1.1. Microscopic Traffic Flow Simulation for Volcanic Evacuation Planning

Great-scale volcanic eruptions can cause massive ash fall influenced by wind directions and speeds ¹⁴. Compared to other natural disaster types as earthquakes, tsunamis, floods, landslides or hurricanes, volcanic eruptions might reveal a long sequence of complex processes and events, such as massive tephra fall influenced by weather and climate. Living conditions during a volcanic disaster become increasingly difficult over time, as continuous ash fall restricts the usage of (1) water, (2) electricity, (3) communication and (4) transportation networks, simply due to the physical abilities of volcanic ash, including its difficulty of removal from solid and liquid bodies and its increased capillary abilities ¹⁵.

Due to the high density of volcanic ash, it is a threat for the stability of the built infrastructure. Therefore, citizens affected need to evacuate the areas immediately as the life quality decreases rapidly ¹⁶.

The novel guidelines for evacuating more than 300,000 households from Kagoshima City explain how each household may use a private passenger car to leave the city via dedicated evacuation routes ³. This implies that every household has the possibility of using one passenger car to leave the city borders. Nevertheless, we assume that this is challenging at selected times of a typical working day: children are at school, elderly people in nursing homes, and numerous adults working in the central business district in the center of Kagoshima City. The first two groups mentioned can be described as vulnerable road users, as they rely on active travel modes and mobility aids, such as bicycles, walking, walkers, or wheelchairs. Another alternative would be public transportation, although its use may be restricted during volcanic eruption events ^17,18.

This study develops a microscopic traffic simulation testbed to identify efficient volcanic-evacuation strategies for vulnerable groups, aiming to move people beyond Kagoshima’s city limits as quickly as possible. Because rail evacuation is critical at city scale, the Kyushu Shinkansen is explicitly modeled. Its use is examined as a conditional pre-impact evacuation option rather than as a mode assumed to remain available under all volcanic-hazard conditions: the strategy is relevant only if Shinkansen operations can continue under the applicable safety standards during the assumed evacuation window. The choice between transfer via Kagoshima-Chuo Station and direct evacuation toward locations outside the affected area should therefore be determined by the expected hazard extent and the spatial relation between evacuation origins, safe destinations, and the available transport network. The plan-defined receiving municipalities and target points are located outside Kagoshima City but within Kagoshima Prefecture, and the evacuation across prefectural borders is not anticipated in the Massive Tephra Fall Response Plan ³.

In the present case study, schools are used as points of interest for defining evacuation origins. However, these origin locations are exchangeable within the proposed framework and may also represent other facility types that require organized evacuation support, such as nursing homes or comparable institutional settings.

We ask whether organized bus-based evacuation can rapidly evacuate students, teachers, and other vulnerable populations under weekday traffic. Three scenarios are tested: (1) evacuation buses transporting students and teachers from schools to Kagoshima-Chuo Station followed by Shinkansen evacuation to Kumamoto, (2) Scenario 1 with background traffic from detector-based demand to capture congestion effects, and (3) pedestrian-only evacuation from schools to the station. Accordingly, these scenarios are not designed as a direct modal comparison with the plan’s household-based private-car evacuation, but as a complementary strategy for school-based and other vulnerable populations whose immediate access to private cars may be limited during a typical working day.

In this study, modeling refers to the microscopic representation of the connected evacuation process from facility-based origins to intercity rail departure. This includes evacuation-bus movements on the urban road network, pedestrian access and transfer at Kagoshima-Chuo Station, and Shinkansen transport toward Kumamoto within one simulation framework. The novelty of the approach lies in linking these road-, station-, and rail-based components as one operational chain for preemptive volcanic evacuation, rather than examining them as separate transport processes.

The expected contribution is a quantitative assessment of congestion and multimodal interactions in volcanic disasters and practical guidance for improving large-scale urban evacuation plans.

