Urban Flood Severity and Residents’ Participation in Disaster Relief: Evidence from Zhengzhou, China

Abstract

As global climate change intensifies the frequency of extreme weather events, urban flood control and disaster reduction efforts face unprecedented challenges. With the limitations of traditional, top-down emergency management becoming increasingly apparent, many countries are actively incorporating community-based participation into flood risk governance. While research in this area is expanding, the specific impact of urban flood inundation severity on residents’ participation in relief efforts remains significantly underexplored. To address this research gap, this study employs the Community Capitals Framework (CCF) and a Gradient Boosting Decision Tree (GBDT) model to empirically analyze 1322 survey responses from Zhengzhou, China, exploring the non-linear relationship between flood severity and public participation. Our findings are threefold: (1) As the most direct source of residents’ risk perception, flood inundation severity has a significant association with their participation level. (2) This relationship is distinctly non-linear. For instance, inundation severity within a 200 m radius of a resident’s home shows a predominantly negative relation with participation level, with the negative effect lessening at extreme levels of inundation. The distance from inundated areas, conversely, exhibits an “S-shaped” curve. (3) Flood severity exhibits a significant reinforcement interaction with both communication technology levels and government organizational mobilization. This indicates that, during public crises like flash floods, robust information channels and effective organizational support are positively related to residents’ transition from passive to active participation. This study reveals the complex, non-linear associations between flood severity and civic engagement, providing theoretical support and practical insights for optimizing disaster policies and enhancing community resilience within the broader context of urban land management and sustainable development.

1. Introduction

In the context of global warming and rapid urbanization, the frequency of extreme precipitation events induced by climate change has sharply increased, leading to progressively severe urban floods that pose a significant threat to human life and property. Research indicates that since 2010, the incidence of urban flooding in China has risen rapidly, with approximately 76% of Chinese cities experiencing at least one flood in 2015 [1]. Concurrently, Goal 11 of the United Nations’ Sustainable Development Goals calls for a significant reduction in the number of deaths, affected people, and direct economic losses caused by water-related disasters by 2030. It is therefore imperative to study and address the issue of urban flooding, particularly its implications for urban land use and resilient city planning.

To comprehensively strengthen its capacity to respond to future urban inundation risks and safeguard the lives and property of its residents, China has invested substantial financial and human resources in continuously improving its urban flood emergency prevention and control systems, as well as its disaster relief mechanisms. In this context, disaster relief is defined as the organized efforts and actions aimed at supporting residents in responding to urban floods, including both material assistance and organizational support. In addition, China’s traditional flood governance model has been characterized by a technocratic approach, relying primarily on technical experts and engineering solutions, with limited community involvement. However, as urban floods grow more frequent and severe, purely technical solutions may fail to fully address local needs or promote residents’ active participation in disaster response. In China, legal and institutional provisions such as the Emergency Response Law and Law of the People’s Republic of China on Flood Control provide a framework for public participation in disaster governance. Citizens can engage in activities such as information reporting, volunteer work, and local preparedness drills. However, decision-making authority remains largely centralized within government institutions, so while public participation is encouraged operationally, it has limited influence on strategic planning and policy decisions. Consequently, the effectiveness of China’s conventional governance model has become increasingly constrained.

Therefore, it is necessary to transform the traditional paradigm of government as the sole risk management entity and empower the public with the responsibility and obligation to participate in disaster governance. Building on this, scholars have further investigated the role and impact mechanisms of public participation in the disaster governance process. Resident participation in disaster relief can provide broader perspectives for governance, enabling managers to acquire information from affected areas more rapidly after a flood occurs [2]. Furthermore, resident participation helps to foster positive neighborly relations and improve the efficiency of various community-based initiatives [3].

Existing research has identified that resident participation is influenced by a variety of factors, primarily categorized into individual factors, social factors, and governmental management. Gammoh et al. [4] found that individual capabilities and residents’ perceptions of community organizations are key factors influencing their participation in flood prevention efforts. Through a quantitative analysis of social element characteristics across pre-disaster prevention, during-disaster response, and post-disaster reconstruction phases, Cui and Li [5] systematically explored the dynamic mechanisms and resilience characteristics of diverse social elements in full-cycle disaster management, with a focus on the correlational mechanisms between different social factors and residents’ participation in disaster relief. Based on a comparative study of community governance models in six Chinese cities, Liu et al. [6] systematically revealed the interactive mechanisms between government leadership structures and the level of residents’ participation. Drawing upon this body of research, the present study will incorporate individual, social, and governmental factors into its analytical framework to compare their respective influences on residents’ participation.

To address the challenges of urban flooding, scholars domestically and internationally have conducted research on resilient cities based on resilience theory. According to the theory of Chandler and Coaffee [7], the keyword for resilient cities has evolved from “elastic” to “ductile.” A “ductile” resilient city is required to be able to form a new “steady state” through interaction with the disaster after being affected. “Ductile” resilience emphasizes that exposure is a critical evaluation metric for resilience. Therefore, this study selects “residents’ distance to the inundated area” and “the degree of surrounding inundation at the resident’s address” as specific variables to measure the disaster context, thereby characterizing residents’ level of exposure in an urban flood. The disaster context is a crucial external factor influencing residents’ participation in relief efforts, as it determines the level of risk associated with such participation; in turn, the level of risk directly affects residents’ motivation to engage in relief work. This concept of “ductile” resilience is particularly relevant to understanding how communities adapt and re-engage after a shock, moving towards a new equilibrium of proactive participation.

