Using Evacuation Drills to Improve Tsunami Evacuation Preparedness and Resilience

This paper presents the use of tsunami evacuation drills within a coastal community in the Cascadia Subduction Zone (CSZ) to better understand evacuation behaviors and thus to improve tsunami evacuation preparedness and resilience. Evacuees’ spatial trajectory data were collected by Global Navigation Satellite System (GNSS) embedded mobile devices. Based on the empirical trajectory data, probability functions were employed to model people’s walking speed during the evacuation drills. An Evacuation Hiking Function (EHF) was established to depict the speed-slope relationship and to inform evacuation modeling and planning. The regression analysis showed that evacuees’ speed was signiﬁcantly negatively associated with slope, time spent during evacuation, rough terrain surface, walking at night, and distance to destination. We also demonstrated the impacts of milling time on mortality rate based on participants’ empirical evacuation behavior and a state-of-the-art CSZ tsunami inundation model. Post-drill surveys revealed the importance of the drill as an educational and assessment tool. The results of this study can be used for public education, evacuation plan assessment, and evacuation simulation models. The drill procedures, organization, and the use of technology in data collection provide evidence-driven solutions to tsunami preparedness and inspire the use of drills in other types of natural disasters such as wildﬁres, hurricanes, volcanoes, and ﬂooding.


Introduction
Recent and devastating tsunami events (Mori et al., 2011;Lindell et al., 2015;Sassa and Takagawa, 2019) have caused life loss and financial burdens to individuals and communities.Higher evacuation speed and efficiency mean higher survival rate of at-risk populations in low-lying coastal communities, especially for near-field tsunamis with small evacuation time windows (Wang et al., 2016;Raskin and Wang, 2017).To reduce the time of evacuations and maximize the survival rate during real events, evacuation drills have been used as an effective simulation for public education and city emergency planning purposes.
Evacuation drills are a method to practice evacuation from risk areas with planned scenarios that mimic realistic hazardous situations.The drill can provide participants "impactful field-based learning experience" (Zavar and Nelan, 2020) to be better prepared for future disaster responses.Tsunami evacuation drills have three functions: training, assessment, and information.Training: The goal of the training is to ensure that drill participants can implement (even improve) any evacuation instructions that they have received through brochures, lectures, videos or other media.For example, participants can implement evacuation plans, such as the planned destination and route choice from previous experience or knowledge.They can also gain or revise evacuation knowledge after drills.Assessment: The goal of the assessment is to measure the degree to which the evacuation plans and training materials can evacuate or help the largest number of people possibly before the expected tsunami arrival time.With respect to the assessment function, evacuation drills can be viewed as scheduled simulations of actual evacuations.These simulations are designed to enhance evacuation preparedness by identifying gaps in response performances.For example, emergency managers can assess the effectiveness of the signage by investigating whether the evacuation signage can help participants navigate to safety during the evacuation drills.
Information: The goal of the information is to provide scientific evidence, such as human behaviors in drills, for informing/validating tsunami evacuation studies.Tsunamis can be uncontrolled, unpredictable, and devastating.It is difficult to collect real behavior data during such events.Evacuation drills provide an alternative for researchers and emergency personnel to fill this gap to support evacuation modeling and planning (Poulos et al., 2018).
For instance, walking speed data from an evacuation drill can be used to inform individual movement speed in evacuation simulation models when no real evacuation data is available.
An evacuation drill has two unique features: controlled and partially realistic.
People's responses to a disaster include four sequential components: receiving a warning or disaster cue, decision making, evacuation preparedness, and evacuation movement (Lindell and Perry, 2012).Urbanik et al. (1980) and Lindell et al. (2019a) defined the total evacuation time of individuals or households as a function of those four time-related components: t total = f (t w , t d , t p , t e ) where t w indicates the time of receiving a warning or disaster cue; t d is the decision time; t p indicates preparation time; and t e indicates evacuation travel time.The sum of t w , t d , and t p is normally called "milling time" which represents the period of time people spend before evacuation.The sequence of this process can overlap in some situations, but it is generally not reversible so that it can be meaningfully controlled in evacuation exercises.For example, drill participants will not start to evacuate until they are told of a potential threat.This sequential process in drills is equivalent to people not taking protective action until receiving disaster cues (such as the ground shaking) in real events.Though it is impossible to convey the full stress of a real evacuation due to practical or ethical constraints (Schadschneider et al., 2011), an evacuation drill can provide valuable information reflecting evacuation processes to actual events (Poulos et al., 2018).Evacuation drills, thus, are simulation exercises to replace and amplify real experiences with guided opportunities, often "immersive" in nature, that evoke or replicate substantial aspects of the real world in an interactive fashion.

Research Objectives and Questions
This research (1) provides a tsunami evacuation drill template/model/framework for future research and local education programs, and (2) provides empirical evidence of tsunami evacuation walking speed and route choice by conducting evacuation drills across heterogeneous evacuees.By using a GNSS trajectory dataset, probability functions are employed to model participants' average walking speeds and corresponding probabilities.An Evacuation Hiking Function (EHF), revised from Tobler's Hiking Function (Tobler, 1993), is created to describe the relationship between slope and walking speed during evacuation drills.Regression analysis is used to examine what environmental conditions or spatial characteristics affect walking speed during evacuation.Specifically, the research questions include: 1. What are the participants' walking speeds and distributions during evacuation drills? 2. How is walking speed affected by elevation change, terrain type, time of day, evacuation distance, time spent during evacuation, and previous experience with evacuation routes?
3. To what extent does milling time (if added before evacuation) impact the potential mortality rate?