To answer this research question, we use the microscopic traffic simulator SUMO ¹⁹ to model lane-level congestion, bus operations, pedestrian access, and rail transfer at Kagoshima-Chuo Station. Background traffic demand is estimated from detector counts, while evacuation demand is generated from school locations and assigned group sizes. Performance is evaluated using evacuation completion time, person travel and waiting times, and bottleneck formation at key road links and station access points.

1.2. Structure of This Work

After a brief introduction, Section 2 outlines the methodology for building a traffic flow simulation testbed for urban volcanic-evacuation planning. It covers data preprocessing, bus route and passenger estimation, integration of the Shinkansen link between Kagoshima-Chuo and Kumamoto, and background-demand modeling using real traffic count data. The section concludes with three behavioral scenarios, where Scenario 1 serves as the baseline without background traffic. Section 3 presents the simulation results, and Section 4 discusses the findings and future work.

2.1. Preprocessing Open Data Inputs and Micro-Simulation Network Generation

The simulation network is generated from open road data provided by OpenStreetMap (OSM) (Fig. 2(a)) and converted ²⁰ into a routable microscopic traffic model (Fig. 2(b)) following established workflows ^21,22. Building footprints are used to identify school locations as evacuation origins, while Kagoshima-Chuo Station (Fig. 2(b), within yellow circle) is represented as the main transfer hub to high-speed rail. To enable long-distance evacuation, the required Kyushu Shinkansen corridor segment (Fig. 2(c)) is integrated into the city-scale network and aligned so that rail and road infrastructure overlap consistently in one shared coordinate space. In the model, an express Kyushu Shinkansen service departs every 15 minutes from Kagoshima-Chuo Station and travels non-stop to Kumamoto Station in about 46 minutes, providing the backbone for rail-based evacuation.

2.2. Calculating Optimal Evacuation Routes

Evacuation bus routing is derived using a Traveling Salesperson Problem (TSP) approach based on an open routing service implementation ²³. School locations are represented by centroid points derived from their footprints and linked to the closest traversable road segments to ensure routes follow the street network. We first evaluate a larger candidate set of 26 schools within a 2 km radius of Kagoshima-Chuo Station and then reduce the experiment scope to six schools in the north-western area for controlled simulation runs (Fig. 2(b)). The resulting route connecting the six schools and the station has a length of 5,124 m and an estimated travel time of 745 s (Fig. 3). Throughout the study, we distinguish between (a) routes reflecting guideline-based evacuation intentions and (b) alternative routes produced by network-based optimization for comparison. Due to the specific robustness of the used tracemapper ²⁴ method only an aggregated route gets matched to our SUMO simulation network. We solve this issue by exporting our route into a trace (such as a GPS trajectory) and then between densifying to a 10 m spacing, increasing the number of points from 41 to 534 (\(\sim\)5.1 km total). The matching itself is conducted in a more permissive way including acceptance of small misses (to avoid teleporting), a higher search radius per point of 60 m, a candidate point score smoothing procedure to remain on a corridor, and a pre-query of available junction connectors at every intersection.

The workflow for the route instruction generation can be applied in an automated way in case the converted SUMO microsimulation network is not too altered in space and topology from its original OSM road network representation. After a first visual inspection, we found that several segments used in OpenRouteService (ORS) were not accessible for motorists, such as the circle around Kagoshima-Chuo Station at the origin of the route (see yellow number 1 in Fig. 3). Subsequently, we excluded the route segment from 0 to 1 (white numbers, see Fig. 3, accessible by persons), so that evacuation buses start at the white number 1 in Fig. 3 and passengers exit the bus at number 31 and walk to the Shinkansen platform located at 0. Although 26 schools within a 2 km radius of Kagoshima-Chuo Station were initially identified as candidate evacuation destinations, the final experiments focus on 6 schools to establish a controlled and operationally interpretable simulation setting. Serving all 26 schools would require a substantially more complex bus route system and introduce large differences in route length, cycle time, service frequency, and congestion exposure, making the results more difficult to interpret. The reduced setting is also more consistent with the assumed deployment of 20 physical buses, considering not only vehicle availability but also likely limitations in available human drivers during an emergency. Under this fleet constraint, serving all 26 schools would likely require longer and less efficient routes or lower service frequencies. The 6-school configuration is therefore treated as a controlled experimental Scenario rather than as an optimal destination selection.

