mobility is critical; and the Social Adaptation dimension (RP, EL, MI) captures the information-access barriers and lack of social capital that often prevent marginalized groups from receiving timely early warnings.

There were distinct spatial distribution patterns across the different variables. MI and EL are spatially uniform, whereas UE and DI exhibit extreme variation and skewed distributions, suggesting that their spatial influence on floodplain vulnerability is likely non-stationary.

Data preprocessing

To quantify the social vulnerability variables, we constructed the spatial downscaling through the following three stages:

(1) Population Exposure Identification We identified population distribution relative to floodplains by spatially overlaying the calibrated 1-km WorldPop raster with the GFPLAIN250m dataset. Population grid cells were classified as “inside-floodplain” if their centroids intersected with the floodplain mask, and “outside-floodplain” otherwise. These grid-level counts were then aggregated to the prefecture-level to determine the baseline flood exposure for each city.

(2) Spatial Downscaling of Social Vulnerability To align coarse socioeconomic data with high-resolution floodplains, city-level social vulnerability indicators from the China City Statistical Yearbook (2020) were spatially downscaled to a 1-km raster. Following the top-down approach, we applied the assumption of intra-city spatial uniformity to estimate the population of specific vulnerable groups in each grid cell. This was achieved by multiplying the total grid population by the corresponding city-level proportions, ensuring consistent spatial resolution across all vulnerability layers.

(3) Multi-dimensional Integration and Aggregation The spatially classified exposure data and downscaled vulnerability layers were integrated to form a unified grid-scale database. Finally, these rasterized data were re-aggregated to the city level. The resulting database provides a comprehensive suite of city-level indicators, including total population exposure and the proportions of multiple social vulnerability groups both within and outside floodplain boundaries.

Methods

The overall technical roadmap of this study, ranging from multi-source data integration to the spatial heterogeneity analysis of vulnerability drivers, is depicted in Fig. 2.

Fig. 2.

Methodological framework of the study.

Flood exposure quantification and spatial autocorrelation analysis

(1) Flood Exposure indicators.

To comprehensively characterize urban flood exposure, this study quantified flood exposure from two complementary perspectives: relative exposure and absolute exposure. Specifically, the exposure level was determined by integrating 2020 Census data with 250 m-resolution floodplain maps through a pixel-based spatial aggregation method. This approach ensures that both absolute population counts and relative exposure ratios are accurately captured while eliminating administrative-level estimation biases. These two indicators respectively reflect the structural dependence of urban population on floodplains and the potential scale of population affected by flooding.

Relative flood exposure captures the proportion of the population residing within floodplains, thereby indicating the extent to which urban population distribution is constrained by floodplain, as shown in the Formula 2.

2

Where, denotes the relative flood exposure of city ; represents the population located within floodplains in city ; and is the total population of city.

Absolute flood exposure (Flood-exposed population) measures the total number of people exposed to floodplains, reflecting the potential magnitude of population affected during flood events, as shown in the Formula 3.

3

Where, denotes the absolute flood exposure of city . Unlike relative exposure, this indicator highlights population aggregation exposure to flood, particularly in highly urbanized regions where large populations are concentrated within floodplains.

(2) Bivariate Spatial Autocorrelation.

According to the First Law of Geography, geographical features are spatially correlated with one another¹⁸. To investigate the spatial dependence between flood exposure and social vulnerability variables, the spatial weight matrix was constructed based on the Queen Contiguity criterion.

The global spatial autocorrelation was detected by the Moran’s I indicator (Global Moran’s I), as shown in the Formula 4.

4

Where ; denotes the total number of spatial units; and are the mean values of the absolute and relative flood exposure indicators, respectively; and represents the spatial weight between units and . A significantly positive indicates that cities with high (or low) flood exposure tend to be surrounded by cities with high (or low) vulnerability, reflecting positive spatial coupling, whereas a negative value suggests spatial mismatch. To identify localized spatial association patterns, the local bivariate Moran’s I is calculated as the Formula 5.

5

Where . The local statistic identifies High–High, Low–Low, High–Low, and Low–High spatial association types, revealing clustered and transitional zones of spatial dependence between the two-flood exposure indicator calculated in the Formula 2.

Spatial modeling using the PCA-MGWR

To investigate the social drivers of floodplain settlement, we modeled relative flood exposure as the dependent variable. Unlike absolute counts, relative exposure captures the structural reliance of urban populations on floodplains. We developed an integrated PCA-MGWR framework to overcome the limitations of global regressions and address the inherent multicollinearity and scale-dependent spatial heterogeneity of socioeconomic indicators.

(1) Variable Screening and Multicollinearity Diagnosis.

To preliminarily screen candidate socio-economic variables prior to spatial modeling, Spearman’s rank correlation analysis was employed to quantify the monotonic association between flood exposure and each explanatory variable, as shown in the Formula 6.

6

Where n is the number of spatial units (cities), and d_ij is the difference between the ranks of X_j and for city i.

However, the direct application of MGWR faces significant statistical challenges, specifically multicollinearity. As shown in Fig. 3, nine correlated variables were selected for Pearson correlation analysis. For example, correlation coefficient between LA and RP is as high as 0.83. In the local regression model, the collinearity was magnified, leading to unstable regression coefficients and even sign reversal, which compromised the reliability of the results.

Fig. 3.