Contribution
Working closely with members of the community, state agencies, and scientists who live and work in Newport, Oregon, we designed a series of evacuation drills for the South Beach peninsula.This region provides two formally designated evacuation sites: Safe Haven Hill (SHH) and the Oregon Coast Community College (OCC).The evacuation drills fulfilled the functions of preparedness training, outreach, and education, as well as contributed to our scientific understanding of evacuation behaviors.Residents, potential visitors, students, and researchers engaged in the evacuation drills to better prepare for "The Really Big One" (Schulz, 2015).Moreover, local authorities, agencies, and researchers utilized this opportunity to assess existing evacuation plans and infrastructure (i.e., evacuation signage and shelters).
In this study, empirical evidence was obtained to analyze the walking speed and factors affecting the walking speed in evacuation drills.Utilizing the empirical data, the walking speed distribution and EHF provide evidence for evacuees' traveling behavior to support tsunami evacuation modeling and emergency planning.Previous tsunami evacuation modeling studies used alternative methods when empirical evacuation data was not available, such as forming a normal distribution for walking speed with arbitrary parameters (Mas et al., 2012;Wang et al., 2016;Mostafizi et al., 2019), assigning a maximum free-flow speed (Lämmel et al., 2010;Mas et al., 2012), or using a linear function to describe speed-density relationship (Takabatake et al., 2017).However, the functions established in this study might provide a more realistic solution for the walking speed assignment for evacuation modeling.
Furthermore, the EHF is useful when modelers aim to include the walking-slope relationship in an evacuation model, such as building the cost-distance model by elevations (Wood and Schmidtlein, 2012).
There are additional components to the drills in this study.First, the coordinates and time data were recorded from a downloadable mapping application Strava (©2020) (Strava, 2020) on participants' mobile devices.Using the Strava application, we were able to track participants' route choices, locations, and time.Participants' Strava trajectories were incorporated into an agent-based evacuation model.These anonymous route trajectories were processed and visualized during a debrief session.The visualizations of participants results were coupled with pre-computed tsunami inundation dynamics to provide participants a practical understanding of how their evacuation behaviors -such as milling time, walking speed, and choice of routes -affects their ability to "beat the wave" (Priest et al., 2016).
Second, we conducted a post-evacuation assessment utilizing a Qualtrics online survey to understand how these organized drills motivate people to prepare for future coastal hazards.
Third, in subsequent iterations of the evacuation drills, we added increasing levels of complexity to the decision-making process, such as separating family geospatially at the time of the earthquake, helping injured friends during the earthquake, and nighttime tsunami drills (Cramer et al., 2018).

Literature Review
Previous research on tsunami risk reduction focused on an interrelated set of topics including infrastructure, warning systems, risk assessment, vulnerability, the adeptness of using vulnerability frameworks (Løvholt et al., 2014), evacuation mapping, and modeling.
However, those modeling approaches were based on evacuation assumptions such as consistent or probabilistic walking speed and shortest route choice.For instance, Wang et al. (2016) and Mostafizi et al. (2019) assigned a normal distribution to walking speed for individuals in the evacuation simulation.The mean of the normal distribution was based on a study of pedestriand walking on the street in a non-emergency situation (Knoblauch et al., 1996).The authors assumed that the walking speed distribution in a normal situation could somehow represent the walking speed in an emergent tsunami evacuation.Beyond those assigning a probabilistic distribution, Wood and Schmidtlein (2012) used a hiking function (Tobler, 1993) to build a cost-distance model for tsunami evacuation.This hiking function was able to capture the impact of slope on walking speed, however, in a normal hiking situation.Overall, existing evacuation models assumed the behaviors in a normal situation could represent behaviors in an emergent evacuation, but more recent literature failed to support this assumption due to a lack of adequate evidence from empirical or experimental evacuation behaviors.

Evacuation Drill
Numerous studies focused on in-building vertical evacuation behavior from drills or real events (Proulx, 1995;Kretz et al., 2008;Xu and Song, 2009;Yang et al., 2012;Qu et al., 2014;Poulos et al., 2018), especially for fire threats with topics addressing speed, milling time, pedestrian flow and density, evacuation fatigue, and modeling.For example, a study from Finland collected data from 18 evacuation situations in different building types ranging from a single hospital ward to a stadium (Rinne et al., 2010).With a large sample size, this study provided empirical evidence of milling time, walking speed, and grouping behavior.
While many studies were developed for in-building fire drills, only a few studies documented tsunami evacuation specifically.Sun et al. (2014) (reversion Sun (2020)) used a single person evacuation drill as an educational method to eliminate people's biased attitudes after the 2011 Great East Japan Earthquake, such as overly optimistic, overly pessimistic, and overly dependent.This study provided a new approach for local authorities to initiate community level drills for communities with limited resources.Further, the single person drill attempted to improve personal preparedness from an individual level.The authors concluded that this type of drill could (1) shift the focus of tsunami risk preparedness practice from the community level to the individual level; (2) change "negative attitudes" toward tsunami preparedness; and (3) transform resident's self-view from someone who would need help to someone who would take the initiative in reducing tsunami risks.Sun (2020) emphasized making a multi-screen video to record the evacuation process for future education use; however, this study did not provide an in-depth analysis of human behavior during a tsunami evacuation drill.Poulos et al. (2018) used an in-building tsunami evacuation drill from a K-12 school to validate the agent-based simulation model for indoor evacuation.This study compared the pedestrian flow and evacuation time in the drills with that in the simulations.The results showed that the error between simulated and actual pedestrian flow rates was 13.5%, and the error between simulated and actual evacuation times was 5.9%.This study provided valuable insight drill data to validate evacuation simulation.However, it was for an indoor fire evacuation rather than an outdoor tsunami evacuation.In 2020, Nakano et al. (2020) introduced a "four-way split-screen" evacuation movie method to establish a communication bridge between experts and non-expert residents.The movie clip simultaneously displayed a school evacuation drill and a tsunami inundation.The author argued that this movie clip was a tool to help experts establish scenario-based evacuation strategies and to implement preparedness activities for non-expert citizens.In the same year, Yosritzal et al. (2020) conducted an evacuation drill to analyze the effects on walking speed from age, gender, and walking distance in a community in Indonesia.The authors assigned six observers to record the travel time for 18 evacuees on three designed evacuation routes.This study provided some empirical drill data and discussed the potentials of using the results to inform the evacuation modeling.
In general, current research about tsunami evacuation drills provide an inadequate understanding of evacuees' behaviors, such as walking speed, factors affecting walking speed, participants' abilities to walk to a higher elevation, and participants' feedback (National Research Council, 2011;Cramer et al., 2018).