Nevertheless, the results also suggest a practical direction for extending this framework toward evacuation from all 26 schools. Because the six schools examined here are located relatively close to Kagoshima-Chuo Station, Scenario 3, in which evacuees walk directly to the station, provides a relevant benchmark. Its shortest total evacuation time indicates that walking can be an effective strategy for schools located near the station. Accordingly, under limited bus availability, a hybrid evacuation strategy may be more efficient than assigning buses uniformly across all schools: students and teachers from nearby schools could evacuate on foot, while the available buses are allocated to schools located farther from the station. Such a strategy would preserve the operational simplicity of the current experiment while offering a more scalable basis for future analyses targeting evacuation from the full set of 26 schools.

2.3. Implementing the Base Scenario: Traffic Flow Modelling and Simulation Setup

The base evacuation design assumes a capacity of 50 passengers per evacuation bus and 1,000 passengers per N700 Kyushu Shinkansen (8 cars). A fleet of 20 buses operates as a continuous 24-hour circular service (compare routes in Fig. 4) connecting seven designated bus stops and Kagoshima-Chuo Station.

Bus operations are initialized from the main Kagoshima City bus terminal to reflect realistic deployment conditions, and the effect of this starting location is illustrated by the spatial adjustment of bus demand origins (Fig. 5). The implemented bus circulation and school-to-station evacuation structure are summarized in the experimental routing view (Fig. 4).

2.3.2. Simulating High-Speed Rail Evacuation (Kyushu Shinkansen)

Rail evacuation is implemented through Kagoshima-Chuo Station, using access paths that connect the bus drop-off area near the western station exit to Kyushu Shinkansen Platform 11 (illustrated in Fig. 6; persons are exaggerated in size for visualization purpose).

The represented station access structure follows the observed spatial arrangement of the bus drop-off area, western station entrance, and Shinkansen platform access at Kagoshima-Chuo Station.

To ensure realistic transfer time assumptions, walking durations from two station entry points to the platform area are calibrated using a tracking experiment conducted with Strava ²⁷. The model includes a 5-minute boarding process, meaning each train becomes available shortly before departure, while departures remain fixed at a 15-minute headway. Every 15 minutes an Express Kyushu Shinkansen train is departing from Kagoshima-Chuo Station and requires a 46 minutes non-stop ride to Kumamoto Station. First trial simulations indicate that the resulting walking-time distribution ²⁸ is consistent with the recorded trajectories.

2.3.3. Background Traffic Demand Estimation and Generation

Kagoshima prefecture has 811 cross-sectional traffic count detectors, of which 287 are matching with the generated simulation network (compare to Fig. 2(b)) pictured in Fig. 7. These detectors (pink points in Fig. 7) are video cameras that observe traffic in a specific direction (visualized by blue arrows in Fig. 7) in a cross section of the road space. This means that one edge of our simulation network might correspond with one detector. For each detector monthly 5-minute-resoluted traffic count records are freely available as Open Data ²⁹.

In our approach, we also make use of virtual lane-based induction loop detectors ³⁰ for traffic demand estimation and use the point to lane matching approach presented in Fig. 8. The idea is to find candidate points along road edges with the closest Euclidean distance to the cross-sectional detector locations. Subsequently a query is performed on the edge for obtaining the number of lanes in one direction and the virtual lane-based induction loop detectors are set in parallel to each other in every lane of a specific edge.

The 5-minute counts are then equally distributed among the lanes and in case of vehicle counts lower than the number of lanes, we assign the most central lane with the value lower than the absolute number of lanes of one edge. After matching detector locations to the closest road segments and distributing 5-minute counts across available lanes (Figs. 7 and 8), the resulting demand is converted into passenger-car routes and flows ³¹.

A preliminary simulation run using only passenger cars is conducted to validate whether the generated demand produces a plausible spatial distribution of trip origins and network loading before integrating evacuation agents.

