Integrated Optimization of Emergency Evacuation Routing for Dam Failure-Induced Flooding: A Coupled Flood–Road Network Modeling Approach

Abstract

Floods resulting from dam failures are highly destructive, characterized by intense impact forces, widespread inundation, and rapid flow velocities, all of which pose significant threats to public safety and social stability in downstream regions. To improve evacuation efficiency during such emergencies, it is essential to study flood evacuation route planning. This study aimed to minimize evacuation time and reduce risks to personnel by considering the dynamic evolution of dam-break floods. Using aerial photography from an unmanned aerial vehicle, the downstream road network of a reservoir was mapped. A coupled flood–road network coupling model was then developed by integrating flood propagation data with road network information. This model optimized evacuation route planning by combining the dynamic evolution of flood hazards with real-time road network data. Based on this model, a flood evacuation route planning method was proposed using Dijkstra’s algorithm. This methodology was validated through a case study of the Shanmei Reservoir in Fujian, China. The results demonstrated that the maximum flood level reached 18.65 m near Xiatou Village, and the highest flow velocity was 22.18 m/s near the Shanmei Reservoir. Furthermore, evacuation plans were developed for eight affected locations downstream of the Shanmei Reservoir, with a total of 13 evacuation routes. These strategies and routes resulted in a significant reduction in evacuation time and minimized the risks to evacuees. The life-loss risk was minimized in the evacuation process, and all evacuees were able to reach safe locations. These findings confirmed that the proposed method, which integrated flood dynamics with road network information, ensured the safety and effectiveness of evacuation routes. This approach met the critical needs of emergency management by providing timely and secure evacuation paths in the event of dam failure.

1. Introduction

Reservoirs are a vital form of infrastructure for flood control, irrigation, water supply, and power generation. They play a crucial role in supporting both the livelihoods of individuals and the sustainable development of the national economy in China. According to the 2022 National Water Resources Development Bulletin, China has constructed 95,296 reservoirs with a combined storage capacity of 988.7 billion cubic meters. However, dam failures occur intermittently due to inherent design flaws, extreme weather events, and management errors [1,2]. For instance, on 1 August 2018, the Sheyuegou Reservoir in Xinjiang suffered overtopping damage from a brief but intense rainfall event. The failure of the dam led to 28 fatalities and direct economic losses of CNY 175 million [3]. On 1 May 2020, a section of the Sardoba Reservoir dam in Uzbekistan collapsed. The flooding caused six deaths, environmental damage, and economic losses. Over 100,000 people from Uzbekistan and Kazakhstan were evacuated [4]. Another significant case was the August 1975 collapse of the Banqiao and Shimantan dams in China, which caused the deaths of an estimated approximately 171,000 people and left millions more displaced. This event remains one of the world’s most catastrophic dam failures [5]. In all these cases, timely and well-coordinated emergency evacuation could have significantly reduced fatalities and minimized damage. However, existing disaster prevention and evacuation plans often fail to adapt to rapidly changing flood conditions. They also do not provide timely or effective evacuation guidance for affected populations. Therefore, it is essential to analyze and predict the evolution of floodwaters following dam failures. Additionally, evacuation route planning methods need to be developed to ensure the safe relocation of populations downstream. This would enhance emergency management capabilities in dam breach scenarios, ultimately reducing casualties and property losses.

In the field of flood evacuation planning, numerous studies have employed various computational techniques, such as cellular automata models, and optimization algorithms, like the ant colony optimization (ACO) algorithm [6,7,8,9,10]. For example, Irsyad and Hitoshi studied evacuation route choices for residents in the riverside Kampung areas of Indonesia, analyzing the influence of individual characteristics, route risks, and network structures on evacuation decisions [11]. They optimized evacuation paths using spatial syntax. Rasam et al. combined Geographic Information System (GIS) path analysis with multi-criteria decision analysis to identify safe alternative routes in the Sabah region during floods [12]. Li et al. introduced a method for shortest evacuation path planning based on cellular automata, using a high-precision two-dimensional hydrodynamic model to explore flood characteristics in floodplain areas [13]. Li et al. proposed a Reinforcement Routing optimization model for evacuation paths based on reinforcement learning [14]. This model considers traffic conditions, path risks, and route safety, offering new insights into emergency evacuation path optimization. These approaches have successfully simulated flood dynamics and evacuation strategies, but they often fall short of considering the dynamic nature of real-time flood evolution. The coupling of hydrodynamic models with evacuation routing is essential for improving evacuation efficacy, as it allows for the simulation of real-time changes in both flood conditions and infrastructure disruption, which this study addresses more comprehensively.