Grounded in the Community Capital Framework (CCF), this study aims to examine the associations between the severity of urban flood inundation and residents’ participation in disaster relief, with particular attention to potential non-linear patterns. Specifically, it explores how spatial exposure to flooding, together with other contextual factors, is related to residents’ engagement in disaster governance, thereby offering insights to support community resilience and participatory approaches to urban flood management. It poses the following research questions: (1) How is the severity of urban flood inundation associated with residents’ participation in disaster relief, relative to other influencing factors? (2) What non-linear patterns exist between flood inundation severity and residents’ participation in disaster relief? (3) How does flood inundation severity interact with communication technology, government organization, and other community capitals to shape residents’ participation? To address these questions, this study will conduct a non-linear analysis using a Gradient Boosting Decision Tree (GBDT) machine learning model, based on data collected from 1322 respondents in Zhengzhou, China. Through the GBDT model, this research will reveal how urban flood inundation severity and individual, social, and governmental factors non-linearly influence the level of residents’ participation, thereby enriching the literature in the field of urban flood governance and contributing to land system science by highlighting the spatial and land use dimensions of community resilience.

2. Literature Review

2.1. The Role of Community Capitals in Disaster Governance

CCF is a theoretical approach widely used in community development research [8]. It conceptualizes a community’s assets as seven interdependent “capitals”: natural, cultural, human, social, political, financial, and built. The CCF is not only useful for multidimensional analysis of a community’s assets and deficits but also serves as a theoretical basis for formulating and evaluating community development strategies [9]. Specifically, it helps identify which capitals are strong and which are deficient, allowing administrators to adjust resource allocation to enhance overall community resilience.

Recent studies have highlighted the significant impact of residents’ participation on the effectiveness of urban flood governance [10]. Previous research on community participation has identified several influencing factors, such as individual capabilities, neighborly relations, governmental capacity, information sharing, and the accessibility of emergency facilities [11,12]. These factors can be effectively categorized within the CCF. This study focuses on five of these capitals:

Human Capital encompasses residents’ professional skills, knowledge levels, and socioeconomic status [13]. These elements are manifestations of an individual’s capacity. In the context of urban flood governance, residents with greater capabilities often exhibit a higher sense of self-efficacy—the belief that they can effectively contribute to flood management tasks. This heightened self-efficacy, in turn, fosters a greater level to participate in relief efforts, whereas residents with lower perceived capabilities may be less inclined to participate.

Social Capital functions to connect various disaster governance forces, creating order and structure in the relief process. It primarily includes trust among residents, networks of mutual aid, and information-sharing capacity. A strong sense of community responsibility significantly predicts altruistic behaviors, such as participating in flood relief work.

Financial Capital, which includes residents’ available funds and economic assets, provides the material security necessary for participating in relief efforts. Individuals with sufficient financial capital are more likely to engage in disaster preparedness, such as purchasing insurance or acquiring emergency supplies (e.g., fire extinguishers, life jackets). This level of preparation provides these residents with a greater sense of security [11].

Political Capital (referred to as government capital in some contexts) includes technical, financial, and policy-related support from governmental bodies, representing the primary organizational force in participatory flood governance. For instance, research by Mees et al. [14] indicates that government policies supporting resident participation in flood prevention can significantly increase their engagement in relief work.

Natural Capital in this study is operationalized as the severity of flood inundation. Previous research on natural capital in this context has primarily focused on spatial prediction of flood inundation trends [15,16] or the exposure of residents and built elements to flooding [17,18]. There has been little focus on how flood severity itself influences the level of residents’ participation. This study aims to fill this gap by analyzing its impact within the five-capital CCF.

2.2. The Mediating Role of Risk Perception

To more clearly elucidate the mechanisms underlying these associations, this study incorporates the concept of risk perception. During sudden-onset disasters, the public typically assesses the level of risk based on their subjective cognition and judgment [19]. This subjective evaluation is defined as risk perception, and it critically influences both everyday behavior and decision-making during a crisis [20,21].

According to Slovic (1987) [19], risk perception is primarily shaped by two factors: risk information (objective conditions such as disaster severity and the capacity of risk management organizations) and sense of control (a subjective perspective rooted in factors like individual ability). Research by Frewer et al. [22] highlights sense of control as a crucial dimension in risk judgment; when residents perceive a disaster as uncontrollable, they tend to exaggerate its severity, leading to significantly heightened risk perception. In our model, the independent variables (the community capitals) contribute to either risk information or sense of control, directly influencing risk perception and, through it, indirectly shaping residents’ psychological states and coping strategies. The severity of urban flood inundation is a unique variable in this model, as it is both objective (a measure of physical exposure) and subjective (perceived in relation to one’s immediate surroundings). This dual nature makes it a particularly important factor influencing risk perception and, ultimately, participation.

Varying levels of risk perception have a direct impact on mental health indicators such as psychological stress [23,24] and negative emotional responses [25]. A high level of risk perception can place residents in a prolonged state of anxiety and panic, compelling them to take action to alleviate this tension. Under such stress, behavioral tendencies shift based on an individual’s psychological state. For example, individuals with lower mental health resilience may adopt avoidance-type coping mechanisms [26], while high anxiety can lead to risk aversion. Conversely, positive emotions can facilitate more rational risk-response strategies. The most fundamental way to alleviate this disaster-induced tension is to mitigate the disaster’s negative impacts. Participating in relief work is one effective way to do this, as it can provide residents with a sense of achievement and other positive emotions, thereby improving their psychological state. Therefore, this study uses the level of residents’ participation as the dependent variable to investigate methods for encouraging engagement in urban flood governance, which can help administrators build more effective emergency management systems.