Evacuation Drill Technology
Aforementioned tsunami evacuation drill studies (Sun, 2020;Nakano et al., 2020) used cameras to record evacuees' behaviors.For example, the single person drill study (Sun, 2020) applied multiple small-scale artifacts such as video cameras and GPS devices to record the process of the evacuation drills.This study also used people to document the process.An interviewer and a note taker asked related questions and recorded the evacuees' reactions during the evacuation drills.This process not only provided evacuees a more realistic and real-time scenario but also enabled evacuees to provide immediate feedback on their evacuation efforts.Compared with a post-drill survey procedure, this approach can overcome the issue of losing accuracy of behavior and emotion recollection due to memory decay (Wu, 2020).Nevertheless, this approach requires extended inputs including devices, recording labor, and post-drill editing.
Researches have been applying computer graphical simulations to visualize disaster scenarios to improve realistic quality in drills (Chen et al., 2012;Hsu et al., 2013;Farra et al., 2015;Kawai et al., 2015).Virtual Reality (VR) is gaining increasing acceptance because it retains a considerable cost advantage over large-scale real-life drills (Hsu et al., 2013).To improve the traditional VR system assisting tsunami evacuation drills, Kawai et al. (2015) developed a light weight headset that allowed participants to view digital materials during the evacuation movement.The VR could also generate different scenarios during evacuation.While the authors claimed that they had not fully developed this system, this function could be integrated with existing commercial equipment for evacuation education or training purposes.

Walking/Running Speed
The selection of travel speed of an individual is important in tsunami evacuation modeling because travel speed is one of the critical factors for individuals to "beat the wave", but it is difficult to determine due to competing variables.The walking/running speed varies by individual characteristics (age, mobility, height, weight, etc.) and environment conditions (road surface type, slope, wind, etc.).Depending on the geographic features of communities, the population in tsunami inundation areas may have to travel different distances in a short period of time (Wood and Schmidtlein, 2012).
In general, an unimpaired adult's movement speed between 0.6 m/s and 2.0 m/s is commonly observed in field studies.Speed less than 0.6 m/s is considered as an extremely low walking speed (Wu et al., 2019).An unimpaired adult's preferred walking speed is between 1.2 m/s and 1.4 m/s (Mohler et al., 2007;Perry et al., 2010;Wu et al., 2019).
Many environmental factors and individual characteristics can affect the walking speed.Bohannon (1997) summarized the comfortable and maximum walking speed for people aged 20 to 80, and also summarized the speed differences between genders and heights.While maximum running speed declines with the increase of age, the comfortable walking speed has less variance.Not only age but also surface type impacts the evacuation walking speed.Gast et al. (2019) found that the preferred speed (1.24 ± 0.17m/s) on a smooth surface is significantly faster than the preferred speed on rough terrain (1.07 ± 0.05m/s).The more information gained during the movement also significantly deceases the preferred walking speed (Mohler et al., 2007).Rinne et al. (2010) developed walking speed distributions for in-building drills and found that median values for a non-emergency situation were 1.3 m/s for adults, 1.5 m/s for children, and 2.1 m/s for goal-oriented runners.The median walking speed of primary school children was 1.1 m/s on an incline in a cinema drill.Again, many studies documented walking speed for in-building drills but not for outdoor tsunami evacuations.Fraser et al. (2014) comprehensively reviewed 15 studies and summarized pedestrian walking/running speeds for different age groups.In this study, different age groups' walking speeds were used in a GIS-based least-cost distance evacuation model.Though the walking speed spectrum was not created based on empirical tsunami evacuation scenarios, the established speeds by age groups from non-emergency situation are useful to inform evacuation modeling when demographic variables are available.Tobler (1993) built a non-linear function to describe the relationship between slope and walking speed, speed = a×e −b×abs(Slope−c) where estimated parameters are a = 1.67, b = 3.5, and c = 0.05.This function depicts a maximum speed at −2.9 • and the speed decreases monotonically on either side of the maximum value, as shown in figure 6. Tobler's Hiking Function (THF) is widely used in various fields such as recreation planning, rescuing missing persons, assessment of urban social interaction, pedestrian health care facility accessibility, evacuation route planning, etc. (Campbell et al., 2019).Later on, researchers (Rees, 2004;Campbell et al., 2017;Irmischer and Clarke, 2018;Campbell et al., 2019;Davey' et al., 2020) also developed different models to represent speed-slope relationship.Those existing functions depicted two common speed-slope features: (1) a peak representing the maximum travel rate, and (2) a decline on each side of the peak indicating that the speed reduces with the change of the slope.Differences included where the peak point was and how fast the speed decreased on each side of the peak.
The THF had a significant contribution to the Anisotropic path modeling (Wood and Schmidtlein, 2012;Fraser et al., 2014;Priest et al., 2016).It was used to estimate the evacuees' walking speed and also the minimum "beat-the-wave" speed depending on the path cost with the change in slope.Though using the original THF function seems rational, the parameters estimated in the THF were estimated in a normal walking scenario rather than an emergent evacuation scenario.
While evacuation methods have been developed, they are rarely used by local practitioners due to a lack of systematically consistent data or information (Løvholt et al., 2014).Thus, we provide an example of organizing evacuation drills with a practical way to record detailed evacuation data and also provide empirical evidence to augment other approaches.