Pearson correlation coefficient between DPs variables.

(2) Multicollinearity diagnosis and Principle component analysis.

To diagnose multicollinearity among the social vulnerability variables, the Variance Inflation Factor (VIF) is calculated for each variable. The VIF is defined as the Formula 7.

7

Where denotes the coefficient of determination obtained by regressing variable j on all other explanatory variables. VIF value greater than 10 is commonly regarded as indicating severe multicollinearity.

(3) Dimensionality Reduction via PCA.

To eliminate linear dependence while preserving the dominant variance structure of social vulnerability, we applied Principal Component Analysis (PCA). PCA transforms the standardized variable matrix X into a set of orthogonal (uncorrelated) components Z via an eigenvector matrix L:

8

Each principal component Z_k is a linear combination of the original indicators. We retained a subset of components that captured the majority of the total variance, ensuring a statistically stable and parsimonious set of predictors for spatial modeling.

(4) MGWR Modeling.

Finally, we employed Multiscale Geographically Weighted Regression (MGWR) to examine the spatially non-stationary relationships between flood exposure and the extracted principal components. Unlike traditional GWR, which assumes a single spatial scale for all processes, MGWR allows regression coefficients to vary at variable-specific bandwidths. The model is expressed as:

9

where y_i is the relative flood exposure for city i; (u_i, v_i) denotes the spatial coordinates of the centroid of city i; z_ij is the j-th principal component score; and β_bwj represents the local regression coefficient estimated at an adaptive bandwidth (bwj). This framework robustly addresses multicollinearity while revealing the scale-dependent mechanisms of social vulnerability across China.

Results

Spatial patterns of flood exposure

Main distribution characteristic and pattern of “Hu Huanyong Line”

It can be seen that population exposure in China’s floodplains is significantly spatially unstable. The boundary aligns closely with the Hu Huanyong Line, as revealed by overlaying floodplain and population distribution (Fig. 4a and b). The primary exposure areas are concentrated in the third topographic step east of the line, particularly the flat plains of the NE, North, Central, and East China. These areas are flat plains with a dense water network and centers for agriculture, population, and industry. Conversely, the NW and SW are characterized by high elevations and arid climates, where exposure is minimal except in regional river valleys.

Fig. 4.

(a) Map of the Flood Exposure Ratio; (b) The distribution of the Flood-exposed Population.

Relative exposure: “Locking effect” of the physical environment

According to the Floodplain Coverage Ratio (Fig. 4a and b), cities with high coverage ratio were mainly distributed in Jianghan Plain in the middle reaches of Yangtze River and Huaihe River Basin.

As illustrated in Table 1, the floodplain coverage ratio of Xiantao and Qianjiang reached 100% (1.000), followed by Tianmen at 98.1%. These cities, mainly located in CC, are typical “cities born by water”. Consequently, it is almost impossible for their construction land to avoid the floodplain. This extreme high overlap implies that flood risk is an unavoidable “environmental background” for some small and medium-sized cities, where social vulnerability is primarily derived from Geographical Lock-in.

Absolute exposure: “Aggregation Risk” brought by the urbanization

However, the spatial pattern of the absolute flood-affected population is distinct, as shown in left panel of Fig. 4; Table 3. The cities with the largest number of people threatened by flood are megacities with large population bases, rather than small and medium-sized cities with their entire areas submerged.

Table 3.

Top 10 cities in terms of flood-affected population(left); Top 10 cities in terms of percentage of flood coverage (right).

Geographic region	City	Flood-exposed population	Geographic region	City	Flood exposure ratio
SW	Chongqing	19,690,218	CC	Xiantao	1.000
NC	Tianjin	14,956,967	CC	Qianjiang	1.000
EC	Shanghai	13,830,245	CC	Tianmen	0.981
EC	Suzhou	12,500,023	EC	Bengbu	0.977
CC	Wuhan	10,588,057	EC	Taizhou	0.976
SC	Foshan	9,191,566	SW	Suining	0.968
EC	Nanjing	8,291,991	EC	Huaibei	0.941
SC	Guangzhou	8,045,258	EC	Fuyang	0.941
NC	Beijing	7,984,384	EC	Huainan	0.915
EC	Wuxi	7,335,312	NC	Liaocheng	0.913

Table 3 shows that Chongqing ranks first in China with 19.69 million people, which is associated with its unique mountainous-river terrain and large population size. It was followed by Tianjin (with a population of 14.957 million) and Shanghai (with a population of 13.83 million). It is worth noting that seven of the top ten cities are located along the Yangtze River and developed eastern seaboard, such as Suzhou, Nanjing, Wuhan, and Guangzhou. This finding reveals a typical characteristic of flood risk in China: the “coupling of wealth and risk”. On the one hand, rapid urbanization has attracted a large number of people to riverbanks and coastal areas. On the other hand, once an extreme flood disaster occurs, the potential population exposure and social impacts will be catastrophic, even if high-standard flood defense systems are in place.

Analysis of spatial agglomeration characteristics (LISA analysis)

The Global Bivariate Moran’ s I (0.543) indicate a significant positive spatial association between the flood-exposed population and the flood exposure ratio, suggesting that cities with a high proportion of population residing in floodplains tend to be spatially adjacent to cities with similarly high exposure levels (Fig. 5b). The local bivariate Moran’s I further reveal heterogeneous spatial association patterns between the two variables (Fig. 5a), reflecting distinct regional configurations of exposure intensity and exposed population concentration.