2.3. The Conceptual Framework

This study utilizes a robust conceptual framework, directly derived from Figure 1, to investigate the multi-faceted factors influencing residents’ participation in disaster relief efforts. The framework posits that built capital (e.g., communication technology), social capital (e.g., neighborhood relationships), human capital (e.g., demographics and capabilities), and political capital (e.g., governmental factors) all exert a direct influence on residents’ participation. For instance, readily available communication technology can directly facilitate coordination, while strong social ties foster a sense of collective responsibility, both directly motivating engagement in relief activities. Similarly, individual capabilities and effective governmental support can directly empower and encourage participation.

Furthermore, natural capital (e.g., flood inundation severity) also directly impacts participation, as the immediate threat can heighten perceived risk and urgency. Crucially, the framework emphasizes an interaction effect between natural capital and the other four capitals. This suggests that the severity of a flood event is not an isolated determinant; rather, its impact on participation is significantly modulated by the existing levels of built, social, human, and political capital. For example, a severe flood might elicit a more organized and effective community response if supported by robust communication infrastructure and strong neighborhood networks. This comprehensive approach highlights that residents’ engagement in relief efforts is a complex interplay of direct influences and critical synergistic effects between environmental factors and various forms of community capital.

The study applies this framework to the unique context of the 2021 Zhengzhou flood, an extreme and rapid-onset urban disaster that overwhelmed official response systems and triggered significant community self-organization. By analyzing this event, the research can test whether flood severity acts as a motivator or a barrier, and how resources like social and human capital moderate this relationship. It challenges simplistic assumptions by investigating the specific conditions under which residents transition from passive victims to active participants in an extreme event. The findings offer critical insights into urban resilience by providing empirical evidence on community behavior in a non-Western context characterized by intense, climate-change-driven hazards.

3. Materials and Methods

3.1. Study Area

Zhengzhou, the study area, is a major city in northern Henan Province with an estimated population of approximately 12.7 million in 2021, providing a representative context for examining urban flood governance and residents’ participation in disaster relief in large, densely populated cities. It has an annual average precipitation of 589–669 mm. Between 17 and 20 July 2021, the city experienced persistent heavy rainfall, with 552 mm recorded from 20:00 on 19 July to 20:00 on 20 July, breaking 60-year hourly and daily rainfall records. This extreme event caused severe urban flooding, resulting in significant casualties and economic losses, and highlighted aspects of a “human disaster,” including fatal incidents in underground spaces such as subway systems. Reports also noted deficiencies in emergency response and organizational coordination. This unprecedented flood provides a critical context for this study to investigate residents’ participation in disaster relief.

3.2. Data Collection

Data for this study were collected through a paper-based questionnaire survey from 28 June to 11 July 2023. To ensure the representativeness of the sample, a multi-stage stratified proportional sampling method was employed across five administrative districts in Zhengzhou: Jinshui, Erqi, Guancheng Hui, Zhongyuan, and Huiji. The sampling process involved two stages: an initial random selection of 26 townships and sub-districts as first-level sampling units, followed by the selection of 38 communities within those units as the second-level sampling units. In each community, we randomly selected approximately 40 participants, who were then invited to participate in a structured face-to-face interview conducted by a total of 8 enumerators (including 6 master students and 2 assistant professors).

The primary data collection instrument was a structured questionnaire developed specifically for this study. The design process involved a comprehensive review of relevant literature and consultation with subject-matter experts to ensure the inclusion of key variables. To establish content validity, the questionnaire was reviewed by a panel of experts in survey methodology and the study’s thematic area. A pilot survey was then conducted with a small sample population to identify ambiguous or unclear items, leading to subsequent revisions. Reliability testing was performed using Cronbach’s alpha, which demonstrated acceptable internal consistency across the main constructs.

A total of 1500 questionnaires were distributed, and after excluding invalid or incomplete responses, 1322 valid questionnaires were retained for analysis (see Figure 2), yielding an effective response rate of 88.1%. The combination of paper-based administration, in-person guidance, and community coordination contributed to the high response rate. The spatial distribution of the survey respondents is detailed in Figure 2. Note that the surveyed sample closely reflects the overall population of Zhengzhou. In terms of age distribution, individuals under 18 account for 16.36%, those aged 18 to 50 represent 78.54%, and those over 50 comprise 5.1%. The gender ratio is 54.92% female to 45.08% male. These age and gender characteristics are highly consistent with Zhengzhou’s population structure.

3.3. Variable Selection

Building on the CCF, this study operationalizes 12 factors as independent variables to predict the level of residents’ participation in disaster relief (the dependent variable). These factors are categorized into five dimensions: built capital (communication technology facilities), human capital (individual factors), social capital, political capital (governmental factors), and natural capital (urban flood inundation severity). The detailed definitions and classifications for each variable are presented below in

Table 1. Definitions and Statistical Values of Variables (N = 1322).