Commonly used bionic intelligent algorithms for path optimization include ant colony algorithms, genetic algorithms, and Dijkstra’s algorithm [15,16,17,18,19]. Ozkan et al. analyzed and simulated A* algorithms, genetic algorithms (GAs), and probabilistic roadmap methods (PRMs). They compared the path planning effectiveness of these methods in submerged urban environments to identify the most suitable approach for urban flood disaster path planning [20]. Kumar et al. employed drones for flood disaster monitoring and applied a Firefly Algorithm for path planning, improving rescue efficiency and ensuring real-time information transmission [21]. Zhu et al. introduced an optimal evacuation route scheme based on an improved three-dimensional Dijkstra algorithm, designed for various flood disaster risk levels, ensuring that evacuees reach shelters in the shortest possible time [22]. Li et al. developed a frontier path planning algorithm based on deep reinforcement learning. It incorporates enhanced random reward utilization and heterogeneous reward exploration mechanisms to address simulated rescue path planning under uncertainty and complexity [23].

These methods are typically based on static or simplified road network data. They do not fully account for dynamic flood evolution or road network failure over time. As a result, these methods cannot accurately represent spatio-temporal changes in flood disasters. This leads to evacuation plans that lack timeliness and flexibility. This study hypothesizes that integrating flood simulation data with real-time road network information will enhance the effectiveness of evacuation strategies. Our goal is to develop an optimized flood evacuation model that accounts for flood evolution and dynamic road network failure. Unlike traditional models that rely on static network data or simplified flood scenarios, our model integrates real-time flood dynamics and road network status. This allows it to adapt evacuation strategies on-the-fly, ensuring that paths are not only optimal in theory but also realistic in the face of evolving flood conditions. This study addresses these limitations by considering the uncertainty and time constraints associated with reservoir dam failure floods.

In this study, we employ unmanned aerial vehicle (UAV) measurement technology to obtain high-resolution DEM data. Based on the dynamic evolution of dam breach floods, we integrate flood simulation results with downstream road network data to construct a coupled flood–road network coupling model. This model allows for the analysis of road network failure dynamics and provides a more scientifically robust and practical approach to flood evacuation route planning in dam failure scenarios. The proposed method offers valuable insights for future dam breach emergency management, flood risk mitigation, and emergency evacuation planning in similar sudden water disasters.

2. Material and Methods

The dynamic evolution data of dam breach floods and the road network information model form the foundational data for flood evacuation route planning. The dynamic evolution data serve as constraints for evacuation path planning, while the road network model provides the necessary supporting infrastructure. The integration of these two elements results in the construction of a coupled flood–road network coupling model, which facilitates a more accurate and adaptable approach to evacuation route optimization.

2.1. Flow Control Equation

Reservoir dam failure is a complex fluid dynamic process, often characterized by localized turbulence and significant flow deformation. To accurately capture the three-dimensional flow field of the dam breach flood, this study employs the FLOW-3D hydrodynamic model to simulate and calculate the flood evolution process following a reservoir dam failure [24,25]. Post-processing techniques are then applied to extract relevant flood evolution data. The model utilizes the continuity equation and the momentum equation as the governing equations for fluid motion, as presented in Equations (1) and (2). The RNG laden dam-break flow is shown in Equation (3).

Continuity equation:

$${\displaystyle \frac{\partial \rho }{\partial t}}+\nabla ·(\rho u)=0$$

(1)

Momentum equation:

$$\rho [{\displaystyle \frac{\partial u}{\partial t}}+(u·\nabla )u]=-\nabla p+\nabla ·\tau +\rho f$$

(2)

Governing equations of the RNG k-ε turbulence model:

$$\begin{array}{l}{\displaystyle \frac{\partial k_T}{\partial t}}+{\displaystyle \frac{1}{V_F}}\left(u{A}_x{\displaystyle \frac{\partial k_T}{\partial x}}+v{A}_y{\displaystyle \frac{\partial k_T}{\partial y}}+w{A}_z{\displaystyle \frac{\partial k_T}{\partial z}}\right)=P_T+G_T+DIf{f}_T-{\epsilon }_T\\ {\displaystyle \frac{\partial {\epsilon }_T}{\partial t}}+{\displaystyle \frac{1}{V_F}}\left(u{A}_x{\displaystyle \frac{\partial {\epsilon }_T}{\partial x}}+v{A}_y{\displaystyle \frac{\partial {\epsilon }_T}{\partial y}}+w{A}_z{\displaystyle \frac{\partial {\epsilon }_T}{\partial z}}\right)=\\ {\displaystyle \frac{CDIS1·{\epsilon }_T}{k_T}}·\left(P_T+CDIS3·G_T\right)+DIf{f}_{\epsilon }-CDIS2{\displaystyle \frac{{\epsilon }_T^2}{k_T}}\end{array}$$