Fig. 5.

The LISA maps between the Flood exposure ratio and flood-exposed population (a) Global Moran’s I; (b) The local spatial autocorrelation.

The High-High clusters are primarily concentrated in the EC region, particularly in urban agglomerations located in the middle and lower reaches of the Yangtze River and the lower reaches of the Yellow River. In these areas, cities with large flood-exposed populations are surrounded by neighboring cities exhibiting high flood exposure ratios, forming contiguous zones of coupled demographic and spatial exposure. This pattern reflects both intensive floodplain development and dense population concentration within major river basins.

(2) Low-Low clusters are predominantly distributed to the west of the “Hu Huanyong Line”, covering most parts of Northwest (NW) and Southwest (SW) China. These regions are characterized by higher elevations, relatively arid climatic conditions, and comparatively limited floodplain development. Cities with low exposed population proportions are spatially adjacent to cities with similarly low exposure ratios, indicating a structurally low exposure regime shaped by both natural and demographic constraints.

(3) High-Low clusters are identified in several upstream cities, including Lanzhou (NW), Diqing (SW), and Kunming (SW), which are in the upper reaches of the Yellow River and the Yangtze River basins. In these cases, cities with relatively high flood-exposed populations are surrounded by neighboring cities with comparatively lower exposure ratios, reflecting localized demographic concentration within otherwise low-exposure regional contexts. Conversely, Low-High clusters are sporadically distributed in parts of the CC and EC regions, where cities with relatively low exposed population shares are embedded within broader high-exposure environments, suggesting localized topographic protection or urban form differentiation within high-risk basins.

Identification of vulnerability dimensions via PCA

Variables screening

Prior to PCA, a Spearman’s rank correlation analysis was performed to screen the candidate social vulnerability variables based on their empirical associations with flood exposure.

As shown in Table 4, UE, LA, RP, EL, AC, and DI exhibit statistically significant correlations with flood exposure (p < 0.01), with correlation coefficients ranging from 0.13 to 0.43, indicating moderate and non-redundant relationships. By contrast, MI and SL show very weak correlations (|ρ| < 0.08) and are not statistically significant, while UW also demonstrates a low and non-significant association (ρ = −0.081). These variables were therefore excluded from subsequent analysis.

Table 4.

Correlation of each Variables with flood exposure ratio.

Social Vulnerability Variables (Abbreviation)	Spearman’s correlation coefficient
Medical insurance (MI)	0.048
Unemployment (UE)	0.333**
Living alone (LA)	-0.2**
Unmarried women (UW)	-0.081
Rent population (RP)	-0.24**
Educational level (EL)	-0.134**
Action capabilities (AC)	0.348**
Disposable income (DI)	0.432**
Salary level (SL)	-0.077

Accordingly, the remaining six variables (AC, UE, LA, DI, RP, and EL) were retained for dimensionality reduction and vulnerability structure identification.

Extraction of main components and their variance contributions

To avoid the multicollinearity among socioeconomic variables and to extract the representative vulnerability features, this study analyzed nine original variables with principal component analysis (PCA). The suitability of the data for factor analysis was confirmed by the KMO measure (0.762) and Bartlett’s test of sphericity (p < 0.001).

PCA (Fig. 6) illustrates that the principal components (PCs) can reduce the dimensionality of data. The eigenvalues of the first four principal components are all greater than one, and their cumulative variance contribution rate reach to 90.18%. This indicates that these four synthetic variables explain more than 90% of the information variance in original dataset, with minimal loss of information. This is sufficient to represent the overall social vulnerability in floodplains in China.

Fig. 6.

PCA Variance/Scree plot.

Sociological interpretation of principal components

According to the rotated component loading matrix Table 5, the four principal components were named to represent the key characteristics of China’s social stratification.

Table 5.

Principal component loading matrix extracted by PCA.

Social Vulnerability Variables (Abbreviation)	PC-FP	PC-LI	PC-EL	PC-AC
Action capabilities (AC)	-0.595	0.552	0.191	0.544
Unemployment (UE)	-0.532	0.595	0.366	-0.338
Living alone (LA)	0.898	0.142	0.091	0.047
Disposable income (DI)	0.552	0.633	0.180	-0.217
Rent population (RP)	0.894	0.130	0.215	0.248
Educational level (EL)	-0.084	-0.619	0.773	0.001

PC-FP: Floating population index

Load characteristics: This component has extremely high positive loading on “one-person household” (0.898) and “proportion of renters” (0.894). This clearly outlines the profile of “new residents” or the migrant population in the urbanization process in China. This group is predominantly composed of cross-regional migrants, relies on rental housing with no ownership of private property, and features an atomized family structure (characterized by living alone). In the sociology of disasters, they often become “social islands” due to a lack of community belonging, weak social networks, and limited economic accumulation, making them marginalized in flood warning and evacuation.