Study Site
The Cascadia Subduction Zone (CSZ) megathrust is a 1,000 km long dipping fault that runs from northern California, United States up to Northern Vancouver Island, British Columbia.It is about 100-160 km off the Pacific coast shoreline (Thatcher, 2001;Nelson et al., 2006), as shown in Figure 1.A magnitude 9 (M9) CSZ earthquake can pose significant threats to coastal communities in the U.S. Pacific Northwest (Wood et al., 2020), and its likelihood of occurring in the next 50 years is 7% -25% (Goldfinger et al., 2012).It will generate a near-field tsunami with waves of ten meters or more that would strike coastal communities within 20-40 minutes (Gonzalez et al., 2009;Federal Emergency Management Agency, 2012;Goldfinger et al., 2012).According to a study from the United States Geological Survey (USGS) (Wood, 2007) (Thatcher, 2001) The city of Newport, Oregon, United States with 10,854 residents (2019 Census), has a high number of employees (1,455) exposed to risk, though it has relatively lower numbers of residents in the tsunami inundation zone than other communities in Oregon (Wood et al., 2010(Wood et al., , 2015)).As shown in Figure 2 (Wood, 2007).This recreation site is under a high risk of inundation with the added concern of visitors more vulnerable due to having less evacuation knowledge than local residents.Given this context, this region is an ideal case study location for evacuation drills.

Evacuation Drills Process
To maximize educational impact and understand the variables influencing evacuation, drills were conducted across various occupations, origins and destinations, scenarios, and time of day.We invited participants from schools, government agencies, and non-profit education organizations.The public was also allowed to register on-site.Different days of the year were also selected to cover various seasons and weather without keeping participants in extreme weather or hazardous conditions.Evacuating by foot is officially promoted during tsunami evacuation in Oregon by state education and outreach programs (State of Oregon Department of Geology and Mineral Industries, 2012).All participants in our study were asked to evacuate by foot as fast as possible during evacuation drills to better simulate a real evacuation situation.participants were asked to register and provide basic demographic and evacuation knowledge information, including downloading the Strava app, before being sent to start locations.We assigned participants to four starting points that represent popular trail-heads in the SBSP area (SBSP1, SBSP2, SBSP3, and SBSP4) and three starting points that represent popular working and recreation locations in the Yaquina Bay area (Bay1, Bay2, and Bay3), as shown in figure 2. At a predetermined time, participants were told to imagine a CSZ earthquake, pause (simulates the decision time, but is not captured by Strava app), start the Strava app, and then evacuate to SHH or OCC.After each evacuation drill, a debrief and evaluation session was held onsite.Participants were then invited to submit their downloaded Strava evacuation data and complete an online Qualtrics questionnaire.
One key component to this drill process is providing near real-time results to participants immediately after the drills and encouraging them to evaluate their evacuation behaviors and decisions.Specifically, the research team downloaded participants' evacuation trajectory data and overlayed it to the tsunami inundation model (see section 3.4) (Park et al., 2013;Wang et al., 2016;Mostafizi et al., 2019).The visualization of the comparison was shown to participants to encourage them to evaluate their route choices and walking speeds.For instance, participants were shown whether they would be caught by a tsunami if they evacuated on the routes and with the speeds they used in the drills.Participants saw their collective route choices (we did not identify particular individuals on screen to protect privacy) and how their decision-making affected whether they could reach safety (Cramer et al., 2018).
The official drills were conducted six times from 02/18/2017 to 08/10/2017.The dates were selected by participants' availabilities and conveniences.The weather of the six dates covered sunny, cloudy, and slight rain.We did not organize drills in winter to enhance participants' safety.Table 1 documents   To reflect the diversity of the volunteer participants, we included college students, teachers, state and local government personnel and a range of age and gender demographics.The HMSC, OPRD, and State agency groups represented those who were familiar with the evacuation routes, as they were required to walk the routes as part of their job orientation.The fourth, fifth, and sixth waves of participants consisted of students involved in undergraduate research [SURF, REU, and Oregon Sea Grant students], and we increased the complexity of scenarios of the drills.Please see supplement materials for details of those designed scenarios.Due to limited data points of the last drill, those scenarios cannot be scientifically tested, but they are documented in the supplement material for information and future research.

GNSS Trajectory
Global Navigation Satellite System (GNSS) [Sometime refers to Global Positioning System (GPS)] enabled mobile devices to track and map participants' locations.Those data can contribute to human movement, behavior, and route choice research (Chen et al., 2020b).It has been used in disaster studies such as risk mitigation (Ai et al., 2016) and decision making (Zerger and Smith, 2003).In this study, participants used the STRAVA app in their own GNSS enabled mobile devices to record the latitude, longitude, elevation, and time during evacuation.While the data could be impacted by the mobile devices or Satellite connection signal, this study showed that 74/87 (85%) of participants' trajectory data are valid.GNSS enabled mobile devices are easy to access and may be affordable for small jurisdictions to repeat the drills.The Kalman filtering (or linear quadratic estimation) (Kalman, 1960;Kalman and Bucy, 1961) was used to reduce noise and error from the GNSS data.

