Modeling Culvert Failure and Community Accessibility Impacts During Hazards

ABSTRACT

Culverts (or road‐stream crossings) are prone to failure during extreme events such as flooding. However, culverts are considered small‐scale infrastructure and have been understudied and overlooked. The impact of culvert failure could be far‐reaching and is especially damaging in rural networks where there is less redundancy in the transport system when compared to a dense urban network. Culvert failure can block road infrastructure and disconnect community members from places of interest such as hospitals. Current work on culverts has focused on improving component level condition ratings. This work allows for prioritization at the component level; however, culvert failure should be modeled as a system event where multiple culverts may fail at one time. Community characteristics are also neglected from current work focusing only on culvert condition ratings. Spatial distributions of communities and places of interest must be integrated into risk analyses to holistically understand culvert impact. This paper introduces a new methodology for simulating culvert failure during hazard events. Particular attention is given toward measuring community impact and using publicly available data to ensure transferability. Additionally, infrastructure‐induced disconnection is coupled with community characteristics to identify areas of concern and propose mitigation actions. Finally, culvert importance metrics are introduced to help identify critical culverts within the system. The method is illustrated on a case study network in Western Massachusetts.

1. Introduction

Within a transportation network, road‐stream crossings are vital in facilitating connectivity between two points. Road stream crossings are structures that carry roads over water bodies. They span perpendicular to roadways to allow for safe passage of water and wildlife thus playing a crucial role in maintaining connectivity between points of interest and mitigating flood risk (Hotchkiss et al. 2007). Throughout this article, road‐stream crossings will be referred to as culverts. Culvert failure can cause disconnection between two points of interest. They fail due to various factors including (but not limited to) inadequate structural design, loss of hydraulic capacity due to excessive water flow, material degradation, and scouring of culvert foundation soil (Watkins 1960; Rigby et al. 2002). These failures can cause disproportionate harm in rural locations where there is little redundancy in the system. For example, a collapsed culvert in Vermont led to a detour of up to 70 miles for some residents (Thurston 2021). Failures are further exacerbated by natural hazards such as hurricanes or intense rain events, which can lead to flooding (Pan et al. 2021; Faturechi and Miller‐Hooks 2015). These events can cause loss of property and disrupt infrastructure networks thus creating challenges in accessing hospitals, schools, and other points of interest (Schweikert et al. 2014b; Serdar et al. 2022). To prevent these disruptions in the transportation network, engineers must ensure that the network components (roads, culverts, and bridges) are adequately designed, routinely inspected, and periodically maintained. To this effect, infrastructure managers must develop effective management, inspection, and simulation tools to ensure operational integrity, particularly in disaster scenarios.

Current methods used by infrastructure managers, engineers, and researchers to understand the impact of network component failures use fragility functions or damage functions (Yang et al. 2009; Liu et al. 2018; D. Kim et al. 2011; Shinozuka et al. 2007; Hwang et al. 2001; El‐Taher and Moore 2008). Fragility functions, which are defined as the conditional probability of attaining or exceeding a specified performance level given a (set of) hazard intensity measure(s) have been extensively used in assessing the performance of bridges after a certain damage state has been exceeded. These methods heavily rely on the availability of data, which is typically available for bridges but not culverts (Wyant et al. 2002; Choi et al. 2004; Gardoni et al. 2003; S. Kim and Shinozuka 2004; Benevolenza and DeRigne 2019). In particular, two types of data are needed to create fragility functions. First, data are needed about the location, material, and condition of the bridge structures. Second, data are needed about historical event frequency and bridge damage. Culvert repositories typically contain information about culvert location and condition; however, historical data linking culvert damage to hazard events of varying intensities are not available. Therefore, the creation of traditional fragility models for culverts is challenging. An approach is needed to couple current culvert infrastructure data to hazards such as floods. Furthermore, it is also crucial to quantify how culvert failures impact communities. The approach should not only measure the community disconnection but also explore the impact disconnection has on varying demographic subgroups as natural disasters have been shown to exacerbate existing inequalities among these groups (Gibbs et al. 2019; Vodopivec and Miller‐Hooks 2019; Benevolenza and DeRigne 2019).

This paper addresses these limitations by proposing a simulation‐based methodology to estimate the effect culvert failure can have on community accessibility. Focus is on the immediate impact and the ability of residents to access hospitals. The proposed method couples culvert condition rating with hazard intensity to simulate hazard events where multiple culverts may fail at once. Previous work has focused on failing culverts one at time which is helpful for criticality analysis but not representative of a hazard scenario (Rees et al. 2018). Moreover, the metric attached to culvert vulnerability should be attached to community access. For example, the Geotechnical Extreme Events Reconnaissance (GEER) team documented numerous culvert failures following Hurricane Maria (Silva‐Tulla et al. 2020). Many of the failures led to road damage and subsequent blockages in the transport network further highlighted the need for a holistic risk assessment. The paper is organized into the following sections: Section 2 reviews existing literature on transportation network modeling, culvert condition rating, and social vulnerability; Section 3 details the development of the methodology for the probabilistic simulation framework designed to calculate the probability of failure of each culvert; Section 4 applies the methodology to a case study with a focus on accessibility, criticality, and social vulnerability; Section 5 concludes the study by summarizing significant insights gathered throughout the research.