Variable Name	Definition	Maximum Value	Minimum Value	Mean Value	Median Value
Level of residents’ participation	The degree to which you participated in disaster relief efforts during the Zhengzhou heavy rainfall (1–5)	5	1	3.37	3.38
Communication Technology	Familiarity with the use of communication technology infrastructure (1–10)	10	1.25	5.21	4
Government Organizational Capacity	Satisfaction with the government’s organizational efforts (1–10)	10	0	5.57	5
Flooded Area Distance	Distance from the flooded area (m)	1648	0	281.21	184
Surrounding Flooding Extent	Percentage of the flooded area within a 200 m radius (%)	42.37	0	2.07	0.22
Age	Age (1 = 0–18 years, 2 = 18–30 years, 3 = 30–40 years, 4 = 40–50 years, 5 = 50–65 years, 6 = 65 years and above)	6	1	-	3
Education Level	Education level (1 = No education, 2 = Middle school or below, 3 = High school, 4 = College and above)	4	1	-	3
Shelter Accessibility	Accessibility of nearby shelters to residents	1.33	0	0.22	0.15
Gender	Male = 0, Female = 1	-	-	-	-
Neighborhood Relations	How would you rate the mutual aid relationship with nearby residents? (1–5)	5	1	3.27	4
Presence of Elderly at Home	Whether there are elderly people in the household = 1	-	-	-	-
Presence of Children at Home	Whether there are children in the household = 1	-	-	-	-
Occupation	Whether your occupation is related to disaster relief = 1	-	-	-	-

.

The dependent variable for this study is the level of residents’ participation in disaster relief, which refers to the actual involvement of community members in disaster governance activities, including providing information on flood risks, delivering food and essential supplies, transporting affected residents, participating in medical consultations and treatment, and offering psychological counseling. This was operationalized using a Likert scale ranging from 1 (strongly disagree) to 5 (strongly agree), assessing the level of residents’ intended engagement across seven types of disaster relief activities. The sample’s mean score was 3.37, indicating a moderate level of participation among respondents.

As the primary independent variable of interest, flood inundation severity is operationalized through two indicators, i.e., flooded area distance and surrounding flooding extent. Following the methodology of Ma Wei et al. [27], flooded area distance was measured as the straight-line map distance from a resident’s home to the nearest inundated area. The average distance was 281.21 m (max: 1648 m; median: 184 m), suggesting that virtually all respondents would have visually observed urban flooding during their daily routines. In addition, to assess the surrounding flooding extent, we calculated the percentage of inundated area within a 200 m radius buffer centered on each respondent’s address. The 200 m radius was determined based on the median distance observed across all respondents. We also conducted robustness checks using an alternative 500 m radius, with the results presented in the Appendix A (

Table A1. Ranking of Feature Importance (Flood Extent Within 500 m).

Rank	Feature	Importance	Relative Importance (%)
1	Communication technology	0.1314	33.56
2	Government Organization Capacity	0.0901	23.01
3	Flooded Area Distance	0.0763	19.50
4	Age	0.0224	5.72
5	Education Level	0.0166	4.23
6	Shelter Accessibility	0.0155	3.96
7	Surrounding Flooding Extent (500 m)	0.0137	3.51
8	Gender	0.0080	2.05
9	Neighborhood Relations	0.0080	2.04
10	Elderly	0.0062	1.59
11	Children	0.0026	0.66
12	Occupation	0.0006	0.17

, Figure A1 and Figure A2).

Recognized as a key determinant of participation in social activities [13], individual capacity was measured by two proxies: disaster relief competence and education level. Competence was coded as a binary variable (1 = yes, 0 = no) based on whether respondents had previously engaged in disaster relief-related work. Education level was classified into four ordinal categories (1–4): no formal education, junior high school or below, high school, and college or above.

Representing the immediate social environment, neighborhood relationships can directly influence residents’ participation [2,28,29]. This was measured by assessing the degree of mutual assistance and trust among residents on a 5-point scale (1 = low, 5 = high). The average score of 3.27 suggests a moderate level of neighborhood social capital.

ICT facilities enable cross-temporal and cross-spatial communication. Residents’ ICT usage capability was measured on a 10-point scale, yielding a mean score of 5.21 and a median of 4. This indicates that while most residents can use ICT, a subset of the population may lack sufficient mastery.

Residents’ satisfaction and shelter accessibility were used as proxies for governmental capacity, representing organizational effectiveness and resource allocation, respectively. Shelter accessibility was calculated following Ma Wei et al. [27], using a three-step process: (1) Shelter coordinates from Gaode Maps were processed in ArcGIS 10.8; (2) The Gaode Maps API and Python 3 were used to establish an OD cost matrix and calculate walking times for a 500 m grid. (3) The two-step floating catchment area (2SFCA) method was applied to this data, combined with population data, to derive a final accessibility score for each location.

Recognizing that sociodemographic factors can also influence residents’ participation, this study incorporates gender, age, and household composition (i.e., whether there are elderly people and children in the household) as control variables.

3.4. Methods

3.4.1. Gradient Boosting Decision Tree (GBDT)

Gradient Boosted Decision Trees (GBDT) is a machine learning algorithm that integrates weak classifiers through iterative optimization of the loss function to minimize residuals [30]. In recent years, it has been widely applied due to its strong predictive performance and flexibility. GBDT offers several advantages: (1) it does not rely on the assumption of linear relationships between independent and dependent variables; (2) it effectively captures and explains non-linear relationships, thus broadening its applicability across diverse domains; (3) through stagewise learning, it iteratively adjusts weights and improves accuracy in prediction; and (4) it mitigates the effects of variable interactions and alleviates multicollinearity issues.