In parallel, geometrical and topographical adjustments of the road network representation are applied to guarantee a consistent and routable network structure, including realistic connectivity and stable link geometry for subsequent multimodal evacuation experiments.

2.4. Scenario Design and Analytical Framework

Person capacities are set to 50 passengers per evacuation bus and 1,000 passengers per N700 Kyushu Shinkansen, with 20 buses operating as a 24-hour circular service from the main Kagoshima City bus terminal and serving seven designated bus stops. It is to mention that we assume full train capacity. Nevertheless, boarding and alighting times are not calibrated yet, and the inclusion of assumed, daily train users (commuters) and tourists will likely be taken into account in future research.

To explore optimal configurations, we define three scenarios using the same microscopic traffic simulation network but varying road-user compositions: Scenario 1 simulates only evacuation buses transporting students and teachers from school locations to Kagoshima-Chuo Shinkansen Station, followed by rail evacuation to Kumamoto Station; Scenario 2 adds background passenger-car demand estimated from 287 map-matched detectors (August 5, 2025, 07:15–13:15 JST) to capture congestion and interaction effects; and Scenario 3 replaces bus access with a pedestrian-only approach where all students and teachers walk from school bus stops to Kagoshima-Chuo, while retaining the passenger-car traffic from Scenario 2.

Across all scenarios, key constraints include the spatial distribution of households and activities (homes, schools, CBD), weekday time-of-day conditions, multimodal interactions (buses, private cars, pedestrians, rail), realistic demand representation, and public transport capacity limits. The experiments assume a pre-eruption, imminent-event setting; therefore, ash-related impacts such as road disruption and additional eruption-induced congestion are not considered.

The simulated pedestrians use the default parameterization ³⁰, without yet distinguishing walking speeds or pedestrian types; vulnerable road-user classes as well as age-related differences will be addressed in future work.

We apply a pedestrian width of 0.48 m, length of 0.21 m ³², minimum gap of 0.25 m, maximum speed of 5 m/s, and desired maximum speed of 1.39 m/s ³³. Pedestrians move on footpaths, sidewalks, and walking areas using SUMO’s striping model ²⁸, which divides lane width into discrete lateral stripes with a default stripe width of 0.64 m ³⁴.

The pedestrian router generates bidirectional walking routes and accounts for delays at nearby red-light crossings. For edges reached during a red phase, travel time is computed as

Finally, Strava-based walking trajectories between two station entry points and the platform area were used to identify pedestrian parameter settings that best reproduce observed transfer times relative to the default configuration. Together with the detector-based background traffic demand and the observed station access structure, this provides empirical grounding for the modeled road-traffic and station-transfer components used in the Scenario experiments.

2.4.2. Scenario 2: Bus-to-Rail Evacuation with Background Traffic

Scenario 2 extends Scenario 1 by adding passenger-car demand estimated from the detector dataset, enabling analysis of congestion and bus-car interaction effects.

Figure 10 shows the same bus-based student and teacher evacuation chain as Scenario 1, but with additional background passenger-car traffic (Fig. 10(b)) derived from detector-based demand estimation. The map highlights how evacuation buses and pedestrians interact with general traffic near Kagoshima-Chuo, while detector locations (Fig. 10(a)) support validation of the generated demand.

2.4.3. Scenario 3: Students and Teachers Walking to Kagoshima-Chuo Station with Background Traffic

Scenario 3 assumes all students and teachers walk from their school locations to Kagoshima-Chuo Station instead of using evacuation buses, while background passenger-car traffic remains active as in Scenario 2.

Figure 11 visualizes a pedestrian-only evacuation access strategy, where students and teachers walk directly from school areas to Kagoshima-Chuo Station along shortest-path routes. The zoomed section emphasizes realistic interactions with traffic lights and vehicles under congested urban conditions.