(3)

In these equations, ρ is the fluid density; t is time; u is the velocity field; P is pressure; τ is the stress tensor; and f represents external forces. k_T is the turbulent kinetic energy; ε_T is the turbulent dissipation rate; P_T is the production of turbulent kinetic energy due to velocity gradients; G_T is the production of turbulent kinetic energy due to buoyancy; DIff_T is the turbulent diffusion term; CDIS1 is 1.42; CDIS3 is a model constant; CDIS2 is the shear rate function; and DIff_ε is the diffusion term.

2.2. Road Network Information Model Coupled with Flood Information Construction

The issue of personnel evacuation in dam breach accidents is inherently a complex systems engineering problem, influenced by the evolution of flood dynamics and constrained by the road network structure. To describe the relationships between the nodes and edges in the road network more precisely, this study employs graph theory. This effectively analyzes critical issues such as node connectivity and path selection and constructs a road network information model [26,27,28]. By integrating flood evolution data, the road network structure is digitally represented, leading to the development of a coupled flood–road network coupling model.

$$\left\{\begin{array}{l}G=\left(P\left(G\right),E\left(G\right),{\psi }_G,{\phi }_G\right)\\ P\left(G\right)=\left\{p_i\left|i=1,2,\dots ,n\right.\right\}\\ E\left(G\right)=\left\{〈p_i,p_j〉\left|p_i,p_j\in P\right.\right\}\\ {\psi }_G=\left\{\left(t_i,s_i,v_i\right)\left|p_i\in P\right.\right\}\\ {\phi }_G=\left\{e_{ij}\left|〈p_i,p_j〉\in E\right.\right\}\end{array}\right.$$

(4)

$${\psi }_G=\left(\begin{array}{ccc}{t}_1& s_1& v_1\\ t_2& s_2& v_2\\ ⋮& ⋮& ⋮\\ t_n& s_n& v_n\end{array}\right)$$

(5)

In this equation, G represents the road network information model. Specifically, P(G) denotes the set of road network nodes p_i, which correspond to the intersections in the actual road network, where n is the total number of nodes. Additionally, E(G) represents the set of road segments (p_i, p_j) within the network, with the travel cost (e_ij) determined by Equation (6). Furthermore, Ψ_G is the parameter matrix, which includes the flood arrival time (t_i), flood inundation depth (s_i), flood flow velocity (v_i), and other relevant factors at each road network node. Lastly, φ_G is the adjacency matrix for road weights, which characterizes the travel cost and connectivity relationships between nodes.

$$e_{ij}=\left\{\begin{array}{l}{\displaystyle \frac{d_{ij}}{v_0}}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\mathrm{Segments} 〈p_i,p_j〉 \mathrm{connected}, h=0\\ {\displaystyle \frac{d_{ij}}{\eta \left(v_0-3.1h\right)}}\hspace{1em} \mathrm{Segments} 〈p_i,p_j〉 \mathrm{connected}, 0<h\le 0.2\\ {\displaystyle \frac{d_{ij}}{0.5\eta v_0}}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\mathrm{Segments} 〈p_i,p_j〉 \mathrm{connected}, 0.2\le h\le 0.8\\ \inf \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\mathrm{Segments} 〈p_i,p_j〉 \mathrm{disconnected}, h>0.8\end{array}\right.$$

(6)

In this context, d_ij represents the length of the arc 〈p_i, p_j〉, d_ij ≠ 0 with the calculation formula provided in Equation (7). Additionally, v₀ denotes the normal walking speed of individuals unaffected by the flood, while η is the reduction coefficient, set at 0.9 to ensure pedestrian safety. h represents the flood depth (m) of the segment, and inf indicates infinity.

$$d_{ij}={‖p_i-p_j‖}_2=\sqrt{{\left(x_j-x_i\right)}^2+{\left(y_j-y_i\right)}^2}$$

(7)

2.3. Path Planning Method Based on the Coupling of Flood and Road Network Information

To determine the shortest path in the flood–road network information coupling model, the Dijkstra algorithm is employed [29,30]. Starting from a designated origin, the algorithm follows a greedy approach to traverse the network, calculating the closest neighboring node to the current node while excluding previously visited nodes. The process terminates once the destination node is reached. The specific implementation steps are as follows:

Step 1: Construct the flood evolution simulation model for the reservoir dam breach. Using the FLOW-3D v.11.2 flood evolution simulation and FlowSight post-processing, gather essential data such as the downstream flood depth distribution, the flow velocity distribution, the inundation extent, and the arrival time of the flood front.