2. Literature Review

Culverts are important components of transportation networks. Methods are needed to mathematically model and access transportation network performance. This performance must be a function of culvert condition and possible damage following a hazardous event such as flooding, intense rain, or mud slides. This infrastructure damage must then be translated into societal impact intersecting infrastructure with social vulnerability. This section reviews past work related to (1) transportation network modeling, (2) culvert condition assessment and performance, and (3) social vulnerability and transportation infrastructure.

2.1. Transportation Network Modeling

Transportation networks can be idealized as a collection of links and nodes arranged and designed to achieve a desired function with acceptable performance and reliability (Haynes et al. 2023). These networks enable business operations, sustain supply chains, and ensure access to essential services. The resilience of a system is fundamentally related to its ability to prepare for, withstand, and recover following disruptions (Armitage et al. 2012; Meerow et al. 2016). Transportation networks are particularly vulnerable during natural hazards such as earthquakes, floods, and landslides, since these natural events induce excessive loads on the infrastructure (Khademi et al. 2015; Chang et al. 2000). This damage is either documented in cases of past hazards or simulated for planning purposes. Several studies have then used Monte Carlo methods to simulate damage to the network (Hassan et al. 2022). These methods rely on accurate estimations of component (i.e., link or nodal) failure probability. This failure probability should be attached to the condition of the component. Previous work on the condition of culvert components is further discussed in Section 2.2.

Network performance measures are then found after damage to individual components is measured or estimated. These measures are based on the ability of a transportation network to maintain acceptable levels of service during disruptions is a critical measure of its design and performance (Wakabayashi and Iida 1991; Gu et al. 2020; Soltani‐Sobh et al. 2016). A network is then judged for its ability to consistently operate efficiently and manage disruptions effectively (Khademi et al. 2015; Faturechi and Miller‐Hooks 2015). To systematically quantify network performance, researchers have developed models that simulate potential failure mechanisms and define operational benchmarks. These metrics are often based on connectivity or travel time/distance to places of interest (Logan and Guikema 2020; Boakye et al. 2022; Miller and Baker 2016). Accessibility metrics can be further stratified across different social vulnerability groups or geographic areas to identify areas of concern, propose mitigation actions, and optimize recovery activities (Gardoni and Murphy 2020; Coleman et al. 2024).

Performance metrics based on accessibility to critical services have different computational times and accuracies. Metrics which measure travel time are the most accurate; however, they may be computationally expensive as they should account for factors such as traffic and congestion. Connectivity‐based metrics often neglect traffic and focus on network topology. These metrics can be seen as lower bound estimates of transportation impact. The computational efficiency of connectivity metrics makes them suitable for large network simulation. Connectivity can be conceptualized primarily as a topographical problem within network design (Wakabayashi and Iida 1991). A general undirected network can be defined by nodes or vertexes and links or edges connecting the nodes (Watts and Strogatz 1998). There are two common measures of network connectivity in the literature, the diameter or the efficiency (Latora and Marchiori 2001). The diameter of a network is defined as the average length of the shortest paths between nodes. The efficiency represents the inverse of the shortest path distances and can be used in a disconnected network. Equations for network diameter ($\delta$) and network efficiency ($\eta$) are shown in Equations (1) and (2) where $n$ is the number of nodes in the network, $w_j$ is the normalized weight at node j (e.g., fraction of nodal population), and $d_{ij}$ is the shortest path from node $i$ to node $j$. Performance metrics can also be based on disconnection from critical services. Percolation‐based methods offer a complementary topological approach to network robustness assessment, modeling the progressive loss of network connectivity as nodes or links are probabilistically removed to represent disaster‐induced failures (Dong et al. 2020). This framework has been extended to replace the traditional largest‐connected‐component metric with a robust component (the subset of nodes that retain access to at least one critical facility), thereby grounding topological robustness measurement directly in post‐disaster service accessibility (Dong et al. 2019). All accessibility metrics can be used to create geographic maps predicting areas of large disconnection or inaccessibility. These maps can be used for prioritization and planning.

$$\delta =\frac{1}{n(n-1)}\sum _{i=1}^n\sum _{\begin{array}{c}j=1\\ j\ne i\end{array}}^n{d}_{ij}·w_j$$ (1)

$$\eta =\frac{1}{n(n-1)}\sum _{i=1}^n\sum _{\begin{array}{c}j=1\\ j\ne i\end{array}}^n\frac{w_j}{d_{ij}}$$ (2)

2.2. Culvert Condition and Performance Assessment

Culverts, like bridges, can be modeled as nodal elements in a graph network. Nodal failure may cause disconnection to places of interest impacting community resilience. Accurate information about the probability of nodal failure is dependent on information about culvert location, condition, and likelihood of failure. The availability and format of this information is based on the size of the culvert. The federal highway administration defines any structure over 20 feet which has supports over a depression/obstruction and has a track or passageway for carrying traffic or moving loads as a bridge. Bridges are subject to regular inspections and national ratings (Federal Highway Administration 2023). This information is publicly available in the national bridge inventory. However, many culverts which carry traffic are under 20 feet. This is especially true in rural areas where roads are narrower and support lower traffic. Rural areas are especially vulnerable to disconnection as mentioned in Section 1.