Alternative methods were also considered. For instance, Random Forests, another tree-based ensemble method, construct predictions by averaging multiple independent trees rather than sequential boosting. While Random Forests are robust and less susceptible to overfitting, their predictive performance is often slightly inferior to GBDT in scenarios involving strong non-linear dependencies. In addition, as a representative of classical regression approaches, Polynomial Regression extends linear regression by introducing polynomial terms to capture non-linear trends. However, it is vulnerable to overfitting as the polynomial degree increases and struggles to scale effectively when dealing with multiple predictors. Besides, we have also supplemented the analysis with other potential complementary approaches and presented the findings in the Appendix A (

Table A3. Results of Ordered Probit Regression Analysis.

Variable	Coefficient	Standard Error	Z-Score	P > |Z|
Age ***	−0.075	0.029	−2.594	0.009
Gender *	−0.102	0.056	−1.829	0.067
Education Level ***	0.148	0.035	4.260	0.000
Occupation	−0.067	0.107	−0.621	0.534
Elderly	0.042	0.080	0.528	0.598
Children	0.010	0.067	0.144	0.885
Neighborhood Relations	0.017	0.032	0.532	0.595
Shelter Accessibility	0.048	0.130	0.372	0.710
Government Organization Capacity **	0.073	0.028	2.561	0.010
Communication Technology ***	0.300	0.034	8.833	0.000
Flooded Area Distance ***	0.000	0.000	4.485	0.000
Surrounding Flooding Extent ***	−0.041	0.006	−7.280	0.000

Note: */**/*** indicates that the result is statistically significant at the 0.05/0.01/0.001 level, respectively. The overall model fit indices are as follows: Log-Likelihood = −3487; AIC = 7056; BIC = 7269; Cox-Snell R² = 0.421.

and

Table A4. Double Machine Learning Results for Key Variables.

Treatment	ATE	Standard Error	95% CI
Communication Technology	0.127 ***	0.016	[0.096, 0.159]
Government Organization Capacity	0.036 ***	0.014	[0.096, 0.159]
Flooded Area Distance	0.000	0.000	[−0.000, 0.001]
Surrounding Flooding Extent	0.018	0.025	[−0.030, 0.066]

Note: *** indicates that the result is statistically significant at the 0.001 level. ATE represents the Average Treatment Effect, where a statistically significant value indicates the presence of a causal effect. The model was implemented using the Python linear DML package, with the nuisance models specified as RandomForestRegressor. Model fit indices were as follows: R² = 0.909, RMSE = 0.173.

, Figure A3)—including baseline ordered logit/probit models for the 1–5 ordinal outcome, GAMs to capture smooth nonlinearities, and causal forests or DML for robustness checks in causal interpretation.

Furthermore, the GBDT model is well-suited to addressing all three research questions. For Research Question 1, the GBDT model can assess the relative importance of flood inundation severity compared with individual, social, and governmental factors. For Research Question 2, Shapley Additive Explanations derived from the GBDT model can illustrate the marginal effects of flood severity on residents’ participation, revealing patterns and thresholds not captured by linear models. For Research Question 3, interaction effects within the GBDT framework are analyzed to examine how flood severity interacts with other factors to influence participation. This approach ensures that each research question is explicitly addressed and allows for a nuanced understanding of the drivers of residents’ engagement in disaster relief.

This study will use relative importance to rank the importance of each variable, using this method to identify factors that can significantly influence residents’ participation in disaster relief. The calculation equations for GBDT are as follows [3]:

$$F\left(x\right)={\sum }_{m=1}^M{f}_m\left(x\right)={\sum }_{m=1}^M{\beta }_{m}h\left(x;a_m\right)$$

(1)

$$F_m\left(x\right)=F_{m-1}\left(x\right)+\eta ·{\beta }_{m}h\left(x;a_m\right),\eta \in \left(\mathrm{0,1}\right]$$

(2)

In Equations (1) and (2), F(x) represents the predicted value of sample x, which is the weighted sum of m decision tree predictions. represents the prediction value of the m-th decision tree for the sample, which is a classifier based on decision trees, and represents the weight of the m-th decision tree. After iterative optimization of the model, a learning rate η (0 < η ≤ 1) is introduced to determine the contribution of each decision tree and control overfitting problems.

Python code is used to implement the GBDT model, with key parameters including learning rate, maximum tree depth, and number of trees. Using grid search and k-fold cross-validation, the study determines the use of a learning rate of 0.01, a maximum tree depth of 2, and 200 trees for model fitting. The model was implemented with k = 3. Its fit indices were as follows: MAE = 0.328, RMSE = 0.176, and R² = 0.421.

3.4.2. Shapley Additive Explanations (SHAP)

This study employs SHAP to evaluate the contribution values of 12 independent variables in GBDT model fitting results to reflect the impact of different independent variables on residents’ participation. This helps us study the influence of each factor on residents’ participation in disaster relief from perspectives beyond importance and relative importance.