Step 2: Convert the road network into a mathematical model suitable for the Dijkstra algorithm using graph theory. Integrate the downstream flood evolution data with the transportation network, thereby constructing a coupled road network model that incorporates flood evolution information.

Step 3: Extract the dynamic failure process of the road network from the flood–road network information coupling model. By incorporating the flood front arrival time, divide the road network failure process into several characteristic periods. For each period, reconstruct the road network topology and update the corresponding adjacency matrix.

Step 4: Using the downstream flood evolution data, population distribution, and terrain information, identify the starting point s (affected area) and the destination e (shelter) for path planning. Load the appropriate road network adjacency matrix and apply the Dijkstra algorithm to optimize the shortest path, determining the optimal route to the destination for each time period.

Step 5: Finalize the path planning process and develop an evacuation plan for the affected populations downstream of the dam breach.

The implementation process of the path planning method based on the coupling of flood and road network information is illustrated in Figure 1.

3. Results

3.1. Case Study Application

The Shanmei Reservoir, situated midstream of the Dongxi River, a tributary of the Jinjiang River in Nan’an County, Fujian Province, is a large-scale, multi-functional water conservancy project primarily designed for water supply. It also serves additional purposes, including flood control, power generation, irrigation, and ecological regulation. The controlled catchment area above the dam site spans 1023 km², with a total reservoir capacity of 578 million m³. The basin’s average annual precipitation ranges from 1200 to 1900 mm. The reservoir’s main dam is an earth-rock embankment with a clay core wall, featuring a crest elevation of 105.48 m, a maximum dam height of 75.5 m, a crest width of 8.0 m, and a crest length of 305 m. Since its construction in 1972, the Shanmei Reservoir has experienced 47 small-to-moderate floods with peak flows ranging from 1000 to 2000 m³/s, and 7 major floods with peak flows exceeding this range. Note that all the data used in this study are sourced from the Fujian Provincial Water Resources Management Center.

3.2. Acquisition of Terrain and Road Network Information

This study focuses on the Shanmei Reservoir and its downstream area, utilizing drone-based oblique photogrammetry technology to collect terrain and road network data downstream of the reservoir. Following field surveys, flight zones were defined based on the locations of residential clusters and key facilities in the town and surrounding valleys, covering an area of 3.49 km² and a river stretch of 3.4 km. A drone equipped with a multi-lens high-performance camera was used to capture overlapping multi-angle images of the study area, with an image resolution of 1:500. Due to the steep terrain in the surrounding valleys, the forward overlap was set to 80%, and the side overlap was set to 70%. In the flatter downstream town areas, the forward overlap was adjusted to 70%, and the side overlap was adjusted to 60%. The drone flight height was set at 100 m.

Through a series of processes, including image data acquisition, data preprocessing, aerial triangulation, and 3D product generation, a high-resolution 3D terrain reconstruction of the Shanmei Reservoir and its downstream area was produced. The resulting digital 3D terrain model is shown in Figure 2.

Based on the high-resolution terrain model generated through 3D reconstruction, road network information is extracted and integrated with the raw dataset collected by drones. After resizing the images, the main road lines are extracted, and intersections are identified to preliminarily construct the road network topology for the downstream area of the Shanmei Reservoir. To ensure the model’s completeness and accuracy, a topological analysis and verification of the road network are conducted. The connections between each node and its neighboring nodes are clarified, and detailed connectivity information is compiled into node and road segment tables. The final road network model for the downstream area of the Shanmei Reservoir, as shown in Figure 3, includes 142 nodes and 81 road segments, ensuring both the clarity of node connectivity and the integrity of the network topology.