PC-LI: Livelihood instability index

Load characteristics: This component is dominated by “disposable income” (0.633) and “proportion of registered urban unemployed population” (0.595). This represents a complex dimension with distinct Chinese characteristics. Interestingly, income, as the core indicator variable characterizing economic strength, and unemployment rate, as the key proxy variable representing economic risks, exhibit a positive correlation in the same component dimension. This reflects the dual characteristics of “high risk - high return” in China’s urban economy: economically developed cities (high income) often have higher living costs and more intense employment competition, which is manifested as a high unemployment rate in statistics. In comparison, there’s lower income alongside lower unemployment rates in rural areas. Therefore, the PC-LI fundamentally reflects the livelihood fragility in a highly marketized economy, which is that even this group earns a high income, they face a financial predicament caused by cash flow interruption if work stops due to a flood.

PC-EL: Low educational level index

Load characteristics: This component is strongly driven by “illiterate people aged 15 and above” EL(0.773). This dimension directly points to cognitive vulnerability. Low education not only constrains this group’s ability to acquire and understand complex disaster warning information but is strongly related to low socio-economic status. This group represents an “information blind zone” in disaster prevention and mitigation, and they require special communication strategies utilizing visualization or graphical representation.

PC-AC: Lack of action capabilities index

Load characteristics: This component is obviously loaded on the “proportion of elderly and children aged over 65 or under 5” (0.544). This represents physical vulnerability. The elderly and children have limited power to transfer by themselves and thus have to rely on families and public assistance. With the accelerated arrival of China’s aging society, PC-AC has become a key indicator for measuring the pressure of regional flood control and evacuation.

Based on above, PC-FP, PC-LI, PC-EL, and PC-AC have constructed the social vulnerability profile of China’s floodplain population from perspective of social structure, economic livelihood, cognitive level and physiology function. This framework further constitutes the core variables for the MGWR model analysis.

Model performance comparison and optimal model selection

To investigate the effectiveness of the PCA-MGWR framework in assessing flood fragility, this study constructed six regression models to conduct a systematic analysis, including OLS、PCA-OLS、GWR、PCA - GWR、MGWR, and PCA - MGWR. R², Adjusted R², and AICc for all the models are shown in Table 6.

Table 6.

Comparison of the effects of different regression analysis models.

Statistical data	OLS	PCA-OLS	GWR	PCA-GWR	MGWR	PCA-MGWR
R2	0.159653	0.213824	0.5029	0.5799	0.6955	0.713
Adjusted R2	0.150131	0.193492	0.3704	0.4404	0.6041	0.6290
AICc	16.850589	29.318825	911.2513	880.5548	808.4552	781.0033

Characteric differences between the global and locol scale models

First, the conventional global regression model (OLS) exhibits extremely weak explanatory performance. As shown in Table 6, the R² of OLS model is only 0.159653, and the AICc is as high as 16.85. (Note: The AICc value is scaled here for comparison with R² and differs from other spatial model.) This indicates that if the relationship between economic variation and flood exposure is assumed constant across the entire country, as the homogeneity assumption, the model could explain only 16% of the data variance. The results demonstrate the presence of spatial non-stationarity between flood exposure and social vulnerability, indicating that the global linear model could not capture these complex spatial patterns.

Effects of spatial regression model

After being incorporating the spatial perspective, the model’s performance has been improved significantly. GWR Model: The R² could be increased to 0.5029 with this model, which indicates that considering local spatial differences can explain more than half of the variable relationships.

MGWR Model: The R² could be further increased to 0.6955 with this model, and the AICc could decrease from 911.25 in the GWR model to 808.45 (a lower AICc value indicates a better model fit). This significant improvement indicates that the influence range (bandwidth) of different social vulnerability factors, such as floating population vs. aging, on flood exposure varies. Compared with the GWR that forces a uniform bandwidth, MGWR allows variables to have distinct spatial scales, which makes it more in line with the real operation mechanism of the socio-hydrological system.

Optimization of PCA in multicollinearity reduction

Socioeconomic indicators describing social vulnerability are often intrinsically correlated, which poses a challenge for regression-based inference. In this study, variance inflation factors (VIFs) were used to diagnose multicollinearity before and after PCA transformation (Tables 7 and 8).

Table 7.

The VIF of the social vulnerability variables.

Variables	VIF
AC	1.649
UE	1.596
LA	4.003
DI	1.817
RP	4.154
EL	1.722

Table 8.

The VIF of the principal components.

Principle components	VIF
PC-FP	1.693
PC-LI	1.325
PC-EL	1.217
PC-AC	1.969

For the original variables, relatively elevated VIF values are observed for LA (VIF = 4.003) and RP (VIF = 4.154), indicating notable redundancy among demographic and housing-related indicators. Although these values do not exceed critical thresholds, they suggest potential instability in coefficient estimation when directly incorporated into spatial regression models.

After PCA transformation, the VIFs of all retained principal components decrease substantially, ranging from 1.217 to 1.969, well below conventional multicollinearity thresholds. This confirms that PCA effectively removes linear dependence among explanatory variables, providing a statistically stable input structure for subsequent MGWR modeling. Importantly, this dimensionality reduction preserves substantive vulnerability dimensions while filtering out redundant information.

In addition, the benefit of minimizing multicollinearity is clearly observable. Incorporating the PCA-processed variables into the MGWR model (PCA-MGWR) proves to be the best approach. Its R² reached the highest value in this study (0.7125), with an adjusted R² of 0.6290, and the AICc further decreased to 781.00. Compared with the standard MGWR, PCA-MGWR not only enhances model fit, but also minimizes the collinearity disturbances among variables (such as income and unemployment, rent and living alone) to stabilize the local regression coefficients through the input of original PCs.