Many culverts, unlike bridges, are not regularly inspected or maintained. Despite this, the impact culvert failure can have on community mobility and accessibility has led to ad hoc culvert condition and performance assessments. Several states have created their own culvert rating systems. For example, Utah has a 5‐point rating scale while Massachusetts has a 9‐point rating scale. Work has been done to normalize these scales based on the descriptions given by state DOTs (Mohammadi et al. 2024). The ratings are based on a multitude of factors including deterioration, culvert size, and surrounding condition (e.g., debris or stream flow). States have also created their own culvert management and data sharing systems since these structures are not in the national bridge inventory (Najafi and Bhattachar 2011; R. Kim 2009).

One such initiative is led by the Vermont Agency of Transportation, which collaborated with researchers, engineers, and scientists to develop the Vermont Culvert Geomorphic Compatibility Screening Tool (R. Kim 2009). The geomorphic incompatibility screening tool was created because existing culvert data largely focused on design guidelines and aquatic organism passage but lacked information geomorphic incompatibility (i.e., how well culverts align with the natural stream process and channel characteristics). Five key variables are used to determine geomorphic compatibility: percent bankfull width, sediment and debris continuity, slope, approach angle, and bank erosion. Culverts are scored on a scale of 0 to 5, with 5 indicating full geomorphic incompatibility between the culvert structure and stream channel.

Culvert failure factors are also complex and variable in nature. The three main culvert failure factors are inadequate flow capacity (i.e., hydraulic factors), prolonged erosion around the culvert (i.e., geomorphic factors), and structural collapse (Yanqing and Chie 2003; Rigby et al. 2002). These failure factors are also further exacerbated by flooding (Chinowsky et al. 2013; Vajjarapu et al. 2019; Quinn et al. 2018; Schweikert et al. 2014a). Culvert assessment protocols that directly consider structural, hydraulic, and geomorphic factors have been developed by previous researchers and state agencies (University of Massachusetts 2024). Designed to mimic the bridge inspection program, inspectors were sent out to collect information on the existing structural, hydraulic, and geomorphic conditions of culvert. This ends‐only approach evaluates culverts from the outside and draws on the foundational work by Hunt (2010), which laid the groundwork for culvert condition evaluation (Hunt et al. 2010). The assessment methodology developed evaluates three mechanisms for culvert failure: structural failure (failure due to deficiencies in the structural integrity of crossings); hydraulic failure (inability to pass enough water to avoid structural damage, washouts, or water overtopping the road); and geomorphic failure (loss or substantial damage due to blockage with woody debris, sediment blockage, or scour). The scoring/ rating system for evaluating the risk of culvert failure is designed to provide a range from 0, indicating low risk, to 1, indicating high risk for each failure mechanism. The overall risk of failure (R) is a combination of the three component scores (structural, hydraulic, and geomorphic risk) discussed into a single score representing the overall potential for crossing failure. The overall risk of failure is defined as the greatest of the structural (S), geomorphic (G) and hydraulic (H) scores and ranges from 0 to 1 as shown in Equation (3) below. Although this protocol was developed and deployed in Massachusetts (Rees et al. 2018), it can easily be adapted to other states.

$$R=\max (S,G,H)$$ (3)

It is important to note that these risk of failure scores are not matched to flood (or any other hazard) events. Therefore, the score should not be confused with a probability of failure. The method for using culvert condition data in regional risk simulation is discussed in Section 3. To the authors' knowledge, these scores have not been used to conduct regional risk assessments in the past.

2.3. Social Vulnerability and Transportation Infrastructure

Disasters do not affect all communities equally. The capacity to withstand and recover from infrastructure disruptions is shaped in part by the social and economic characteristics of a community's residents. Vulnerable populations, including elderly individuals, low‐income households, and racial and ethnic minority groups, consistently face amplified consequences from natural disaster events (Cutter et al. 2003; Benevolenza and DeRigne 2019). The concept of social vulnerability captures these differential susceptibilities: communities with higher concentrations of vulnerable residents face greater obstacles in preparing for, responding to, and recovering from hazard events (Cutter et al. 2003). Among vulnerable groups, the elderly population is of particular concern in the context of transportation network disruption. Older adults are disproportionately likely to have preexisting health conditions (including cardiovascular disease, diabetes, and chronic respiratory illness) that create an urgent and recurring need for access to medical facilities (Benevolenza and DeRigne 2019; Waddell et al. 2021). For such individuals, a loss of road network connectivity during or immediately following a hazard event can carry life‐threatening consequences that extend well beyond those experienced by the broader population. This is especially concerning as climate change induced sea‐level rise has been shown increase isolation risk and disconnect vulnerable individuals from critical services (Logan et al. 2023).

Recognizing that infrastructure disruptions impose unequal burdens, researchers and public agencies have developed quantitative tools to identify and characterize socially vulnerable populations at the geographic scale relevant to planning and emergency management. Flanagan et al. (2011) introduced the Centers for Disease Control and Prevention Social Vulnerability Index (CDC SVI), which aggregates 15 census‐derived indicators across four thematic domains to produce tract‐level scores widely used in disaster preparedness and response planning. States have also created their own tools. For example, the Massachusetts Department of Transportation (MassDOT) has institutionalized these considerations through an Environmental Justice Policy (Massachusetts Department of Transportation 2019) that directs transportation investment and planning decisions to explicitly account for the needs of communities defined by high minority populations or below‐median household incomes.