3.3. Flood Evolution Simulation Analysis

The model is divided into six grid blocks of varying sizes, as shown in Figure 4, labeled Grid 1 through Grid 6, from right to left. Given that turbulent flow is likely to occur as the reservoir water flows downstream, Grid 1, located at the Shanmei Reservoir’s dam site, is locally refined. The refined grid cell size in this area is 2 m × 2 m × 2 m, while the other grids have a cell size of 6 m × 6 m × 6 m. The grid partitioning of the model is illustrated in Figure 4. The inflow boundary is set at the inlet point Y_max of Grid 1, and free outflow boundaries are placed at the surface Y_min of Grid 5 and the surface Y_max of Grid 6. The Z_max surface of all grids is set as a pressure boundary condition, while the connection surfaces between grids are defined with symmetry boundary conditions. The remaining grid boundaries are specified as wall boundary conditions. The boundary conditions for each mesh are set as shown in

Table 1. Boundary conditions for each mesh.

	${\mathit{X}}_{\mathbf{min}}$	${\mathit{X}}_{\mathbf{max}}$	${\mathit{Y}}_{\mathbf{min}}$	${\mathit{Y}}_{\mathbf{max}}$	${\mathit{Z}}_{\mathbf{min}}$	${\mathit{Z}}_{\mathbf{max}}$
Grid 1	W	W	S	Vfr	W	P
Grid 2	S	W	W	S	W	P
Grid 3	S	S	W	W	W	P
Grid 4	S	S	W	W	W	P
Grid 5	W	S	O	S	W	P
Grid 6	W	W	S	O	W	P

.

The simulation models the scenario in which the Shanmei Reservoir is subjected to a 10,000-year design flood, characterized by a peak inflow discharge of 8040 m³/s and a corresponding water level of 102.28 m. The initial reservoir water level is set at 102 m, and the reservoir inflow follows the calibrated design flood hydrograph. Due to the unavailability of detailed bathymetric data for the reservoir area, the upstream boundary condition is represented by a breach outflow hydrograph derived from regulated flood routing calculations. The dam breach is assumed to initiate once the reservoir water level rises to the dam crest, with the breach initiation level set to the design flood level of 102.28 m. The total simulation time for the dam failure process is 2 h. The erosion rate of the dam body is assumed to be positively correlated with the velocity of overtopping flow. Simultaneously, the breach is allowed to expand both laterally and longitudinally. The soils on both sides of the breach undergo intermittent instability and collapse as the breach deepens.

To assess the downstream inundation at various time points following a dam breach, eight monitoring points were strategically positioned based on the distribution of downstream villages for flood depth and flow velocity monitoring. These monitoring points, listed in order, include Shanmei Reservoir Management Office, Yanjiang Cement Company, Sanluo Village, Wancheng Paper Industry, Shinan Village, Jun Village, Xiawei Village, and Shikou Village, corresponding to Monitoring Points 1 through 8. The locations of these monitoring points are illustrated in Figure 5.

Flood depth and flow velocity were recorded at three specific time intervals following the breach: 15 min, 45 min, and 75 min. These data were used to analyze the evolving characteristics of the downstream flood. As shown in Figure 5, the breach progressively widens and deepens due to the erosive force of the water flow, while the floodwaters extend downstream. The distribution of flood depth generally follows a symmetrical pattern, with the deepest water in the center and progressively shallower water on either side. Meanwhile, the position of the maximum flow velocity shifts downstream as time progresses.

The flood depth and flow velocity data from the eight monitoring points are presented in Figure 6. Then, 45 min after the dam breach, all monitoring points detected the presence of water flow, with the flood reaching its peak at 75 min. Monitoring Point 1 (Shanmei Reservoir Management Office) is located at a sharp U-shaped bend in the downstream river channel. Due to the effects of vortex circulation and flood fluctuations, the flow velocity at this point can reach up to 22.18 m/s. Monitoring Point 7 (Xiawei Village) is positioned just before a T-shaped river bifurcation downstream, where the flow exhibits complex behavior, characterized by high kinetic energy at the front. The nearby riverbank experiences a water blockage, leading to a rise in water levels. The maximum flood depth reaches 18.65 m. Monitoring Points 3 (Sanluo Village), 4 (Nan’an Wancheng Paper Industry Co., Ltd., Nanan, China), 5 (Shinan Village), and 8 (Shikou Village) are located near the riverbank at lower elevations. This results in similar characteristics for flood depth and velocity. In contrast, Monitoring Points 2 (Nan’an Yanjiang Cement Co., Ltd., Nanan, China) and 6 (Jun Village), situated at relatively higher elevations, are less impacted by the flood, with the maximum water level not exceeding 3.45 m.