SHAP is a theoretical model based on Shapley values in game theory. This method was first applied to machine learning model prediction result interpretation in 2010 [31]. In 2017, Lundberg and Lee [32] detailed the feasibility of SHAP in machine learning model result interpretation. This method treats each independent variable in every sample as a contributor, calculating each variable’s Shapley value as its contribution value to the prediction result.

$${\Phi }_j={\sum }_{S\subseteq N\backslash \left\{j\right\}}{\displaystyle \frac{\left|S\right|!\left(n-\left|S\right|-1\right)!}{n!}}\left(f\left(S\cup \left\{j\right\}\right)-f\left(S\right)\right)$$

(3)

In Equation (3), n represents the total number of factors, N\{j} represents the set formed by all subsets in set N except j, S is a subset in set N\{j}, f(S) is the output of S in the model, and f(S∪{j}) refers to the output of the union of S and j in the model.

The GBDT model prediction results expressed using SHAP are the sum of SHAP values of each factor plus a baseline value:

$$f\left(x\right)={\Phi }_0+{\sum }_{j=1}^n{\Phi }_j$$

(4)

In this study, φ₀ is the mean of all results output by the sample in the model. Φi quantifies the contribution of independent variables to residents’ participation; if Φi > 0, it indicates that the value of this independent variable has a positive effect on the model’s predicted value, while the opposite indicates that the value of this independent variable will reduce the model’s predicted value.

4. Results

4.1. Relative Importance of Independent Variables

The analysis of variable importance, detailed in

Table 2. Ranking of Feature Importance.

Rank	Feature	Importance	Relative Importance (%)
1	Communication Technology	0.1375	34.37
2	Government Organizational Capacity	0.0921	23.02
3	Flooded Area Distance	0.0465	11.63
4	Surrounding Flooding Extent	0.0439	10.98
5	Age	0.0215	5.40
6	Education Level	0.0154	3.87
7	Shelter Accessibility	0.0145	3.64
8	Gender	0.0090	2.25
9	Neighborhood Relations	0.0087	2.20
10	Presence of Children at Home	0.0061	1.53
11	Presence of Elderly at Home	0.0039	0.98
12	Occupation	0.0004	0.12

, reveals the primary factors shaping residents’ participation in disaster relief. Institutional and technological factors emerge as the most significant predictors, with communication technology (34.37%) and government organizational capacity (23.02%) exerting the strongest influence. This highlights how advanced communication technologies enhance the dissemination of accurate and timely disaster information, while effective grassroots government organizations play a crucial role in mobilizing relief efforts through their proactivity, flexibility, and local knowledge. Consistent with prior work on the importance of mobilization and communication [6], our results extend this literature by demonstrating that their positive influence is magnified under high-risk conditions.

Notably, the direct experience of urban flood inundation severity (22.61%) also ranks as a key determinant of residents’ participation. Among individual attributes, age and education level are the most influential significant factors. Younger or more highly educated residents tend to demonstrate greater self-efficacy and a more comprehensive understanding of disasters, which increases their likelihood of engaging in voluntary participation.

4.2. Non-Linear Relationships Between Independent Variables and Residents’ Paticipation in Disaster Relief

To interpret the non-linear relationships captured by the GBDT model, this study utilizes SHAP values to generate dependence plots for each independent variable against residents’ participation in disaster relief. In these plots, the X-axis represents the value of an independent variable, while the Y-axis represents its corresponding SHAP value, indicating the magnitude and direction of its impact on the model’s prediction for participation level. The trend line reveals the complexity and non-linearity of these relationships.

As shown in Figure 3a, communication technology exhibits a generally positive correlation with the level of residents’ participation, characterized by a distinct threshold effect. When the communication technology score is below 4, its impact on participation level is negative (SHAP value < 0). As the score increases from 4 to 7, the positive impact grows slowly. However, beyond a score of 7, the positive influence is significantly enhanced. A plausible explanation is that advanced communication capabilities expand the scope and speed of information dissemination, allowing residents to rapidly synchronize flood-related information. This can foster a “sense of controllability” over the situation, which in turn significantly promotes their level of engagement in relief efforts.

According to Figure 3b, government organizational work demonstrates a similar positive correlation with participation level. However, its impact follows a more linear-like trend, lacking the sharp inflection point or “threshold” observed with communication technology. This suggests a steady, consistent positive return on perceived governmental capacity. The two indicators for flood severity reveal more complex, non-linear relationships with participation level.

Figure 3c illustrates the effect of the degree of inundation surrounding a resident’s address. The plot shows a predominantly negative relationship. While there is a marginal positive effect at near-zero levels of inundation, it rapidly decays. As the inundation percentage increases beyond 1%, its impact becomes consistently negative, indicating that greater direct exposure discourages participation. This negative effect is most potent when the surrounding area is approximately 7.5% inundated (where the SHAP value is lowest). Interestingly, after 10% inundation, the negative impact begins to lessen, suggesting a potential adaptation or “resignation” effect at extreme levels of saturation.

The effect of distance from inundated areas, as shown in Figure 3d, follows a distinct “S-shaped” curve that can be interpreted through three psychological-geographical zones. In the immediate impact zone (0–300 m), the level of residents’ participation decreases with distance. Being within the “affected area” generates the strongest sense of urgency, making their engagement highly sensitive to their proximity to the hazard. A positive effect emerges and then plateaus within the anxiety buffer zone (300–1000 m), where participation level remains relatively stable at a moderate level, possibly because the fear of sudden changes in the flood situation suppresses further increases in participation. Finally, in the safe engagement zone (beyond 1000 m), the participation level rises significantly again. At this distance, residents are far enough to observe the flood situation while maintaining a sense of relative safety, a security that allows them to transition from a state of potential victimhood to one of proactive assistance.