A growing body of research has examined whether socially vulnerable populations experience systematically worse outcomes from transportation infrastructure failures. Okte et al. (2024) introduced a methodology to quantify the impact of pavement condition on Environmental Justice (EJ) communities in Massachusetts, finding that EJ communities were approximately twice as likely to reside near roads in poor condition and to bear a disproportionately greater burden in vehicle operating costs relative to non‐EJ communities. Genc et al. (2025) developed a network‐based framework for evaluating access to critical and essential services in coastal communities subject to natural hazards and demonstrated that social vulnerability characteristics of the surrounding population meaningfully shape estimates of service access risk. Anderson et al. (2022) similarly identified equitable access to community services, rather than aggregate network functionality alone, as the more policy‐relevant measure of transportation system performance during disruption events. These studies collectively support the view that the consequences of infrastructure failure are not uniformly distributed across a population and that equity considerations are an important dimension of quantitative risk assessment.

However, the relationship between social vulnerability and transportation infrastructure is not universally consistent. For example, previous research found no statistical difference between pavement conditions in vulnerable versus non‐vulnerable groups (Hendricks 2017). This finding carries important methodological implications for how social vulnerability is incorporated into risk analysis. When a systemic relationship between vulnerability indicators and infrastructure outcomes exists, composite indices such as the CDC SVI or MA EJ communities can provide an efficient basis for directing mitigation resources toward the most affected groups. When no such systematic relationship is identified, however, relying on composite indices alone may obscure meaningful spatial variation in individual vulnerability factors. In such cases, a disaggregated analysis that maps individual demographic indicators against estimated connectivity outcomes provides a more actionable basis for identifying specific communities of concern.

3. Proposed Methodology

The goal of this study is to evaluate the impact of culvert failure on a community's transportation network. Previous work has been dedicated to documenting culvert locations and creating a rating system to better characterize culvert condition as detailed in Section 2.2. These databases are helpful; however, they do not provide information to communities about failure impacts. Therefore, this study focuses on the use of existing culvert databases (with culvert condition ratings) to simulate different hazard events, predict community consequences, and identify critical culverts which contribute to the defined community consequences. These consequences will be geospatial and a major goal is to identify locations which may have disproportionate consequences or failure. Social vulnerability factors can then be used to recommend mitigation actions. Achieving the study goal requires (1) the creation of a geospatial network, (2) a method for simulating hazard occurrences, and (3) selecting performance metrics for both the community and infrastructure. Each of these parts is discussed below.

3.1. Creation of Geospatial Network

The first step is to create a geospatial network containing all information needed to model the transport network and surrounding communities. It is important to define the geographic extent or resolution of the network. This definition should be based on the community of interest. Infrastructure on which the community relies on for daily functioning should be modeled. The resolution of the infrastructure modeling can be chosen based many factors including spatial extent, computational efficiency, and accuracy (Nocera and Gardoni 2022). The network can be mathematically modeled as a graph system with nodes and edges. Nodes represent places of interest such as hospitals, shelters, schools, and population centers. Nodes would also represent culverts. Edges would be defined based on road geometry. The length of edges would be the distance of the roads. The boundary of the network also needs to be defined. This is often based on analysis area (e.g., watershed, county, city, etc.). However, there are often places of interest just outside of the network which residents may be able to reach. For example, an additional hospital may be located right outside of the city boundaries that residents can easily reach if they can make it to a highway exit in the event of a hazard. Therefore, exit nodes are often added to the network as additional places of interest (Boakye et al. 2022). This is especially salient in small or rural communities where there are a fewer number of destination points as compared to urban networks. Edge failure would represent the event that road is damaged or out of service while node failure would represent the event that the culvert failed.

The network can be created using publicly available data. Population count data from the U.S. Census Bureau can be used to create origin nodes. The US census provides demographic information at both the census block level which is the smallest geographic unit defined by the census bureau and the census block group level which aggregates information about clusters of census blocks (Bureau 2020). Demographic information such as age, gender, and race is provided at the census block group level, which is especially valuable to assess vulnerability and reveal discrepancies in access between different community groups (Boakye et al. 2022). For example, understanding the age distribution is relevant in planning for areas with higher concentrations of elderly people who might need special transportation arrangements during an emergency. Information about roads and culverts (including condition ratings) can be found from local departments of transportation.

3.2. Simulating Hazard Occurrences

The next step is to simulate hazard occurrences on the geospatial network. Particular attention is paid to moving from the culvert condition rating to the culvert failure probability during different hazards events. Once the probabilities of failure are known, Monte Carlo simulation (MCS) techniques can be used to simulate different events (Dunant et al. 2021). Higher intensity events should generate more network failures when compared to lower intensity events. MCS has been used in a number of studies to understand transportation network failure (Taghizadeh and Mahsuli 2025; Anderson et al. 2022). An important consideration for any MCS analysis is the number of simulations or convergence criteria. Convergence is typically computed by tracking the coefficient of variation (CoV) of some network metric such as the mean percentage of the population without access to the population or the weighted global efficiency as shown in Equation (2). Thresholds below 5% CoV are typically used, consistent with best practices in probabilistic risk (Rubinstein and Kroese 2016), indicating that additional iterations would produce negligible variation in the output. Transportation network modeling methods (discussed in Section 2.1) can be used to create metrics related to community connectivity and mobility impacts. Aggregate measures across events can be used to identify locations of disproportionate and/or consistent disconnection.