3.4. Flood–Road Network Information Coupling

To facilitate the swift, organized, and safe evacuation of downstream residents, the first step is to identify the affected areas. This involves overlaying the results of the dam breach flood evolution simulation with downstream imagery data, elevation information, and road network details. The affected areas are then delineated by village. These include the Shanmei Reservoir Management Office in Quanzhou, Yanjiang Cement Company, Sanluo Village, Wancheng Paper Industry, Shinan Village, Jun Village, Xiawei Village, and Shikou Village, totaling eight affected units. Based on the latest government statistical reports and local village committee registration data, the estimated number of people at risk is 6091. The population affected by the flood, the estimated arrival time of the flood front, and the available time for safe evacuation for each affected unit are presented in

Table 2. Information on each affected unit and the time available for safe evacuation.

Affected Unit Name	Affected Population (People)	Flood Arrival Time (min)	Available Time for Safe Evacuation (s)
Shanmei Reservoir Management Office	124	21	1260
Nan’an Yanjiang Cement Co., Ltd.	261	43	2580
Nan’an Yanjiang Cement Co., Ltd.	197	20	1200
Nan’an Wancheng Paper Industry Co., Ltd.	207	22	1320
Shinan Village	717	21	1260
Jun Village	1093	37	2220
Xiawei Village	1475	25	1500
Shikou Village	2017	26	1560

.

The selection of relocation sites was comprehensively evaluated based on elevation and road network data around the affected areas, considering two key factors: terrain flatness and proximity to roads. The downstream area of the Shanmei Reservoir is situated in a remote suburban region with limited economic development and few large-scale facilities. Given the sudden and destructive nature of a dam breach flood, an alternative relocation strategy was adopted to move the affected population. The selection of relocation sites took into account factors such as disaster risk, site scale, and emergency support conditions. By integrating remote sensing imagery, population data, and road network information from the downstream area of the Shanmei Reservoir, a GIS spatial analysis approach was employed to identify suitable emergency evacuation and relocation sites. This ensured the safety, accessibility, and spatial suitability of the relocation points. A total of eight affected points and five relocation points were designated, as shown in Figure 7. The affected points—corresponding to the Shanmei Reservoir Management Office, Yanjiang Cement Company, Sanluo Village, Nan’an Wancheng Paper Industry Co., Ltd., Shinan Village, Jun Village, Xiawei Village, and Shikou Village—are sequentially numbered from Affected Point 1 to Affected Point 8.

The post-processed flood progression simulation results can be converted into flood evolution data with coordinates, which are then integrated with the road network model. Information such as affected points and relocation sites is incorporated into the model. For affected and relocation points not directly connected within the road network’s topology, the shortest straight-line segments are used to link them to the road network. This forms the coupled flood–transportation model for the downstream area of the Shanmei Reservoir. Based on this model, the flooding and failure processes of the downstream road network are analyzed.

Ten minutes after the dam breach, the downstream road network remains largely unaffected by the flood. At 20 min post-breach, the 192 Township Road Bridge, located at the U-shaped curve of the river, becomes submerged. At 30 min, the Shixia Bridge, Gangzai Bridge, and Boersi Bridge, near the river’s T-junction, are also submerged. By 50 min, most of the village roads near the riverbank are inundated. These roads are vital for the evacuation of downstream residents during the flood. The failure characteristics of these roads are categorized into four typical time periods: 10, 20, 30, and 50 min after the breach. The flood–transportation coupled models for each time node are shown in Figure 7.

3.5. Emergency Flood Evacuation Path Planning Analysis

From an emergency response perspective, and considering the dynamic evolution of floodwaters following a dam breach, emergency evacuation path planning is conducted using the Dijkstra algorithm. The objectives are to minimize evacuation time and reduce risks to evacuees. In the worst-case scenario simulation, it is assumed that evacuation begins immediately after the breach occurs. The walking speed of evacuees, assuming they are not hindered by floodwaters, is set at 1.6 m/s [31,32]. Emergency evacuation path planning is then carried out based on the coupled flood–road network model for four time intervals: 10, 20, 30, and 50 min post-breach. The planning results are overlaid on topographic maps, as shown in Figure 8, which illustrate the available road networks and evacuation routes for each time period.