Tsunami Inundation and Participants' Milling Time
Beyond the traveling data collection process, two other critical components affect the success of an evacuation: how a tsunami inundates and how long evacuees spend before evacuation.This study, therefore, incorporated (1) tsunami inundation, (2) milling time, and (3) the empirical drill evacuation GNSS trajectories into an Agent-based Tsunami Evacuation Model (ABTEM) created by the OSU research team (Wang et al., 2016;Mostafizi et al., 2017Mostafizi et al., , 2019) ) to articulate the effectiveness of participants' evacuation.
Tsunami inundation layer: Tsunami inundation time series data was developed by Park et al. (2013) and represented an extreme scenario generated by a M9 Cascadia Subduction Zone event.We used a 0.5-meter water depth as the threshold to indicate that participants were caught by the wave.
Milling time: All participants evacuated immediately during the drills, while in the real events people tend to spend time on decision making, collecting and confirming information, collecting necessities, contacting family, or picking up family before evacuation (Lindell and Perry, 2012).Those psychological and physical tasks can be represented by the aforementioned milling time people spend before evacuation.Due to the scope of this study, drill participants did not experience the actual milling process, so the GNSS data only recorded the evacuation movement.Thus, to understand how the milling time affects By varying milling time, this analysis demonstrated whether and where evacuees would be caught by waves based on their current walking speeds and route choices when exposed to a near-field tsunami caused by the M9 earthquake in CSZ.

Walking Speed Distribution during Evacuation
Walking speed during the evacuation drill had a mean of 1.58 m/s and a standard deviation of 0.62, as shown in figure 4. The mean walking speed was slightly faster than the "fast" walking speed (1.52 m/s) for unimpaired adults that was observed in previous literature (Wood and Schmidtlein, 2012;Fraser et al., 2014).Most of the time (83%) during the evacuation drills, people walked faster than a "moderate" walking speed of 1.22 m/s (Knoblauch et al., 1996;Langlois et al., 1997).As expected, the result indicates that on average people moved faster in the drills than in normal situations.(Langlois et al., 1997;Knoblauch et al., 1996); Moderate = 1.22 m/s (Langlois et al., 1997;Knoblauch et al., 1996); Fast = 1.52 (Wood and Schmidtlein, 2012;Fraser et al., 2014); Transition to Run = 1.79 m/s (Fraser et al., 2014) difference in the walking speed between groups, the average value and the regression analysis show a relatively clearer pattern: Participants evacuating from HMSC to OCC, on average, moved more slowly than others (β = 0.04, p < 0.01), as shown in Figure 4a.The total evacuation distance from HMSC to OCC was also longer than the other evacuation scenarios (for example, from HMSC to SHH and from SBSP to SHH).For occupation groups, REU and Sea Grant students tended to evacuate faster than the others (β = 0.21, p < 0.01), as shown in figure 4b.Those students were on average younger than other groups of people, and the age was, according to previous studies, negatively correlated with the walking speed (Gast et al., 2019).We employed probability methods to model the average walking speed from the drills.
These fitted probability models can inform evacuation simulation and modeling research.
Probability models were fitted using package "tdistrplus" in R (Delignette-Muller and Dutang, 2015).Based on the right-skewed shape of the average walking speed of participants, three distributions were selected as candidates for the model fitting process: Log-logistic, Gamma, and Burr distributions, as shown in Figure 5. Figure 5b demonstrates that all three estimated models fit the empirical data well for the range below 2 m/s.A reasonable explanation is that the this dataset provided more data points for the range below 2 m/s than the other ranges.Because the majority of participants (93%) evacuated with the average speed at the range from 1.2 to 2.0 m/s, those functions can describe the overall data accurately.Log-logistic function shows the best goodness-of-fit statistics of the three candidates, as illustrated in Table 3.Thus, the Log-logistic function is selected to model people's average walking speed during the tsunami evacuation drills.The Log-logistic distribution has a cumulative density function: whereas F (x) represents the cumulative probability of having speed less or equal to x. Maximum likelihood estimation shows that estimated β = 12.3267 and α = 1.5490.
Therefore, the function can be simplified as: 12.3267  (2) This walking speed function with the best-estimated parameters can be used to inform the individual's average walking speed for tsunami evacuation modeling.For example, it can describe the average walking speed of an individual in agent-based tsunami evacuation models presented by Mostafizi et al. (2019) and Mas et al. (2012).The emergency management practitioners can also use this function to estimate the pedestrians' evacuation travel time and assess the current evacuation plans.However, this function does not represent all situations and all population segments due to the limited sample size and representativeness.
Using walking speed for each population segment was summarized in Fraser et al. (2014) and might be more useful when the demographic data and elevation data are available.This limitation of our walking speed function will be discussed in section 6 in detail.(3) The fitted EHF (R 2 = 0.10, MAE = 0.31) produced less error than the THF (R 2 = 0.06, MAE = 0.37) when applied to this drill dataset.The THF shows that the peak walking speed (1.67 m/s) occurs on a slight downhill slope of -0.05 (-2.86 • ), whereas the EHF shows the peak appears on the approximately flat slope of 0.004 (0.23 • ) in the evacuation drills.
Some data points appearing above the peak point indicate that evacuees sometimes run at a fast speed [a speed from 1.79 to 2.11 m/s is the common transition from walk to run (Mohler et al., 2007;Fraser et al., 2014)] on flat areas during the drills.Indeed, we observed some participants started with a fast run at the beginning of the evacuation drills.
The estimated walking speed at the positive slope of the EHF is faster than the speed of the original THF, consistent with expectations.It is likely that participants moved faster under the pressure of an emergency evacuation situation than a normal hiking scenario.
There are limited data points at negative slope (due to participants evacuating uphill most of the time) which creates the difficulty in determining the usefulness of this model for the negative slope range.Nevertheless, the positive slope section of the function is more useful in practice because of the assumption that people move uphill most of the time in a real tsunami evacuation.
In the regression analysis, as shown in Table 4, the slope (change in elevation divided by change in horizontal distance) shows a negative impact (β = −0.78,p < 0.01) on the walking speed during the evacuation drills, and that is consistent with the results from EHF when the slope > 0. The multiple linear regression is suitable to justify this relationship because the majority of data points are located at a slope > 0 (Slope > 0 means participants evacuated uphill.),so the slope-speed relationship is approximately monotonic.In this dataset, the most data points are at the positive slope range; however, this may not be consistent with other communities where evacuees have to move downhill before going uphill.In that case, many data points would be located at the negative slope range so the linear regression would not be able to capture the non-monotonic slope-speed relationship.
Using the EHF would be more useful and reliable than a multiple linear regression analysis for evacuation dataset from those communities.The initial distance (when the trajectory line intersects with the y-axis) represents the distance to destination when a participant is at the start location.The end distance (when the trajectory line intersects with with the x-axis) represents the point that participants arrive at the destination.
The majority of participants of the group that evacuated from Bay to SHH took less time than the other two groups.Participants who evacuated from Bay to OCC had longer total evacuation distance than participants of the other two groups.However, some of them arrived at the destination sooner than some participants that evacuated from SBSP to SHH.
Indeed, the space-time figure shows that some participants who evacuated from SBSP to SHH moved in the opposite direction from the destination sometime during the evacuation drills, which resulted in the late arrival.Two participants who evacuated from SBSP to SHH spent a longer time than others (total 35 and 37 minutes).Their space-time trajectories ascended from 15 to 20 minutes and then declined again, which indicates that they walked in the opposite direction to the destination or wandered around during the evacuation but eventually went in the correct direction.Two other participants (two red outliers) evacuated to neither SHH nor OCC, so the two trajectory lines depart from the destinations.The two participants reported to researchers that they initially planned to evacuate to SHH, but lost their ways and then evacuated to the direction they believed to be safe.We verified that the destination they evacuated to was outside of the tsunami inundation zone; however, their data points were excluded from the analyses for data consistency purposes.