3.1. Per-Person Access-to-Boarding Durations, Simulation Time and Inter-Boarding Durations

The plot in Fig. 12 compares the total per-person time required to reach Shinkansen boarding across the three scenarios, measured from evacuation-bus boarding in Scenarios 1 and 2 and from school departure on foot in Scenario 3 (sorted by ascending person-level access-to-boarding duration). Scenario 1 (bus-based evacuation without background traffic) yields the shortest and most consistent person-level access-to-boarding durations, with most persons boarding the Shinkansen within roughly 10–30 minutes. Scenario 2 (bus-based evacuation with background passenger-car traffic) shifts the curve upward, indicating longer person-level access-to-boarding durations caused by congestion and interactions with general traffic before Shinkansen boarding, and increases the upper range to about 40 minutes for persons with the longest access-to-boarding durations. Scenario 3 (walking access from school under background traffic) produces the longest person-level access-to-boarding durations overall and the widest spread, with times ranging from about 20 minutes up to nearly 70 minutes, showing that walking-based access substantially increases the time required to reach Shinkansen boarding and leads to stronger inequalities between evacuees with shorter and longer access-to-boarding durations.

The plot in Fig. 13 shows the distribution of total time from initial access to Shinkansen boarding for all simulated persons. Scenario 1 concentrates strongly around shorter person-level access-to-boarding durations (roughly 15 to 25 minutes), indicating a short and stable transfer process when only evacuation buses operate. Scenario 2 shifts the distribution toward longer and more variable person-level access-to-boarding durations (about 20 to more than 35 minutes), reflecting congestion effects when background passenger-car traffic is included. Scenario 3 forms a clearly separate peak at substantially higher person-level access-to-boarding durations (around 40 to 55 minutes) with a long tail up to nearly 70 minutes, demonstrating that pedestrian-only access produces the longest and most variable person-level access-to-boarding durations before Shinkansen boarding.

The plot in Fig. 14 shows how the average time required to reach Shinkansen boarding changes over the course of the simulation. Scenario 1 remains relatively stable at around 18 to 23 minutes, indicating a robust bus-to-rail transfer under uncongested conditions. Scenario 2 fluctuates more strongly and reaches higher peaks (up to \(\sim\)33 minutes), suggesting that background passenger-car traffic introduces time-dependent delays and occasional congestion phases before Shinkansen boarding. Scenario 3 has the highest average person-level access-to-boarding duration, and this duration increases steadily over time (from \(\sim\)17 minutes to >50 minutes), indicating that walking-based access is associated with increasing person-level durations before Shinkansen boarding, likely due to longer access distances and increased interactions at crossings and signals as background traffic demand accumulates.

Accordingly, Figs. 12–14 should be interpreted as results for person-level access-to-Shinkansen-boarding durations, while the system-level evacuation completion time is reported separately in Table 2.

Table 2 summarizes the evacuation completion time for each scenario. It reports (i) the arrival time of the last occupied bus at Kagoshima-Chuo Station (if buses are used), (ii) the arrival time of the final evacuee at the Shinkansen destination (Kumamoto Station), and (iii) the number of passengers on the last non-empty Shinkansen. In Scenario 1, the last bus carrying passengers reaches Kagoshima-Chuo Station after 137.87 min. The final person arrives at Kumamoto Station after 191.40 min, and the last non-empty Shinkansen carries 30 persons. In Scenario 2, background traffic causes congestion and slows down the evacuation. The last occupied bus arrives after 163.30 min, and the final Shinkansen arrival occurs after 211.05 min. In this case, the last non-empty Shinkansen carries 534 persons. In Scenario 3, no buses are operated. Instead, evacuees reach the rail system on foot and continue by Shinkansen. The last person arrives at Kumamoto Station after 149.32 min, with 306 persons on the last non-empty Shinkansen.