Culvert condition ratings are often not attached to any hazard of interest, and a lack of historical data makes traditional fragility assessment difficult. Therefore, we propose to couple hazard intensity (I) with the risk of failure score (R) to get a probability of failure $P_f$ as shown in Equation (4). R is a factor that ranges from 0 to 1 which describes the geomorphic, hydraulic, and structural factors, which could lead to culvert failure during a hazard event as discussed in Section 2.2 while I is a factor ranging from 0.1 to 1 meant to describe the event intensity. The risk of failure score can be updated as culverts get inspected or using data driven prediction models (Mohammadi et al. 2024). As events occur, data on culvert failure can be documented and attached to specific hazard intensities. Given the current lack of data, this method is proposed to represent a spectrum of hazard intensities. It is assumed that the risk stratification remains intact through multiple hazard intensities.

$$P_f=R\times I$$ (4)

It is important to note that this method is proposed as a first step to simulating hazard occurrences and quantifying culvert impacts. Culverts are small‐scale infrastructure which are often neglected from regular maintenance and inspection. Therefore, focus is on immediate impact and identification of large consequences. As more data become available, more complex methods should be developed to better understand the relationship between hazard occurrence, culvert failure probability, and current culvert condition rating systems.

3.3. Performance Metrics

After simulating various hazard instances, it is important to create performance metrics for both the community and infrastructure. Community performance should be related to the effect the hazard has on individual's well‐being. This has been previously studied in past research related to a capability approach (CA) for risk analysis (Murphy and Gardoni 2006; Tabandeh et al. 2018). In a CA to risk analysis, the impact of infrastructure failure is measured as the impact to individuals' capabilities such as the capability to be healthy. Previous research on transportation network failure has analyzed this as the change in efficiency to essential services (Boakye et al. 2022). Beyond a CA framing, access to critical services (measured by changes in travel time/travel behavior) has been proposed as a performance metric for community resilience (Logan and Guikema 2020; Logan et al. 2023; Genc et al. 2025). It is important to note that these studies frame community resilience as a function of the ability of a community to access essential services following disruption, a move away from infrastructure centric metrics which focus on the number of assets damaged or the downtime of individual infrastructure components. Therefore, changes to accessibility are proposed as community performance metrics. The choice of exact performance metric is left up to the decision maker; however, it should generally be a function of disconnection or travel time/distance changes as a function of infrastructure damage.

Infrastructure performance metrics are also salient. They can measure things such as the number of links/nodes disrupted and provide asset managers with a better prediction of rehabilitation needs following different hazard impacts. It should be noted that for large networks, the same number damaged assets may result in a range of community impacts. Therefore, it is helpful to (1) examine the relationship between asset damage and community impact and (2) identify infrastructure components which would cause disproportionate harm to the community. The second step is especially salient if the relationship between the number of components damaged and community impact is nonlinear or exhibits a large degree of variation (i.e., the same number of damaged assets results in a wide range of community impacts). Additionally, criticality rankings of infrastructure components can be helpful for culvert managers to prioritize their limited resources. It should be noted that the current risk stratification system can still lead to a large amount of culverts, which are categorized as high risk (i.e., $R=1$). Past research has identified leave‐one‐out vulnerability as a computationally tractable way to identify component importance (Jenelius et al. 2006; Rees et al. 2018; Gu et al. 2023). Some research has proposed MCS based criticality allowing for the consideration of simultaneous component failure; however, the convergence of such criticality can be computationally intractable for very large networks and has only been tested for small or theoretical networks (Scherb et al. 2017; Agyekum et al. 2026).

4. Case Study

This section presents a case study of the proposed method. The case study was developed using only publicly available data to ensure replicability and transferability. The study uses culverts as the only source of failure to better understand the risk of culvert failure. This study only models the immediate impact phase of the event (where the focus is on life safety). Recovery is neglected from this case study. Social vulnerability factors are used to identify areas of concern and recommend mitigation actions to showcase method's potential benefit to emergency managers and planners. The section is split up into three parts as follows.

4.1. Definition of Study Area and Model Characteristics

The communities within the Deerfield watershed in Franklin County, Massachusetts, are used as the study area to demonstrate the proposed approach. Franklin County was one of the many counties affected by Hurricane Irene. Wind gusts recorded in the county were as high as 80 mph and rainfall of 3–10 in. was recorded between August 28 and August 29, 2011 within the area. The storm left 657,000 homes and businesses flooded and without electricity (Bent et al. 2016; Lunderville 2012). The area also experienced several culvert washouts which led the Franklin County Regional Council of Governments (FRCOG) to obtain grant funding to conduct several culvert assessments (Franklin Regional Council of Governments 2025). This area is chosen for our case study given the vast number of culvert condition assessments available within the county. A total of 783 culverts are used in this study. The distribution of risk of failure scores (R) is shown in Table 1. The study area also showcases the value of citizen science efforts in assessing local risk. Small infrastructure is generally not prioritized for maintenance and inspection despite the community consequences their failure can bring. Citizen science led by local governments can generate very valuable data and promote hazard mitigation (Oti et al. 2019).

TABLE 1.

Culvert risk of failure scores.