Milling Time Impact
Figure 8 illustrates the percent of evacuees caught by the tsunami when milling time varies, if participants took the walking speeds and route choices from the drills in a real event.The X-axis represents the amount of artificial milling time that is added before evacuation.For example, x = 10 means the participants spent 10 minutes before evacuation and then started to evacuate based on their empirical drill trajectory.The corresponding y = 13% when x = 10 means that 13% of participants evacuating from SPSP to SHH would be caught by the tsunami if participants spent 10 minutes milling before evacuation.A more immediate impact occurs on the mortality rate for evacuees from SBSP to SHH than the other two groups when milling time increases.This can be explained by the fact that the origins are located closer to the ocean when participants started from SBSP than other two groups.A steep rise in mortality rate between 10 to 15 minutes indicates that 10 minutes is a critical milling time for people to evacuate from SBSP to SHH.For the other two groups evacuating from the bay, milling time shows no impact on mortality rate until it reaches 15 minutes; however, the sharp increase curve from 15 to 25 minutes indicates that the impact rises dramatically during this time range.When adding more than 30 minutes of milling time before the evacuation, all participants would be potentially caught by the tsunami wave if they used the same walking speed and route choice from the drills.In addition to the factors discussed in the previous sections, variables such as walking surface type, night/day, time spent during evacuation, and the distance to destination also have a potential impact on the walking speed during the evacuation drills.A regression analysis was applied to investigate the impact of those factors.
A hypothesis suggests that the roughness of a terrain correlates to the walking speed.
The results support this hypothesis.In the drills, evacuees' walking speed on the rough terrain (sand, non-paved trail, or natural trail surface) is on average 0.11 m/s slower (β = −0.11,p < 0.01) than the walking speed on the smooth surface terrain (paved trail, side-walk, or motorized vehicle lane).This finding is consistent with previous research (Schmidtlein and Wood, 2015;Gast et al., 2019).The result also provides empirical evidence for the evacuation modeling by varying land cover types.Schmidtlein and Wood (2015) explored how different land cover types influence anisotropic least-cost-distance model outcomes.In their study, for instance, dirt/gravel/grass/sand lands were assigned with a lower walking speed value than paved roads.The authors admitted that analysts need to arbitrarily decide which speed value (proxies) is assigned to which land cover type, but the empirically derived values on land covers from actual evacuations would be ideal.The empirical relationship from this present study is closer to this "ideal".For example, evacuation walking speed on the rough surface is 0.11 m/s slower than that on the smooth surface in the drills, and this information can be used to inform a simplified dichotomy of walking speeds by the land cover types for evacuation models.
Walking speed may also be impacted by evacuating in the day or at night.The result indicates that evacuees moved more slowly at night than in the day by -0.21 m/s on average (β = −0.21,p < 0.01).A rationale assumption involves lower visibility of routes and evacuation signage at night, so participants spent more time on navigating or looking for routes and intersections.This finding is consistent with the conclusion from a self-assessment study (Sun and Sun, 2020) that people need longer mobilization time and longer clearance time to reach safety at night.The time spent during the evacuation drills has negative impacts on the walking speed by controlling other variables constant, as expected.The result shows that every increase of one standard deviation in time (443 seconds) is associated with 0.17 decreases in the walking speed (standardized β = −0.17,p < 0.01) on average.An interesting finding is the negative association between the shortest distance to destination and the walking speed.This study calculated the shortest distance from every point on the evacuation route to destinations for each evacuee.An evacuee may not follow this theoretical shortest route; however, it indicates how far away the evacuee is to a destination if taking the shortest route.Every increase of one standard deviation in the distance to destination is associated with 0.10 decreases in the walking speed (standardized β = −0.10,p < 0.01) on average.In other words, an evacuee moved faster when being closer to the destination, even controlling the time spent during the evacuation drills.It should be noted that the R 2 of the fitted model is low, maybe because of a large amount of inherently unexplainable variation.However, this low value may not be a critical concern since the purpose of this regression analysis is explanation rather than prediction.