This result reveals an important distinction between the person-level and system-level performance metrics used in this study. Figs. 12–14 and Table 3 describe the total per-person time required to reach Shinkansen boarding, measured from evacuation-bus boarding in Scenarios 1 and 2 and from school departure on foot in Scenario 3. For this metric, Scenario 3 performs less favorably than the bus-based scenarios. In contrast, Table 2 reports the overall evacuation completion time, measured by the arrival of the last simulated evacuee at Kumamoto Station. According to this system-level criterion, Scenario 3 has the shortest overall evacuation completion time. This difference occurs because all students and teachers in Scenario 3 can begin moving toward Kagoshima-Chuo Station immediately, whereas Scenarios 1 and 2 depend on finite bus fleet capacity, repeated route circulation, and sequential collection from multiple school locations. Scenario 3 therefore yields longer person-level access-to-boarding durations, but a shorter overall evacuation completion time in the present short-distance school setting.

Additionally, Fig. 15 shows the number of evacuated persons boarding each Shinkansen departure over time for all three simulation scenarios. Each bar corresponds to one scheduled Shinkansen departure (15-minute headway), and the bar height represents the total number of boarded passengers at that departure.

To complement the distribution-based comparison of the total per-person time required to reach Shinkansen boarding, measured from evacuation-bus boarding in Scenarios 1 and 2 and from school departure on foot in Scenario 3, Table 3 summarizes the corresponding descriptive statistics. Scenario 1 shows the shortest durations, with a median of 19.52 min and a standard deviation of 5.43 min. Scenario 2 produces longer and more variable durations under background passenger-car traffic, with a median of 23.75 min and a standard deviation of 8.54 min. Scenario 3 results in the longest person-level access-to-Shinkansen-boarding durations, with a median of 43.72 min and a maximum of 67.35 min, while its standard deviation of 8.26 min remains close to that of Scenario 2. These statistics confirm that Scenario 3 performs less favorably for this person-level duration metric, even though Table 2 shows a shorter overall last-person arrival time (at Kumamoto Station) for Scenario 3. The reported person counts correspond to the valid duration records extracted from the same parser output used for Figs. 12–14; for Scenario 3, this results in 4,822 records. The three missing records correspond to pedestrians that were teleported during the simulation after remaining at the same location for more than 240 s, which can occur in rare cases when local pedestrian interactions on sidewalk polygons cause temporary blocking, particularly near intersections. Since these three records represent only 0.06% of the total Scenario 3 demand, and no complete access-to-boarding duration could be computed for them, they were excluded from the person-level duration statistics. This small number of missing records does not affect the interpretation of the distributional results.

3.2. Walking to Shinkansen Platform and Person Activity Durations

The plot in Fig. 16 shows the distribution of walking times from the station access area to the Shinkansen platform for Scenarios 1 and 2.

Both scenarios exhibit a clear unimodal, near-normal shape, with most pedestrians requiring roughly 160 to 190 seconds to reach the platform and a peak around 170 to 180 seconds. The strong overlap between the two histograms indicates that the platform access walk is highly consistent and only marginally affected by background traffic, suggesting that the main differences in total evacuation duration between Scenarios 1 and 2 are primarily driven by bus travel and road congestion, rather than by station-transfer walking time.

The plot in Fig. 17 presents the total evacuation time by activity phase, aggregated across all persons in person-hours. For a phase \(p\) (e.g., waiting for the bus, riding the bus, walking to the platform, waiting for the Shinkansen), the total person-hours are defined as

The unit of \(\mathit{PH}_p\) is

Example: If \(100\) persons each wait \(10\) minutes (\(600\) s), then

In Scenarios 1 and 2, the largest share is waiting for evacuation buses, indicating that bus availability and boarding processes dominate overall delay. Scenario 2 shows a higher total time budget due to additional congestion effects, mainly reflected in increased bus waiting and in-vehicle travel. In Scenario 3, bus-related phases are absent and the total time is largely driven by walking to the platform and waiting for the Shinkansen, demonstrating that pedestrian-only access shifts delays from road operations to station access and rail waiting.

4.1. Discussion

As a preventive measure for a potential large-scale eruption of Sakurajima Volcano in Kagoshima City ¹, evacuation plans assume that residents leave the city either by private passenger car or with support from organized public transport services such as evacuation buses. This study focuses on the second strategy and demonstrates how bus-based evacuation routes can be designed using open data. We also show that the target group of an evacuation strategy can vary depending on the planning goal and available information; in our case, we focus on evacuating school students and teachers located near Kagoshima-Chuo Station.