Based on the above, the ranking of model performance is: PCA-MGWR > MGWR > PCA-GWR > GWR > PCA-OLS > OLS. The PCA-MGWR model is proven to be the best statistical approach to explain the spatial heterogeneity of social vulnerability in China’s floodplains. Therefore, the following analysis of the driving mechanism will be conducted based on the regression coefficients of the PCA-MGWR model.

Analysis of spatial heterogeneity of driving mechanisms

The primary advantage of the PCA-MGWR model is its ability to generate local regression coefficients, thereby revealing the law of how variable relationships change with spatial locations. Figure 7 illustrates the distribution density and range of all the principal component coefficients. It can be seen that the coefficients of PC-FP and PC-AC cross zero, with their values spanning a wide range (e.g., PC-FP ranges from − 1.42 to 0.59). This confirms that the driving mechanisms of social vulnerability exhibit significant spatial instability in China. Subsequently, these data were further mapped into geographical space (Fig. 8), revealing three key driving models. Notably, cool colors (blue and purple) represent positive correlation, while warm colors (red and yellow) represent negative correlation.

Fig. 7.

Standardized distribution of DPs principal component coefficients

Fig. 8.

Spatial distribution of coefficients of DPs principal component variables.

Bandwidth of MGWR

One of the distinct advantages of the MGWR model is its ability to assign independent optimal bandwidths to different socioeconomic variables, thereby reflecting their unique spatial scales of influence. To quantify these variations, the bandwidths of the four principal components were extracted and compared in Table 9.

Table 9.

Optimal bandwidths of variables in the PCA-MGWR model.

Variable	Description	Bandwidth (Number of Cities)	Spatial Scale
PC-FP	Floating Population	30	Highly Local
PC-LI	Livelihood Instability	30	Highly Local
PC-EL	Education Level	358	Global
PC-AC	Action Capabilities (Aging)	358	Global

As shown in Table 9, “livelihood instability (PC-LI)” and “floating population (PC-FP)” exhibit the smallest bandwidth (30 cites), indicating that their influence on flood exposure is highly sensitive to local micro-topography and community-level social structures. In contrast, “the aging population (PC-AC)” and “education level (PC-EL)” show a near-global bandwidth (358 cites). This suggests that “demographic and educational drivers” follow a broader, relatively uniform national urbanization logic across China, whereas “economic and mobility factors” are dictated by immediate, localized environmental factors.

PC-FP: Positive driving effect in Inland areas and risk aversion effect in central-northern region

Figure 8a illustrates the obvious “NW-SE difference” in the effect of mobile population indicator on flood exposure. The NE and SW areas (positive correlation): The coefficients appear in purple and blue (positive values) across most of the NW and SW regions. This means a higher proportion of mobility population is associated with a higher risk of flood exposure in the western inland. This is likely due to the scarcity of plains in the inland mountainous terrain. Migrants often concentrate in low-lying areas or “urban villages” on the edges of river valleys, which face relatively high risks of waterlogging and mountain torrents.

The NC and CC areas (negative correlation): In contrast, coefficients appear in yellow and red (negative values) in the NC and CC regions. This reveals a “risk-avoidance effect”. In these plain cities, the mobility population is pushed to live in newly developed areas or suburbs far from the city center (which are typically non-historical floodplains) due to limited housing costs. Statistically, the higher the floating population, the lower the flood exposure.

PC-LI: “Wealth Exposure” in the Yangtze river basin

Figure 8b reveals the complex relationship between livelihood instability (high income/high risk of unemployment) and flood exposure. Urban agglomeration in the middle of Yangtze River (positive relationship): The coefficients are purple (high positive values) in parts of CC and the SW This result proves the unique “wealth exposure” in China’s floodplains. Compared with the “slum-floodplain” in the western world, the Yangtze River region exhibits a “hydro-dependent” development pattern. High-income groups and high asset valuations are not negatively constrained; instead, they concentrate positively for zonal benefits (transportation, views, and convenience) brought by water. This concentration, coupled with high exposure, implies a greater need for financial resources for disaster recovery.

Northern zone (negative relationship): However, the coefficients in the NC and NE regions are red (negative relationship). This represents that high-income communities are often located in areas of high altitude with adequate drainage facilities in Yellow River area and northern cities with limited water resources, leading to effective flood avoidance.

PC-EL: Spatial mismatch between cognitive vulnerability and human capital

Figure 8c reveals the effect of education level, predominated by proportion of illiterate people, on the flood exposure. This exhibits an obvious pattern of “NW-NE dichotomous opposition”.

The “classical fragile trap” of the northwestern inland (positive relationship): The coefficients in the NW and SW regions are purple and blue (positive values). This means that in the developing western areas, there is a positive relationship between low-education populations and flood risk. This aligns with the classical disaster vulnerability theory: groups with relatively low education levels are often constrained by living hood patterns such as traditional agriculture and animal husbandry. They not only have difficulty conveniently accessing disaster early-warning information, but also lack sufficient cognitive and adaptive capacity to support relocation for risk avoidance, and thus have no choice but to remain in ecologically fragile areas such as river valley fringes and floodplains.