Ten minutes after the dam breach, as shown in Figure 8a, the floodwaters have not yet reached the 192 Township Road Bridge, and the downstream road network remains largely unaffected, with most roads still passable. Following the proximity principle, evacuees from the affected points should proceed swiftly toward the nearest shelter points. Twenty minutes after the breach, as shown in Figure 8b, the floodwaters have submerged the 192 Township Road Bridge, rendering Path 1 inoperable. Evacuees at Affected Point 1 who have not yet moved can no longer reach Shelter Point B and must instead take Path 2 to reach Shelter Point A. Thirty minutes after the breach, as shown in Figure 8c, water depths exceeding 0.8 m at Shixia Bridge, Gangzai Bridge, and Bosi Bridge near the T-shaped river junction cause Paths 6 and 11 to fail. Evacuees at Affected Points 5 and 8 can no longer move toward Shelter Point E. Those at Affected Point 5 can take Path 7 to reach Shelter Point A, while evacuees at Affected Point 8 can use Path 12 to reach Shelter Point D. Fifty minutes after the breach, as shown in Figure 8d, most village roads near the riverbank are submerged with water depths exceeding 0.8 m, causing Paths 7 and 12 to become impassable. Evacuees at Affected Point 5 can still reach Shelter Point A via Path 8, and those at Affected Point 8 can continue to Shelter Point D via Path 13. In total, 13 emergency evacuation paths are identified, with details on path lengths and evacuation times provided in

Table 3. Information on emergency flood shelter routes.

Pathway Name	Starting Point	End Point	Distance (m)	Time (s)
Pathway 1	Affected Point 1	Relocation Point B	1290.4	806.5
Pathway 2	Affected Point 1	Relocation Point A	1925.6	1242.3
Pathway 3	Affected Point 2	Relocation Point A	1070	668.8
Pathway 4	Affected Point 3	Relocation Point B	803.8	502.4
Pathway 5	Affected Point 4	Relocation Point D	1037.8	648.6
Pathway 6	Affected Point 5	Relocation Point E	1579.4	987.1
Pathway 7	Affected Point 5	Relocation Point A	1588.2	1018.1
Pathway 8	Affected Point 5	Relocation Point A	1600.8	1026.5
Pathway 9	Affected Point 6	Relocation Point C	1212.6	757.9
Pathway 10	Affected Point 7	Relocation Point E	1225.2	765.8
Pathway 11	Affected Point 8	Relocation Point E	1385.4	865.9
Pathway 12	Affected Point 8	Relocation Point D	1766	1146.8
Pathway 13	Affected Point 8	Relocation Point D	2173.8	1358.6

.

The evacuation plans for the four time intervals are presented in

Table 4. Personnel evacuation plans for different time periods at each disaster site.

Affected Point Name	0–20 min After Dam Break	20–30 min After Dam Break	30–50 min After Dam Break	50 min and After
Affected Point 1	Pathway 1	Pathway 2	Pathway 2	Pathway 2
Affected Point 2	Pathway 3	Pathway 3	Pathway 3	Pathway 3
Affected Point 3	Pathway 4	Pathway 4	Pathway 4	Pathway 4
Affected Point 4	Pathway 5	Pathway 5	Pathway 5	Pathway 5
Affected Point 5	Pathway 6	Pathway 6	Pathway 7	Pathway 8
Affected Point 6	Pathway 9	Pathway 9	Pathway 9	Pathway 9
Affected Point 7	Pathway 10	Pathway 10	Pathway 10	Pathway 10
Affected Point 8	Pathway 11	Pathway 11	Pathway 12	Pathway 13

. The transfer paths are dynamically adjusted based on road network failures to mitigate the impact of flooding on the evacuation process, thereby ensuring the safety of evacuees. By comparing the safe transfer time for each disaster point with the required time to traverse each emergency evacuation path, it is concluded that all disaster victims at the eight affected points can successfully reach their designated shelter points within the safety time window, without any incidents of blockages caused by path failures due to flooding. This dynamic adjustment strategy enhances the reliability of the emergency evacuation path planning, ensuring a high level of safety for evacuees at various stages of flood evolution. As a result, the strategy effectively reduces the risk of flood-related disasters during the evacuation process.

4. Discussion

The dam breach path planning method proposed in this study is based on the coupling of flood and road network information. It effectively captures the spatial distribution characteristics of flood depth, flow velocity, and inundation extent through a three-dimensional simulation. This significantly enhances the accuracy and reliability of dam breach flood simulations in complex terrain scenarios. For instance, our model captures flood dynamics with high precision, including localized variations in flow velocity and inundation depth, unlike previous two-dimensional models. By considering the dynamic nature of flood evolution and the continuous updates to the road network state, the method allows for flexible adjustments to evacuation strategies. It provides valuable decision-making support to emergency management departments and serves as a foundation for optimizing regional infrastructure. Our findings show that by integrating flood simulation data with road network evolution, evacuation routes can be dynamically adjusted, leading to safer, more efficient evacuations, even in rapidly changing flood scenarios. For example, it can guide the addition of emergency routes or the reinforcement of critical road segments’ flood resilience in flood-prone areas, thereby improving the overall resilience of the road network.