R	Frequency	%
0.0	0	0.00
0.1	1	0.13
0.2	46	5.87
0.3	72	9.20
0.4	107	13.67
0.5	104	13.28
0.6	114	14.56
0.7	31	3.96
0.8	92	11.75
0.9	44	5.62
1.0	172	21.97

The study area includes 1483 census blocks with a total population of 42,498. The focus is on accessibility to the one hospital located within the study area. Additionally, four exit nodes were placed along the study area's boundaries, where the watershed and state highway infrastructure interact. There are southern and northern hospitals that we assume residents can reach if the highway exits remain operational. The total road length in the study area is 1351.1 miles. Data for the analysis were obtained from various publicly available sources, including the National Census Bureau, the Homeland Infrastructure Foundation‐Level Data (HIFLD), the Massachusetts Bureau of Geographic Information (MassGIS), and the North Atlantic Aquatic Connectivity Collaborative (NAACC). This information is summarized in Table 2, and the geospatial network created from these sources is displayed in Figure 1a while the geographic distribution of the culvert risk of failure scores is displayed in Figure 1b.

TABLE 2.

Information about Franklin County Deerfield Watershed.

Catchment property	Information	Source
Total population	42,498	National Census Bureau
Number of hospitals	1	HIFLD
Number of exit nodes	4	MassGIS
Number of culverts	783	NAACC
Total length of roads	1351.1 miles	MassGIS

FIGURE 1.

Study network and culverts.

The geographical location of the roads is used to model the transportation infrastructure system. The transportation infrastructure is represented as an undirected graph comprising 5881 nodes and 7039 links. Key elements such as culverts, census blocks, the hospital, shelter locations, and exit nodes are integrated as nodes within this network. This process converts the entire study area into a model to simulate connectivity in flood scenarios, assuming that the entire population in the origin (census blocks) relies on the transport network to reach the hospital or the exit node. This setup allows for tests to be performed in situations where the study area is completely isolated versus when the study area is not. Including exit nodes in the analysis simulates a scenario in which the population in the study area can have access to a point of interest outside of the delineated study area (such as a hospital outside of the study area). In contrast, excluding exit nodes mimics the situation in which people are confined to the area. This approach simulates different real‐life scenarios in extreme cases.

4.2. Results

The methodology detailed in Section 3.2 is then used to simulate different flood events within this case study area. Culverts are considered the only vulnerable component in the network. In reality, floods can temporarily block roads or damage bridges. This is considered outside the scope of this analysis as the focus is on understanding the impact of culvert damage. These additional damages could be added in a more comprehensive analysis. In addition, emphasis is placed on the connection of the community to a hospital. The study area is located in western Massachusetts and there is only one hospital within the study region. Two kinds of analysis are performed: (1) access to the hospital when exit nodes are functional and (2) access to the hospital when exit nodes are not functional. In both analyses, the connectivity of the entire population is studied. MCS were used with a convergence criteria of 5% CoV on the mean disconnection percentage. This analysis was done for various intensity factors (I). For this study, each culvert was modeled independently and the probability of culvert failure was determined using Equation (4). A uniform distribution was used to determine if a culvert failed or survived using its probability of failure.

Culvert failure has a significant impact on community connectivity as displayed in Figure 2. In Figure 2a,b, each point corresponds to a simulation and it is clear that there is a larger degree of spread when the exit nodes are not functional. This is reinforced in Figure 2c,d displaying violin plots in the scenarios where the exit nodes are functional and not functional respectively. In both cases, the uncertainty is smaller on the extremes of I‐values. Regardless of the spread, there are many simulations where disconnection percentage is above 30%. This percentage is concerning, as population disconnection is a lower bound of the impact of culvert failure. Much of the population might be connected, but experiencing longer travel distances and travel times in flooding events. Another goal of this analysis was to better characterize the network when exit nodes are not operational. The loss of access is larger compared to the case where exit nodes are operational (Figure 2) as expected. Generally, there is large uncertainty when examining the disconnection as a function of failed culverts. This spread means that some culverts may be more critical to the system than others. This is especially clear when looking at the case where the exit nodes are not operational. When the intensity factors are low, the behavior of the network when the exit nodes are operational is similar to the case when the exit nodes are nonoperational as illustrated in Figure 3. In this figure, the mean percentage of the population is plotted as function of the hazard intensity factor $I$. The bounds in Figure 3 represent the 15th and 85th percentiles from the MCS. These results also show that network metrics (i.e., the number of culverts (nodes) damaged) is not a representative measure of hazard impact given the large degree of variability. This is concerning for validation efforts as information is often reported as the number of infrastructure components damaged.

FIGURE 2.

Comparison of mean disconnection for the network with operational versus nonoperational exit nodes.

FIGURE 3.

Percentage of the population disconnected from the hospital.

The percent of the population disconnected from the hospital is helpful to understand the scale of culvert failure impact. However, it is also important to better understand the geographic distribution of disconnection. This can help aid mitigation action. Additionally, areas which have high disconnection from the hospital with lower hazard intensity factors are especially concerning. This is examined through Figure 4. In both cases, population centers which are close to the hospital have little disconnection from the network. The existence of the exit nodes helps the populations on the bottom left parts of the study area. When the exit nodes are not operational, these areas also experience high disconnection probabilities as evidenced by Figure 4b. In both scenarios, the northern part of the study area has higher disconnection and would be a place of concern. It is important to note that the population is not evenly distributed throughout the study area. Instead the places with larger population centers are located closer to the hospital and experience less disconnection. The areas of concern tend to have lower population and are geographically sparse. Therefore, a targeted geographic mitigation would be difficult. To address this, we include demographic factors to better target mitigation actions and identify areas and populations of concern. This is discussed in Section 4.3.