“Labor force concentration”of the NE and coastal areas (negative relationship): However, the coefficients appear red and yellow (negative relationship) in the NE, EC and SC regions. This means that a higher proportion of the population with low education level is associated with lower flood exposure. In contrast, people with higher education level face higher exposure risk.

This counter-intuitive phenomenon highlights a unique characteristic in China known as the “human capital is concentrating on water”. In the old industrial bases of the NE (e.g., Harbin and Shenyang) and the developed cities of the southeast, educational resources and high-tech jobs are clustered in central urban areas along rivers and coasts. Despite being distinct from conventional floodplains, these areas attract a highly educated workforce. Conversely, the less-educated population tends to reside in mountainous or agricultural hinterlands. This spatial mismatch of the “highly educated population - high exposure” reminds us that disaster prevention education should not be overlooked in developed regions either, as a high academic degree does not necessarily mean having a high level of knowledge in hydrology and geography.

PC-AC: Significant “East-west Gradient Disparity” (in mobility and aging)

Figure 8d exhibits a distinct geographic dividing line in the effect of the elderly and activity indicators on flood risk.

Western areas (negative relationship): The coefficients are red (negative relationship) in most parts of the NW and SW regions. This is due to terrain constraints. In the western mountainous areas, conventional settlements with a high proportion of elderly residents are often located in relatively safe areas at higher elevations of the river valley to avoid mountain torrents. Conversely, there are uninhabited areas in the low-lying river floodplains that carry a high risk of flooding. Therefore, areas with a high concentration of the elderly are low-exposure areas.

The eastern plains (positive relationship): This relationship reverses in the areas east of “Hu Huanyong Line”. The coefficients become blue and purple (positive relationship) in the NE, NC and EC regions. There is observed phenomenon of “Aging in Place”. In the old urban areas of eastern plains (often located in high-risk areas along rivers), there is a large population of long-term elderly residents, while many young people have moved to new cities. This phenomenon of intergenerational residential separation has resulted in elderly groups being “trapped” in high-risk flood areas, with a significant positive correlation observed between the two.

Discussion

The “Risk-Development” paradox in the Chinese context: agglomeration and exposure of wealth

Research on environmental justice in the West suggests that the “poverty trap” emerges because disadvantaged populations (low-income groups and minorities) are forced to live in areas of high environmental risk^6,7. However, based on PCA-MGWR analysis, this study reveals a distinct pattern of floodplain distribution in China, characterized as the Paradox of Wealth Exposure. Unlike the “passive marginalization” mechanism prevalent in Western environmental justice literature-where socio-economic exclusion pushes the poor toward hazardous fringes -the driving force in China is a “proactive utility-seeking” mechanism. In this context, the pursuit of transport efficiency and high asset valuations in river-front areas outweighs perceived flood risks. This transforms the survival-based vulnerability of traditional agricultural societies into a systemic, asset-based vulnerability characteristic of rapid urbanization. This finding stands in contrast to the conventional “poverty trap” narrative often observed in the Western world, where disadvantaged populations are forced into high-risk areas due to economic marginalization. While global mappings by Fox et al.⁴ suggest a complex interplay between wealth and risk, our results demonstrate that China’s unique “water-dependent” development model has led to a proactive aggregation of high-value assets in floodplains, transforming survival crises into systemic economic risks. Historically, Chinese culture depended on the development of big river agriculture. Modern industry and urbanization have maintained this route, leading to a high concentration of population and assets in plains that are rich in water resources but also prone to floods.

This “wealth exposure” does not imply a decrease of fragility, but rather a transformation of risk. It represents a shift from the “survival risk” of conventional agricultural societies to the “systemic economic risk” of modern cities. Although these areas possess strong economic recovery capabilities, the high density of assets means that astronomical economic losses can result from damage to flood control engineering projects, such as “black swan” events triggered by extreme rainfall, leading to chain reactions within supply chains. This systemic risk shift highlights the critical need for advanced predictive modeling to evaluate hazard risks under climate change, ensuring the integrated protection of both regional economic assets and cultural heritage²⁰. Therefore, flood management in China must move beyond the perspective of pure poverty relief and address the reality of “high risk anchored by high economies”, aiming to establish a disaster risk-sharing mechanism suitable for the scale of the economy. This pattern reveals a distinct flood “exposure–development” paradox shaped by China’s long-term “human–water” relationship. On the one hand, legacy agricultural settlement along major rivers has led to persistent population concentration on floodplains, where communities have gradually transformed into socially vulnerable groups—such as the elderly and low-income residents—who face “passive entrapment” in high-risk zones due to limited housing options and economic constraints. On the other hand, rapid, asset-intensive urbanization since the 21st century has reinforced flood exposure in developed metropolitan areas, where high-income groups proactively settle in floodplains to capture “waterfront” benefits, such as transport convenience and aesthetic value, often underestimating hazards. These two mechanisms produce a heterogeneous exposure pattern: while wealthy regions require market-based risk transfer like catastrophe insurance to protect concentrated assets, impoverished areas necessitate state-led social safety nets to break the “poverty-hazard” cycle. Furthermore, the adoption of explainable analytical frameworks and causal inference provides a robust pathway for unveiling the complex regional drivers of environmental hazards, thereby supporting more precise and science-based management strategies²¹.

Hidden social vulnerability: The double challenge of mobility and aging

Although “wealth exposure” is observed at the macro scale, society exhibits high heterogeneity and hidden vulnerability at the micro community scale. Here, the model identifies two core vulnerable populations, each driven by distinct mechanisms.