Sociodemographic attributes, particularly age and household composition, also significantly influence residents’ participation in disaster relief. Figure 4a illustrates a distinct non-linear, downward trend between age and participation level. The intention to participate is highest among children and adolescents (0–18 years old) and lowest among the elderly. The impact of age on participation level becomes negative after approximately 30–40 years old, as indicated by the SHAP value falling below zero.

Household composition reveals contrasting effects. As shown in Figure 4e, the presence of elderly members in the household reduces participation level. This is likely attributable to a heightened sense of family responsibility and greater risk aversion concerning high-risk activities like disaster relief. In direct contrast, Figure 4f shows that having children in the household has a slight promoting effect on participation level.

The relationship between shelter accessibility and the effect on residents’ participation in disaster relief presents an “inverted U-shaped” curve. In Figure 4b, when the shelter accessibility is between 0.2 and 0.8, it has a positive impact on residents’ participation, and the impact peaks when it reaches 0.5. Whereas, when the shelter accessibility is less than 0.2 or more than 0.8, it has a negative impact on residents’ participation. This is because residents who live far from shelters often experience a reduced sense of security. In such cases, appropriately increasing the number of shelters as placement points can significantly enhance their perceived safety, thereby encouraging a greater level of participation. Conversely, residents living in close proximity to shelters, while benefiting from a heightened sense of security, may develop an over-reliance on them. This dependence can lead to a preference for remaining sheltered rather than actively engaging in disaster response efforts.

Figure 4c,d, respectively, reflect the non-linear relationships between education level, neighborhood relationships, and residents’ participation in disaster relief. The two are quite similar, showing that as residents’ education level/neighborhood relationships improve, residents’ participation first increases significantly, then remains basically stable, and then further increases. Disaster relief, as collective action that essentially requires individuals to bear time, energy, and even safety risks, benefits from stronger personal capabilities that give residents higher autonomous judgment of risk information and stronger self-efficacy regarding their participation in disaster relief. When residents lack identification with their community, their participation in investing in collective affairs such as disaster relief naturally decreases.

4.3. Interactive Effects Between Variables

To gain a deeper understanding of how urban flood inundation affects residents’ participation in disaster relief, this section explores the interaction effects between flood severity and other key factors. Figure 5 reveals a significant synergistic interaction between communication technology and both flood severity indicators (Figure 5a,b). Specifically, residents’ participation is greatest when high levels of communication technology are combined with conditions of lower risk (i.e., minimal surrounding inundation and greater distance from flooded areas). A parallel synergistic effect is observed for government organizational work (Figure 5c,d), where its positive impact is most pronounced under similar low-risk conditions. This synergy can be understood as the interplay between risk signals and support signals. Flood severity acts as an intuitive risk signal. When this threat diminishes, it works in concert with the positive information and mobilization capacity provided by communication technology and government organizations, resulting in a substantial enhancement of residents’ participation.

5. Discussion

This study set out to examine the associations between the severity of urban flood inundation and residents’ participation in disaster relief, with particular attention to potential non-linear patterns and interaction effects. Our findings enrich a growing body of scholarship on disaster governance that highlights the roles of social capital, governmental capacity, and information systems in shaping civic engagement [4]. Whereas prior studies have often emphasized institutional or social determinants, our results underscore the importance of disaster severity as a contextual factor that operates in complex, non-linear ways. By identifying threshold effects of flood exposure and demonstrating how institutional and technological supports condition these associations, this study extends earlier work that conceptualized participation primarily through social or governmental lenses. The analysis yields three critical insights for theory and practice.

First, this study shows that the physical disaster environment is a key factor associated with residents’ participation, operating through complex psychological associations with risk perception. We found a non-linear relationship between urban flood inundation severity and residents’ engagement. This is not a simple dose–response curve; rather, it is shaped by a perceived “sense of uncontrollability.” On one hand, close proximity to inundated areas initially suppresses the level of participation, as a low sense of control dominates. On the other hand, as inundation severity increases past a certain threshold, residents may undergo a psychological “desensitization” or adaptation, allowing them to regulate their cognitive response, regain a sense of agency, and engage in more rational, proactive behaviors.

This dynamic resonates with Resilience Theory, which conceptualizes community responses as adaptive cycles rather than linear trajectories [7]. Our results suggest that participation can oscillate between suppression, adaptation, and proactive engagement depending on hazard intensity. Within the CCF, flood inundation severity represents natural capital, functioning as the external stressor that interacts with other capitals to shape willingness to participate. Whereas earlier studies often assumed linear exposure–engagement links [18], our findings reveal threshold effects, underscoring the need to integrate psychological adaptation dynamics into resilience-based models of participation.

This dynamic is further nuanced by the spatial layout of emergency facilities. Our findings show that shelter accessibility has a significant “inverted U-shaped” impact on the level of resident participation. Shelters act not only as physical refuges but also as psychological anchors, providing a sense of security that can stimulate positive participatory behavior. However, the “inverted” nature of this relationship suggests a critical threshold: while a baseline of facility access is crucial, an overly dense or saturated layout may inadvertently decrease residents’ motivation to participate by fostering a sense of dependency and weakening individual perceptions of their own agency. Therefore, disaster governance must move beyond purely technical responses to incorporate psychological interventions. Policies should focus on enhancing residents’ sense of controllability through transparent information and community coordination, while urban planning must optimize facility layouts to balance the provision of security with the need to encourage proactive civic engagement.