FIGURE 4.

Spatial distribution of population disconnection at multiple intensity factors.

Culvert behavior is further examined given the large amount of disconnection present in the network. The number of culverts which act as articulation nodes is of interest. An articulation node is defined as a node whose removal increases the connected component count in graph (West 2001). This behavior is of interest to better understand if disconnection is possible with singular culvert failure or if clusters of culverts must fail for disconnection to occur (e.g., parallel vs. series behavior). This analysis revealed that 242 of the 783 culverts considered are articulation points (i.e., their failure results in disconnection for at least one population node in the network) in the scenario where the exit nodes are nonoperational. We note that this finding is consistent with previous flooding hazards in this area. According to USGS, Hurricane Irene destroyed more than 960 culverts and isolated 13 communities for days in the Green Mountains of Vermont (just north of our study area) (U.S. Geological Survey 2011). The geographic distribution of culverts which are articulation nodes is displayed in Figure 5. This information is further explored to suggest mitigation actions in Section 4.3.

FIGURE 5.

Distribution of culvert articulation points.

4.3. Mitigation Recommendations

This section presents mitigation recommendations based on the results. The method presented here illustrates the potential of the proposed method to help improve community resilience. Community resilience is measured through a community's ability to access essential services. This example focuses on accessing the hospital. Mitigation is focused on identifying demographic and geographic groups of concern. This section also discusses culvert criticality (i.e., which culvert(s) should be better monitored, maintained, or upgraded). Given the large number of culverts in the system and the large degree of uncertainty (evidenced by Figure 3), the simulation results here are not robust enough to inform criticality decisions. Previous research has also shown that identifying the most critical segment may not be possible or relevant for large networks (Alderson et al. 2013). The most relevant information would be the set of critical segments that cause disproportionate disconnection from critical services. However, this is a computational and technically challenging problem that we leave toward future research. Therefore, we focus on culvert graph vulnerability techniques (i.e., fail one culvert at a time measure the impacts). Leave‐one‐out vulnerability is computationally stable and does not exhibit the same noise issues that are present with MCS.

To assess culvert criticality, a leave‐one‐out analysis was performed in which each culvert node was sequentially removed from the network graph and two vulnerability metrics were evaluated. The scenario where the exit nodes are not functional is used (i.e., the hospital is used as the only destination node in the analysis). Ranking A quantifies the absolute change in network efficiency following the removal of culvert $i$:

$$\mathrm{\Delta }{\eta }_i=\left|{\eta }_h-{\eta }_{h,-i}\right|,$$ (5)

$${\eta }_h=\frac{1}{n}\sum _{j=1}^n\frac{w_j}{t_j}$$ (6)

where ${\eta }_h$ is the baseline network efficiency (the hospital is used as the only destination node) as given in Equation (6) and ${\eta }_{h,-i}$ is the network efficiency computed after culvert $i$ is removed. $w_j$ is the fraction of population at node $j$ and $t_j$ is the shortest travel time from node $j$ to the hospital. Ranking B is a composite score combining a disconnection component and a change‐in‐travel‐time component, each assigned an equal weight of 0.5. To ensure comparability, both components are normalized prior to aggregation using min–max scaling:

$${\stackrel{\sim }{x}}_i=\frac{x_i-x_{\min }}{x_{\max }-x_{\min }},$$ (7)

where $x_i$ is the raw score for culvert $i$, and $x_{\min }$ and $x_{\max }$ are the minimum and maximum values observed across all culverts. The composite vulnerability score for Ranking B is then defined as:

$$V_i^B=0.5{\stackrel{\sim }{D}}_i+0.5{\stackrel{\sim }{\mathrm{\Delta }t}}_i,$$ (8)

where ${\stackrel{\sim }{D}}_i$ is the min–max scaled disconnection score and ${\stackrel{\sim }{\mathrm{\Delta }t}}_i$ is the min–max scaled increase in travel time for culvert $i$.

A comparison of the top 20 culverts identified by each ranking reveals substantive differences in which structures are flagged as critical (Figure 6a,b). Ranking A identifies culverts that are spatially concentrated along the same road corridor, reflecting the tendency of efficiency‐based metrics to highlight bottlenecks on heavily used routes where multiple culverts govern access along a shared path. While this behavior makes Ranking A effective at detecting critical road segments, it captures relatively few articulation points. This limitation stems from the scale sensitivity inherent in efficiency‐based measures: the absolute change in efficiency caused by removing an articulation point that serves a small or sparsely connected subgraph may be numerically modest, even when that removal renders portions of the network entirely unreachable. Ranking B, by contrast, produces a more spatially dispersed set of critical culverts and identifies a higher proportion of articulation points. By explicitly incorporating a disconnection term, the composite metric ensures that culverts whose failure severs network connectivity are appropriately elevated in the ranking, independent of the magnitude of the associated efficiency loss. These findings underscore that the choice of vulnerability metric carries substantive implications for infrastructure prioritization. Efficiency‐based metrics are well‐suited to identifying corridor‐level degradation; however, in networks where complete disconnection poses a material risk, composite metrics that assign explicit weight to connectivity loss provide a more robust and comprehensive characterization of culvert criticality.

FIGURE 6.

Culvert criticality based on Ranking A and Ranking B.