“Institutional Segregation” of The PC-FP

PCA results reveal that the migrant population (renters and living alone) has become the primary dimension explaining China’s flood vulnerability. MGWR analysis (Fig. 8a) reveals the double burden faced by this group: in underdeveloped inland cities, they are forced to concentrate in low-lying urban villages, facing direct physical exposure. In contrast, in developed coastal cities, while they may be pushed to live in outer suburbs by exorbitantly priced housing, which appears as a “risk aversion effect” in statistics, the effect is passive and renders them vulnerable. In first-tier cities such as Shanghai and Shenzhen, housing market sorting plays a pivotal role. Historic city centers (often at high risk of flooding) force the migrant population to relocate to higher elevations or the later-developed suburbs due to their distance from water systems. This “risk aversion” is not an autonomous choice based on risk perception, but rather a passive replacement driven by the cost of living. This suggests that low exposure does not equate to low vulnerability. They are confronted with difficulties in commuting and the risk of marginalized social isolation.

Non-local long-term mobile tenants often lack the motivation to actively invest in disaster prevention facilities due to their lack of historical memory regarding local disasters. More importantly, the existing community disaster reduction system is often based on household registration management, which causes the migrant population to become “invisible people” in terms of accessing early warning information and the allocation of emergency shelters. This “institutional isolation” is even more dangerous than physical exposure. This phenomenon resonates with the environmental injustice reported in Miami and Houston⁹, where Hispanic immigrants face heightened risks due to social and linguistic barriers. However, our study identifies that the driving mechanism for China’s floating population is rooted in “institutional segregation” rather than cultural barriers. The Hukou (household registration) system functions as a socio-political filter that renders mobile individuals administratively “invisible”. This institutional driver creates a unique vulnerability where migrants are physically present in high-risk zones but are excluded from the official community-based emergency management and early warning frameworks anchored to permanent residency status. Our study extends this by illustrating how China’s Hukou (household registration) system functions as a similar socio-political filter, rendering the floating population ‘invisible’ within official emergency management frameworks.

“Spatial Lock-in” of the PC-AC

As China enters a deeply aging society, the overlap of aging and flood risk (Fig. 8d) poses a safety hazard. The results showed that in the eastern plains and old industrial bases, there is a significant phenomenon of “in-situ aging”. Elderly residents remain in old urban areas along rivers where facilities are aging, while young people with mobility move out.

This intergenerational residential separation leaves the elderly in an “empty nest” predicament when floods strike, making it difficult for them to receive timely assistance from their families. Compared with the elderly in the western mountainous areas who naturally live in highlands due to terrain restrictions, the elderly in the eastern plains are actually “locked” in high-risk flood areas. If no targeted intervention is carried out, this group will be the one with the highest casualty rate in future extreme climate events. The identified “spatial lock-in” of the elderly in China’s eastern plains corresponds to the vulnerability patterns reported in other aging societies like Japan¹⁷, where physical impairments hinder disaster preparedness. A critical distinction in the Chinese driving mechanism is the “intergenerational hollowing” effect triggered by hyper-scale migration. While vulnerability in Japan is primarily driven by “demographic maturation,” China’s is driven by “residential decoupling”. The rapid outflow of young labor to new urban districts leaves the elderly “trapped” in decaying, high-risk riverine cores without traditional familial support networks. This “passive retention” mechanism creates a more severe deficit in immediate emergency assistance than in more stable aging societies. However, the Chinese context is further intensified by rapid urbanization-induced intergenerational separation, which effectively “traps” the elderly in aging urban cores without traditional familial support networks.

Differentiated governance strategies based on spatial heterogeneity

Given the distinct spatial heterogeneity of these driving mechanisms, traditional “one-size-fits-all” flood control policies are no longer suitable. Based on the spatial patterns recognized by LISA clustering (Fig. 5) and the driving factors revealed by PCA-MGWR, three distinct targeted recommendations have been proposed:

(1) High-high Concentration Areas (the Middle and Lower Reaches of Yangtze River and Huaihe River Basin): Based on Risk-oriented Planning and Market-oriented Risk Transfer.

This area is the core zone of “wealth exposure” driven by PC-LI (livelihood/wealth). The policy focus should shift from simple engineering defense to the combination of “Risk-informed Urban Planning” and “Catastrophe Insurance” system. At the planning level, high-precision flood risk assessment must be incorporated into the national spatial planning system. For new development in high-risk areas, the principle of “avoidance” should be strictly implemented. For areas with high asset inventory, promote the design of “resilient cities”, such as sponge city facilities and underground water storage systems, to alleviate the pressure of urban flooding. This governance shift requires a deep understanding of the spatial antagonism between urban morphology and ecological infrastructure to mitigate environmental sensitivities in high-density areas²². Furthermore, the effectiveness of these facilities should be ensured through integrated frameworks - combining hydrodynamic models and GIS - to achieve the precise spatial optimization of sponge structures and functional areas²³. At the institutional level, given the limitations of physical defense and the systemic risks of asset exposure, a catastrophe insurance system is needed urgently. The government should encourage high-income families and enterprises to purchase flood insurance through fiscal subsidies, and utilize market mechanisms to socialize and disperse the huge potential losses caused by “excessive accumulation”, preventing poverty due to disasters or regional economic shutdown.