Second, our research demonstrates that efficient information flow and effective government mobilization act as powerful institutional buffers that not only directly drive participation but also mitigate the “sense of uncontrollability” induced by the disaster. In acute crises like floods, the dual guarantee of information (via communication technology) and organization (via government capacity) is the key pathway for transforming residents from passive bystanders to active participants. Advanced communication technology enhances residents’ ability to obtain and disseminate critical information, while government organizations, as institutionalized intermediaries, facilitate the transition from isolated responses to collective action through early warnings, resource allocation, and collaborative platforms. This finding underscores the importance of built capital (communication technology) and political capital (governmental capacity) in the CCF. From the perspective of Resilience Theory, these factors represent institutional capacities that enhance adaptive governance by sustaining information flows, coordination, and collective action under stress.

Crucially, this study reveals a strong synergistic interaction: the positive effects of information and organization become even more significant in high-risk environments. This indicates that in contexts of severe inundation, robust technological and institutional interventions can effectively counteract the negative coping psychology (e.g., helplessness or avoidance) caused by escalating risks. This highlights a critical policy lever for building a multi-dimensional emergency governance system. Rather than relying on physical defenses alone, community resilience hinges on the synergistic integration of communication systems, institutional arrangements, and social psychological support. Priority should be given to the simultaneous enhancement of communication infrastructure and grassroots organizational capacity to create community governance units with high responsiveness and coordination.

Third, a pivotal and somewhat counter-intuitive finding of this study is that personal capacity (e.g., education) and neighborhood relationships (social capital), while having positive effects, were not the most significant predictors of participation level in this acute disaster context. While higher education can foster a more rational understanding of disaster situations, and strong neighborhood ties can theoretically facilitate synergy, their influence was overshadowed by the immediate disaster environment and institutional factors. When faced with severe, life-threatening risks, the lack of trust and weak emotional connections in communities with poor relationships make it difficult to form the spontaneous synergy needed for effective relief.

This finding does not diminish the long-term importance of human and social capital but rather contextualizes their role. In an acute crisis, the presence of clear information, reliable government organization, and a tangible sense of safety or risk may be more immediate drivers of action than pre-existing social ties or individual skills. This suggests that community governance should adopt a dual-track approach. While long-term efforts must focus on rebuilding community trust and identity through community activities and mutual assistance relations, emergency preparedness must prioritize the institutional and infrastructural factors that are most effective during a crisis. By actively disseminating disaster prevention knowledge and building robust communication and organizational systems, communities can internalize a collective action logic that enhances overall resilience when it is needed most.

6. Conclusions

This study investigated the non-linear relationship between urban flood inundation severity and residents’ participation in emergency response. Integrating the Community Capitals Framework with Risk Perception Theory, we employed a GBDT model to analyze data from 1322 surveyed residents, aiming to enrich flood governance research from a resident-centric perspective. The main conclusions are as follows:

First, urban flood inundation severity is a key determinant of residents’ participation in disaster relief, accounting for 22.61% of its relative importance. Our findings enrich a growing body of scholarship on disaster governance that highlights the roles of social capital, governmental capacity, and information systems in shaping civic engagement [4,6,10]. Whereas prior studies have often emphasized institutional or social determinants, our results underscore the importance of disaster severity.

Second, using two indicators—inundation within a 200 m radius of a residence and distance to the nearest inundated area—the study demonstrates distinctly non-linear effects: proximity to inundation generally reduces participation, though this effect diminishes at extreme levels, while distance exhibits an S-shaped relationship, reflecting a more complex influence pattern. Unlike earlier studies that assumed proportional relationships between exposure and engagement [18], our findings reveal threshold effects and psychological adaptation dynamics, underscoring that exposure does not automatically translate into higher participation. This adds nuance to resilience theory by showing how behavioral responses can shift from suppression to adaptation as risks escalate.

Moreover, synergistic interactions with communication technology and government organizational mobilization enhance residents’ coordination and engagement. Overall, these findings highlight the complex, non-linear dynamics between flood severity and civic participation, offering empirical support for developing community-based resilient governance systems.

Building on the above findings of this study, several practical and policy implications can be drawn. Authorities should tailor engagement strategies to local flood exposure, recognizing that residents’ participation varies with inundation severity. Enhancing real-time communication channels and digital platforms can ensure the timely dissemination of information and improve coordination during emergencies. Strengthening community-level emergency response capacity and volunteer networks further supports effective disaster relief. By treating residents as active contributors rather than passive beneficiaries, policymakers can improve the efficiency of flood response efforts and foster community-based resilience in urban flood management.

Despite these contributions, this study has several limitations that point to avenues for future research. First, the reliance on post-event retrospective survey data may introduce recall bias; future studies should prioritize real-time data collection during disaster events to capture more objective dynamics. Second, the focus on a single city limits generalizability; comparative studies across multiple regions are needed to examine how factors such as regional culture and prior risk experience influence residents’ participation. Third, the operationalization of flood inundation severity could be further refined. Future research using high-precision, real-time data should incorporate multi-dimensional elements—including 2D and 3D spatial structures, flood volume, and morphology—to enhance the scientific rigor and explanatory power of the analysis.