In comparable studies, Dong et al. employed a two-dimensional hydrodynamic coupling model to simulate the evolution of dam-break floods. They also incorporated a road network model to develop environmental scenario models for dynamic evacuation routes [10]. Additionally, Yoon et al. used an artificial neural network model to forecast inundation maps and propose an evacuation route selection method, considering evacuation hazards [33]. However, the accuracy of two-dimensional models is limited in complex terrain, resulting in lower precision in spatial distribution features such as flood depth, flow velocity, and inundation extent. Our approach, utilizing three-dimensional simulations, offers superior accuracy, especially in areas with complex topography, and allows for a more precise representation of flood dynamics. Additionally, we incorporate the dynamic failure of the road network, which has often been overlooked in previous studies. This dynamic coupling enables evacuation plans to be adapted in real-time to the evolving flood conditions, mitigating the risk of evacuation route failure due to infrastructure disruptions.

Hu conducted overtopping dam-break simulations using FLOW-3D v11.2, followed by evacuation path planning for downstream populations [34]. Matar et al. developed an optimization model for evacuating people from flood-risk areas at the first sign of a flood, using GIS tools and traffic modeling apps. Their method improves evacuation efficiency, helping to minimize the negative effects of flooding [35]. However, they overlooked the impact of road network status changes during flood evolution, which led to scenarios where evacuation routes failed. In contrast, our simulation dynamically adjusts evacuation routes based on real-time flood evolution data, offering a more robust and flexible solution.

Despite the contributions of the method proposed in this study, there are certain limitations and areas for future development. First, the efficiency of three-dimensional simulations in large-scale scenarios requires improvement. We plan to optimize the efficiency of these simulations using high-performance computing (HPC) to reduce computational time and handle larger-scale scenarios. Additionally, the impact of heterogeneous human behavior (e.g., variations in age and mobility) on evacuation efficiency has not been adequately addressed. In future work, we will integrate an agent-based modeling (ABM) framework to simulate human behavior heterogeneity. This will allow us to account for differences in individual movement speeds and decision-making during evacuation.

Future research could also focus on integrating multi-source data, such as satellite remote sensing and drone monitoring, to enhance the real-time capabilities of the model. By incorporating these data, we can improve the accuracy and timeliness of evacuation route planning during dam failure events. Furthermore, developing multi-hazard coupling models could increase the general applicability of emergency planning. This would enable the model to address various disaster scenarios, not just dam failures. Incorporating multi-objective optimization techniques, such as Pareto front analysis, would allow for a comparison of evacuation strategies across different scenarios. This approach would provide decision-makers with more flexible and effective solutions.

5. Conclusions

This study presents an optimized approach for emergency evacuation route planning in dam failure-induced flooding scenarios, emphasizing the integration of flood dynamics and real-time road network conditions. Using UAV-based photogrammetry and flood data, a coupled flood–road network model was developed, which forms the foundation for evacuation route planning with Dijkstra’s algorithm. This methodology was validated through a case study of the Shanmei Reservoir, offering insights into improving evacuation strategies.

(1)

A coupled flood–road network information coupling model was developed by simulating the dam breach and integrating the downstream flood evolution with the road network. The Dijkstra algorithm was used to identify optimal evacuation paths, resulting in an effective method for dam breach path planning.

(2)

The flood–road network coupling model for the downstream area of Shanmei Reservoir enabled the analysis of the road network’s failure process. Maximum floodwater levels of 18.65 m and flow velocities of up to 22.18 m/s were observed, demonstrating the dynamic nature of flood’s impacts on evacuation routes.

(3)

Four evacuation strategies were developed for eight affected locations in the downstream area of Shanmei Reservoir, identifying 13 emergency flood evacuation paths. This approach improved evacuation planning for flood prevention, mitigation, and disaster relief.

In conclusion, the proposed methodology enhances evacuation route planning by integrating dynamic flood evolution and road network data. The case study results from Shanmei Reservoir show that all evacuees were able to reach designated shelters using optimized routes. This research has important implications for flood risk assessment, hazard mitigation, and emergency planning, suggesting that similar methods can be applied to other regions at risk of dam failure, improving public safety and disaster preparedness.