To understand the impact of transportation network disruptions on different sociodemographic subgroups within the study area, a spatial analysis is conducted using data at the census block group level. Key vulnerability indicators such as race, income, and the proportion of the population aged above 65 serve as the primary indicators used. These factors were identified based on past research at the intersection of transportation accessibility and vulnerability (Delgado et al. 2023; Burton et al. 2018; Kostopoulou et al. 2024). This initial examination serves as a crucial step toward identifying and addressing disproportionate impacts on these populations after such events.

Income data are obtained from the Census Bureau median household income data set. This data are primarily collected through the American Community Survey (ACS). Within the study area, the median household incomes range from a minimum of $21,830, with an average median income of $66,794. Age and race data were also obtained from the ACS. In this study, the races of interest are White, Black or African American, and Hispanic. The racial composition of the area includes 91.03% White, 4.80% Hispanic, and 1.35% Black or African American. The ages of individuals older than 65 are used to represent older populations, comprising 25.93% of the total population.

A Spearman rank analysis (Spearman 1904) is conducted to understand if there is any systematic relationship between community disconnection from the hospital and social vulnerability factor. This analysis was done for four hazard intensity factors (0.1, 0.5, 0.7, 1) and used in analyzing the 36 census block groups in the study area when the exit nodes are nonoperational (worst scenario). The Spearman rank correlation coefficients calculated for the vulnerability indicators were found to be generally weak in magnitude as shown in Table 3. This finding suggests there is not a systematic bias between hospital disconnection due to culvert failure and the social vulnerability factors checked. Given the weak correlation, we focus on identifying areas with high vulnerability and high disconnection (e.g., a census tract with a large elderly population and large disconnection percentage). These areas are shown using bivariate maps as shown in Figure 7 where darker colors represent areas with higher vulnerability. Both maps are generated for an intensity factor of 0.5. When looking at the maps, we can see that different areas are showing up critical for the different vulnerability factors. For instance, there are areas where the age of the population is high and the population is more likely to be disconnected. Similarly, there are areas where income is low and the population is more likely to be disconnected These areas can be target locations for targeted mitigation action and risk communication prior to flooding events.

TABLE 3.

Correlation coefficients.

I ‐ value	0.1	0.5	0.7	1
Income	0.364	0.443	0.394	0.386
Age > 65	0.147	0.208	0.242	0.239
White	0.151	0.105	0.095	0.058
Black	0.142	0.173	0.195	0.203
Hispanic	0.204	0.167	0.179	0.179

FIGURE 7.

Percentage of the population disconnected from the hospital.

5. Conclusion

This article explores the role of culvert infrastructure on connectivity to essential points of interest post‐disaster. The study aims to propose a transition from the existing methodology which focuses on singular culvert performance rating systems to network simulations. This shift allows for a focus on community consequences and can be used to inform risk analysis and mitigation. The methods developed are conducted using public data to ensure transferability and intended to be used by local officials with vulnerable culvert infrastructure.

The proposed method uses information from existing culvert rating systems, the US census, state agencies, and open homeland foundation data to create a geospatial network. This network is then used to simulate hazard events with different intensity measures. A method is proposed which converts culvert condition ratings to failure probabilities as a function of an intensity factor. More research is needed to calibrate the proposed intensity factors to historical and future flooding events. The current study aims to analyze the effect that different events (and associated culvert failures) can have on the population.

The proposed method is illustrated on case study of the Deerfield Watershed Community. Focus is on connectivity to the one hospital within the study area. The case study only looks at disconnection and does not consider excess travel distance/time within the simulation. The proposed method can easily be adopted to include travel effects; however, the computational time would increase. Exit nodes are placed at highway intersections and evaluated as operational or nonoperational. Findings show that culvert failure can result in a large amount of access loss for community members. As expected, nonoperational exit nodes exacerbate this disconnection. Moreover, we see that the northern part of the study area has high disconnection for lower intensity values. The study also examined correlations between vulnerability indicators and saw low correlations with hospital disconnection. The results from the study show that:

• Culvert condition data could be used to approximate the probability of failure of a culvert through simulations.

• By applying network analysis and graph theory, a transportation network system can be modeled to simulate natural disasters.

• Leave‐one‐out vulnerability analysis can be used to identify critical culverts based on hospital disconnection and/or travel time increase.

• By using social factors and simulation results, vulnerable populations can be identified.

Understanding culvert criticality beyond leave‐one‐out vulnerability is an important future work. The case study shows high variation in population disconnection for a given number of culvert failures. This variation suggests that some culverts will be more critical to the network than others. Identifying these culverts is crucial for local infrastructure managers who have limited time and funds to maintain all of the culverts in the system. Complicating this problem is that identifying a set of critical nodes in a large uncertain graph can be a computational challenge. Despite this, it is an important next step in culvert and community resilience.

Author Contributions

Joshua Govina: formal analysis, investigation, methodology, visualization, writing – original draft, writing – review and editing. Jessica Boakye: conceptualization, investigation, methodology, project administration, resources, visualization, writing – original draft, writing – review and editing. Egemen Okte: conceptualization, investigation, methodology, project administration, writing – review and editing.

Funding

This work was partially funded by the UMass Amherst Department of Civil and Environmental Engineering Brack Research Fund and the UMass Amherst Institute for Diversity Sciences Seed Grant program.

Conflicts of Interest

The authors declare no conflicts of interest.