Participants' Feedback
This section summarizes the post drill survey results regarding participants' attitudes, behaviors, opinions, and lessons learned as a part of the evacuation drills.The feedback from participants reflects the training purpose of the drills and indicates potential issues of evacuation strategies.The majority (87%) of participants stated that the drill was useful and they felt more prepared to evacuate to a safe zone after the drills.An overwhelming majority of respondents believed the drill was useful regarding (1) learning about evacuation time (100%), (2) improving their ability to evacuate to safe zones (87%), and (3) learning Model is significant at 0.00 level (F = 370, p < 0.00).R 2 = 0.047 Significant level: * 0.01, ** 0,05, *** 0.001 how to improve evacuation effectiveness (68%).Most respondents (70%) who evacuated from SBSP pointed out a difficulty in finding clear evacuation signage.Examples of participant comments include: "I feel that the signage around the route that I used was not clear, I tried following the signs but I did not make it to Safe Haven Hill,""...There are not many evacuation signs in the trail..." and"I think they should really really really improve the signs at the Newport Campsite...." Most participants stated that they would prepare a "to-go" disaster kit (58%), make an evacuation plan (61%), and participate in additional evacuation drills (71%).Because the post drill survey was voluntary, some participants opted out of the responses, which resulted in a small number of respondents (n = 31).Thus, there was not enough data to do a statistical assessment of the participants' responses; however, we were fortunate to gather enough information to illustrate the important usefulness of utilizing on-the-ground participant information.It indicated opportunities for route signage improvement, evacuation behavior, and the importance of the drill as an outreach and educational program.

Conclusion
This research organized tsunami evacuation drills across heterogeneous evacuees in a coastal city in the Cascadia Subduction Zone and could serve as a tsunami evacuation drill template/model/framework for preparedness improvement.This study also provided evidence of tsunami evacuation behaviors by using a spatial trajectory dataset collected by GNSS embedded mobile devices.The results include the following: 1. Walking speed distribution was created for different groups of participants.In general, the walking speed has a distribution with a mean of 1.58 m/s and a std. of 0.62 during the evacuation.An evacuation walking speed function, using Log-normal distribution, was created based on the empirical data to describe the probability of mean walking speed (section 4.1).This function can be used to inform the individual's average walking speed for tsunami evacuation modeling studies.
2. The Evacuation Hiking Function was built based on Tobler's Hiking Function with three estimated parameters to model the relationship between the walking speed and the slope in evacuation drills.This function can also be applied to evacuation modeling studies such as calculating cost-distance.
3. The evacuation walking speed is negatively associated with slope (section 4.2), time spent during evacuation, rough terrain surface, walking at night, and distance to destination (section 4.5).Participants who evacuated from the Bay to SHH (section 4.5) and REU students (younger age group) were found to move faster than others (section 4.5 and 4.1).4. 10 minutes is a critical milling time for people to evacuate from SBSP to SHH. 15 minutes is a critical milling time for people to evacuate from the Bay to SHH and from the Bay to OCC (section 4.4).
The feedback from participants indicated the evacuation drill could potentially serve as an effective educational activity to discover the preparedness gaps for both participants and emergency planners.Overall, the results from this study can be used for public education, evacuation plan assessment, and supporting evacuation simulation or models.Results from this study indicate that involving local participants in tsunami drills would enhance our scientific modeling endeavors, increase local preparedness knowledge, and enhance the role of agency, thereby contributing to a culture of preparedness and, ultimately, community resilience (Cramer et al., 2018).

Limitations, Future Work, and Recommendations
Participants in the drills are not exposed to the same level of stress that evacuees exposed in a real event, so the walking speed observed in this exercise should not be assumed to be higher than that in a real evacuation (Schadschneider et al., 2011).In other words, people may receive more motivation (from the environment, peers, or themselves) to move faster in a real event.Future studies can develop a calibration factor between the speed in real evacuations and evacuation drills.For example, one can record the walking speed in real events and the walking speed in drills from the same community, and measure the difference between the two walking speeds as a calibration factor.The factor may be generalized to guide simulation for other communities.Human evacuation performance could also be impacted by topological and geological difference between communities.For example, lower-lying communities with large flat areas require evacuees to walk further in distance to reach safe zones.Researchers or local emergency managers may organize drills in their own towns/cities to better achieve the education and assessment purposes.
Due to the long time input for participants to finish the whole drill process, which typically requires half to one day, randomly inviting local residents to participate could result in a demographic representativeness bias.For example, retired seniors are more likely to participate than others because of time availability.Therefore, this study proactively invited participants with various demographics and knowledge backgrounds to mitigate the representativeness issue.Our results were built on the diversity of the volunteer participants who represent a range of age, gender and other demographics, such as college students, teachers, and government personnel.Some biases, still, can be generated from the invitation process.Future work can either (1) provide demographic information to explain how a sample represents a population; or (2) choose a random sampling method and at the same time simplify or shorten the drill process to reduce the time requirement for participants.
Further research can also investigate the mobile phone GNSS signal issue.Given the exploratory nature of these drills, we used participants' owned devices to record evacuation trajectories.Within the 87 cases from participants who submitted STRAVA data, 13 of them were either corrupted or incompleted, therefore were deleted from the dataset.We observed that some GNSS enabled mobile phone devices could not record the trajectory in STRAVA either due to the weak signal or dysfunction.Future research can provide other types of devices to overcome this issue.
Another limitation involves the survey and basic registration.Participants were encouraged to take an online survey after the evacuation drill, but the demographic information was voluntary, with some participants opting out of such responses.As noted above, the drill portion of this study was utilized primarily to ground-truth the process and the evacuation models.We were fortunate to gather enough information to illustrate the important usefulness of utilizing on-the-ground participant information; however, there was not enough survey data to do a comprehensive assessment of the participants.Future research can also collect demographic, knowledge, and physical information (weight, height, and general health levels) at the pre-registration site.Such information may also affect the walking speed in the evacuation drills.Future work can also divide the impact factors to more categories for regression analysis and examine the further impact of each category.For example, while terrain in this study is dichotomized to natural/paved surfaces, future study can divide the terrain to multiple categories such as the example in Schmidtlein and Wood (2015).URL https://linkinghub.elsevier.com/retrieve/pii/S0198971501000217