(2) Fragility-driven Regions (Central and Western Mainland): Senior-friendly Transformation and Community Mutual Aid.

In the NW and SW regions, low PC-EL and PC-AC are the dominant vulnerability factors. The core of governance lies in bridging the “cognitive gap” and the “action deficit”. On the physical level, in response to the “in-situ aging” risk revealed by PC-AC, infrastructure construction must shift from “flood prevention” to “elderly care”. It is suggested that “Age-friendly Evacuation Routes” be constructed in old communities and river valley settlements, including the addition of barrier-free access, night light guidance systems and gentle slope designs, to ensure that groups with limited mobility can evacuate independently or semi-independently in emergency situations. At the social level: to address the blind spots in early warning for low-education groups, a multimodal early warning system for the “last mile” should be established, such as local TV stations, mobile social media, dialect broadcasts, and illustrated manuals. More importantly, informal neighborhood assistance should be institutionalized and a “Community-based Mutual Aid” should be established. Before the flood season, grid-based registration was carried out for elderly people living alone and people with disabilities, and young volunteers or neighbors were assigned as their “First Responders” for emergency evacuation.

(3) The Edge of Metropolis: Inclusive Disaster Prevention and Information Barriers Removal.

To address the “institutional exclusion” of PC-FP in large eastern cities, policies need to embrace the principle of “Inclusive Resilience”. It is necessary to break the restrictions of the household registration system in the field of disaster reduction, incorporate the floating population into the community grid-based emergency management system, and ensure the right to open and equal use of emergency shelters. To tackle information imbalance in the rental market, it is essential to enforce a mandatory system for disclosing potential risks. Landlords and real estate agencies should be required to disclose to tenants the historical waterlogging records and flood risk levels of their houses, and provide basic emergency escape guidelines. These measures stimulate the active risk-avoidance ability of the floating population by eliminating information barriers, and thereby fill the “immunity gap” in disaster prevention and mitigation beyond physical exposure risks through institutional inclusiveness.

Conclusions

Based on the data of 2020 Population Census, this study constructed a PCA-MGWR coupled model to reveal the spatial heterogeneity and driving mechanisms of the fragile communities in China’s floodplains. Verification through methodological comparison demonstrates that the PCA-MGWR model can overcome the challenges of collinearity and spatial instability of muti-socio-economic data, supporting more stable tool for large-scale assessment of hazard risk.

There are three core conclusions in this study:

First, there has emerged a notably pronounced situation of the “Paradox of Wealth Exposure” in China. In contrast with conventional environmental justice theory, the high-exposure areas in China are mainly concentrated in the economically developed middle and lower reaches of the Yangtze River and coastal urban agglomerations. This spatial lock-in caused by the water-dependent development model means that flood risk has evolved from a survival crisis into systemic asset risks.

Second, the core motivation of social vulnerability to floods in China is undergoing a shift: from single factor “poverty” to “motility” and “aging”. It can be recognized by our model that non-household registered mobility population in eastern developed areas face potential risk of institutional isolation, whereas the “in-situ aging” in central and eastern plains exacerbates the risk of physical retention of the elderly.

Finally, based on the significant spatial differences in driving mechanisms, this study advocates a shift from “one-size-fits-all” defense to adaptive precision governance. For affluent high-exposure areas like the Yangtze River Basin, where “wealth exposure” dominates, the priority should be the establishment of market-oriented risk-sharing systems, such as mandatory catastrophe insurance, to mitigate systemic economic risks. In contrast, for fragility-driven regions in the Northwest and Southwest, precision governance must focus on bridging “cognitive gaps” and “action deficits” through senior-friendly infrastructure (e.g., age-friendly evacuation routes) and institutionalized community-based mutual aid. Furthermore, for megacities with high population mobility, policies should focus on “Inclusive Resilience” by removing institutional barriers for the floating population, ensuring they have equal access to early warning information and emergency shelters regardless of their household registration status. These illustrated strategies provide a roadmap for transitioning from blanket defenses to region-specific socio-hydrological interventions.

Although this provides a fine-grained spatial perspective, it is still constrained by the static property of census data, which could not reflect the dynamic population distribution when a disaster occurs. Future research will combine multi-source spatio-temporal big data with composite flood models to further explore the dynamic coupling mechanisms of socio-hydrological systems in the context of climate change.

Author contributions

Conceptualization, L.Y., M.J. and Q.Z.; methodology, L.Y.; software, L.Y.; validation, M.J, Y.Z. and Q.Z.; formal analysis, L.Y.; investigation, Q.Z.; resources, Q.Z.; data curation, L.Y.; writing—original draft preparation, L.Y. and Q.Z.; writing—review and editing, Q.Z.; visualization, L.Y.; supervision, M.J.; project administration, M.J.; funding acquisition, M.J. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by Shenzhen Science and Technology Program (Grant Number JCYJ20250604135903005).

Data availability

Floodplain data：https://github.com/fnardi/GFPLAIN. Worldpop data：https://hub.worldpop.org/geodata/listing?id=135. The population data is derived from the seventh national census data of the National Bureau of Statistics（China）: https://www.stats.gov.cn/sj/pcsj/rkpc/d7c/.

Declarations

Competing interests

The authors declare no competing interests.