The results from the three Scenario experiments provide insights into how vulnerable road users can be evacuated efficiently when two public transport modes with different capacities and headways are combined. In particular, bus operations should be designed in relation to Shinkansen service, because one Shinkansen train (1,000 passengers) corresponds to the capacity of roughly twenty evacuation buses (50 passengers each). From this perspective, the key objective becomes minimizing the time required to transport all simulated evacuees to Kumamoto while ensuring that Shinkansen departures are utilized efficiently and waiting times remain low. This capacity relation emphasizes that the practical value of Shinkansen-based evacuation depends not only on high rail capacity itself, but also on whether station access and feeder bus operations are organized so that this capacity can be used effectively within the available evacuation time.

In Scenario 1, the system performance is evaluated under idealized conditions without background traffic. Only two elements are calibrated: (i) the number of students and teachers per school, estimated from publicly available information, and (ii) walking times from the western entrance of Kagoshima-Chuo Station to Shinkansen Platform 11, derived from a Strava-based tracking experiment. The resulting walking-time distributions are close to normal and likely reflect the underlying pedestrian movement representation (“striping”) used in the simulation model ³⁴.

Scenario 2 extends this setup by adding estimated private passenger-car demand to represent weekday congestion effects in the study area. This allows the analysis of how evacuation buses and pedestrians are delayed by interactions with general traffic, and it also enables a more realistic representation of combined multimodal operations during rush-hour conditions.

Scenario 3 shifts the access mode entirely by assuming that students and teachers walk directly to Kagoshima-Chuo Station instead of being collected by evacuation buses. The person-level results show that this strategy produces substantially longer total per-person times required to reach Shinkansen boarding, measured from school departure on foot, than the corresponding bus-boarding-to-Shinkansen-boarding durations in Scenarios 1 and 2. However, this does not imply that Scenario 3 is uniformly inferior. As shown in Table 2, the last simulated evacuee reaches Kumamoto Station earlier in Scenario 3 than in Scenarios 1 and 2. The reason is that all students and teachers can start moving toward the station immediately, while the bus-based strategies are constrained by fleet size, repeated route circulation, and sequential collection from several school locations.

The Scenario comparison therefore requires a clear distinction between two evaluation perspectives. For assessing overall evacuation completion, this study uses the last-person Shinkansen arrival time as the primary system-level metric. For interpreting the person-level duration before rail departure, we compare the total time required to reach Shinkansen boarding, measured from evacuation-bus boarding in Scenarios 1 and 2 and from school departure on foot in Scenario 3. Under the system-level completion metric, Scenario 3 performs best in the present short-distance school setting. Under the person-level duration metric, Scenarios 1 and 2 perform more favorably. A bus-based strategy may therefore remain preferable when evacuation planning prioritizes organized group transfer, reduced reliance on long pedestrian access, support for evacuees with limited mobility, or origins located farther from Kagoshima-Chuo Station. The results thus do not identify one universally superior access strategy, but instead show that pedestrian and bus-based evacuation perform differently depending on the evaluation criterion and the spatial and operational conditions of the evacuation setting.

From a practical perspective, the proposed simulation framework provides a structured basis for evaluating multimodal evacuation options under urban conditions in which road-based evacuation alone may face substantial operational constraints. Rather than treating evacuation modes independently, the framework enables their combined performance to be examined within one analytical setting and therefore helps identify how alternative evacuation arrangements may contribute to overall system effectiveness. This is relevant for evacuation policy because it supports evidence-based comparison of strategy options, and for operational planning because it allows potential capacity limitations, coordination requirements, and implementation challenges to be examined before emergency measures are adopted. The approach is therefore not limited to reproducing Scenario outcomes, but can contribute to the development of more robust multimodal evacuation concepts for Kagoshima and other urban areas facing similarly complex evacuation conditions.