Figure 1 :
Figure1: Cascadia Subduction Zone and its Impact Area, revised based on(Thatcher, 2001) , restaurants, local marine facilities, and state or national agencies [National Oceanic and Atmospheric Administration (NOAA), Oregon Department of Fish & Wildlife (ODFW), Oregon Coast National Wildlife Refuge (OCNWR), Oregon Coast Aquarium (OCA)] are located in the low-lying inundation area in the south part of Newport.South Beach State Park and surrounding facilities attract an average of 1,135,584 visitors per year and has the second-highest annual average number of day-use visitors among the 66 parks along the Oregon Coast

Figure 2 :
Figure 2: Tsunami Evacuation Route Choice and Tsunami Inundation Area in Newport, OR, USA(State of Oregon Department ofGeology and Mineral Industries, 2012)

Figure 2
Figure 2 illustrates the evacuation origins, destination, and route choice, and choice density of the drills.Three sites around the bay and four sites within the SBSP were selected as origins to represent the heterogeneous land-use locations, including recreational activity locations, work locations, and parking locations.Two high ground areas, Safe Heaven Hill (SHH) and Oregon Coast Community College (OCC), were selected as the evacuation destinations (State of Oregon Department of Geology and Mineral Industries, 2012).While SHH appears within the inundation zone on X-Y surface, it has a higher elevation (>70 ft.) than surrounding flat land and can serve as a vertical evacuation site that facilitates a rapid evacuation from the low-lying South Beach Sate Park and the harbor areas in Newport (Oregon Office of Emergency Management, 2016).OCC is located inland with high elevation and serves as a horizontal evacuation site that provides ample space for establishing a refuge.

Figure 3
Figure3illustrates the evacuation drill procedure.Across the six waves of drills, the number of samples collected for each drill.Within the total 136 participants, 87 uploaded trajectory data and 74 are valid for analyzing.Thirty one participants conducted voluntary post-drill surveys.Various scenarios in those six drills are documented below: 1. 02/18/2017 OSU students and professors participated in the first official drill.2. 05/11/2017 Participants included OSU students and professors, staff from Oregon Parks and Recreation Department (OPRD), Teen Community Emergency Response Team (CERT) from Toledo Junior High School, Hatfield Marine Science Center (HMSC), and general public and volunteers.3. 06/16/2017 Participants included OSU students and HMSC staffs.4. 06/29/2017 OSU Summer Undergraduate Research Fellowship (SURF) program students and professors were the main participants.This drill required participants to give up cell phone and maps and evacuate based on their own knowledge and on-site evacuation signs.Letters in envelopes were giving to participants for different roleplaying scenarios (Fishing from Shore, Whale Watcher, Looking for sea shells.Details can be found in supplement material).5. 07/13/2017 (night drill) Participants included Research Experiences for Undergraduates (REU) students, Sea grant scholars, and HMSC staffs.6. 08/10/2017 (night drill) Participants included REU students, Sea Grant students, the drill evacuation results, a sensitivity analysis of the milling time was conducted in the ABTEM.Specifically, we included different artificial milling time (0 min.-40 min.)before each participant's GNSS trajectory to analyze the effect of milling time on mortality rate [percentage of participants caught by tsunami based on the inundation model from Park et al. (2013)].

Figure 4 Figure 4 :
Figure 4 also shows the walking distribution categorized by origin-destination and by groups of people with different occupations.While the boxplots illustrate that no obvious

Figure 5 :
Figure 5: Model fitting for average walking speed

Figure 6 :
Figure 6: The impact of Slope on Speed during Evacuation

Figure 7 :
Figure 7: Space-time Trajectory for Each Participant during Evacuation

Figure 8 :
Figure 8: The Impact of Milling Time on Mortality The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.This project was approved by the Oregon State University Human Research Protection Program (HRPP) and Institutional Review Board (IRB) and follows the regulations to protect participants, with projects reference number 6779 and 7349.The authors would like to acknowledge the funding support from the Oregon Sea Grant program #NA140AR4170064 and the National Science Foundation through grants: CMMI-HDBE #1563618, #1826407, and #1902888.Any opinions, findings, and conclusion or recommendations expressed in this research are those of the authors and do not necessarily reflect the view of the funding agencies.The authors are also grateful to the generous support and collaborations from Oregon Parks and Recreation Department, the Hatfield Marine Science Center (HMSC) to conduct the drills at the South Beach State Park, Newport, OR and use the HMSC facility for the night drills.We are also thankful to each and every drill participant for their contributions in particular the group of Toledo high school students, their participation have provided a great example on how K-12 students' risk perceptions can be changed through evacuation drills.publisher: Routledge eprint: https://doi.org/10.1080/03098265.2020.1771684.URL https://doi.org/10.1080/03098265.2020.1771684Zerger, A., Smith, D. I., Mar.2003.Impediments to using GIS for real-time disaster decision support.Computers, Environment and Urban Systems 27 (2), 123-141.

Table 1 :
Evacuation Drill Data Collection

Table 2 :
Variable Description

Table 3 :
Statistics for Model Comparison

Table 4 :
The Impact of Variables on Speed