4.2. Outlook

In future work, we will focus on demand-dependent vehicle routing strategies to reduce evacuation travel times and improve the efficiency of bus operations under varying congestion levels. A key methodological extension will be a clearer separation of travel-time components, distinguishing baseline travel time from additional delays caused by bus bunching and delays introduced by background passenger-car traffic. To strengthen the methodological foundation, we also plan to extend the progress documentation with a broader review of existing evacuation-routing and multimodal simulation approaches.

A further Scenario extension is the integration of traffic control measures at major signalized intersections in Kagoshima City. For this purpose, real (current or historical) signal timing information is needed, including coordinated offsets between intersections rather than approximations based on spatial distance. Implementing measured offsets and realistic signal programs is expected to improve the validity of travel-time estimates and the representation of bus reliability during peak demand.

In addition, we plan to evaluate the traffic safety of evacuation routes suggested in the “Massive Tephra Fall Response Plan” ³ using recent road traffic crash records from the National Police Agency. Crash-based indicators can provide complementary insights at the route and network level beyond travel-time performance alone ^36,37. This safety assessment can be further extended by integrating traffic flow simulation outputs with surrogate safety measures to capture conflict-based risk patterns under different evacuation and congestion conditions.

To enhance realism, we aim to incorporate existing public transport services in Kagoshima using freely available General Transit Feed Specification (GTFS) feeds ³⁸ and complementary information derived from OSM ³⁹.

Moreover, we plan to develop improved visualization methods for simulated pedestrian waiting and walking behavior by introducing catchment-area representations linked to relevant points of interest rather than only to public transport stops, following concepts proposed by Keler et al. ⁴⁰.

Another extension is the explicit modeling of heterogeneous multimodal agents ⁴¹ by redefining the scenarios with varying passenger loads, traffic demand levels derived from detector observations, and population distributions informed by high-resolution sources such as call detail records.

Furthermore, we will examine different categories of vulnerable road users and incorporate more realistic numbers of pupils and teachers, based on guidance from the Kagoshima City Board of Education. In addition, the elderly population is a key group in evacuation planning; therefore, we will include realistic estimates and the spatial distribution of relevant institutions in Kagoshima City, such as nursing homes.

In addition, we will investigate advanced traffic signal control strategies as a means to optimize evacuation performance. This includes replacing estimated fixed-time programs with actuated control ⁴², adapting cycle lengths ⁴³ based on established methods such as Webster ⁴⁴, and improving signal coordination through offset optimization ⁴⁵ to support demand-adaptive progression (“green waves”) for evacuation buses ⁴⁶. Even small changes in signal control and coordination may affect bus delays, reliability, and overall evacuation completion times. Beyond traffic control, the study will be extended toward systematically proposing alternative evacuation routes for comparable origin-destination pairs using advanced routing and optimization approaches.

Future work should also examine the role of Kagoshima-Chuo Station as a potential multimodal evacuation hub. This requires considering regular non-evacuee Shinkansen passengers, more realistic boarding and alighting processes, and the spatial assignment of evacuation origins for vulnerable groups such as seniors and children, who may depend on assisted mobility, public transport, or transfer hubs rather than private-vehicle evacuation from home.

For school-based evacuation scenarios, student-guardian handover requires explicit consideration. In line with the evacuation plan’s assumption of one passenger car per household, this could be represented in future simulations by household evacuation trips with intermediate pickup stops at schools. The bus- and Shinkansen-based strategy examined here can therefore be interpreted as a contingency or stress-test Scenario for cases in which guardian pickup is delayed, infeasible within the available evacuation time, or disrupted by road traffic congestion. This Scenario therefore complements, rather than replaces, local guardian-handover planning in school-based preemptive evacuation.

Finally, we plan to examine evacuation-specific traffic states in which disaster-induced congestion differs from ordinary rush-hour traffic because evacuation demand becomes strongly directional and may resemble large-scale post-event dispersal. This requires explicit modeling of the transition from background traffic to evacuation flows and of the operational control of vehicle fleets and public transport services. Together, these extensions would strengthen the framework for evaluating policy-relevant multimodal evacuation strategies and their operational applicability in Kagoshima and comparable urban contexts.