Methodology to Quantify Statewide Evacuations

Abstract

Mass evacuations, particularly those at a statewide level, are among the largest single-event highway traffic events. They can last several days, cover thousands of miles of roadway, and include hundreds of thousands of people and vehicles. Often, evacuations are criticized for their inefficiency and poor management. Despite the critical importance and the potential impact on lives and safety, there are no recognized methods systematically to quantify traffic characteristics at statewide scales. This paper documents the development and application of an analytical method to measure statewide mass evacuations. The proposed approach sought to be both practical and cost-effective. The research methods are based on simple, yet widely available, and easily understood traffic count datasets that support both qualitative and quantitative analyses. By spatially and temporally arranging sensor-based statewide traffic volume data from the hurricane Irma (2017) and Michael (2018) evacuations, these methods are applied to describe and answer several key questions about statewide mass evacuations. The methods developed in this research are able to estimate the start and end of the auto-based evacuation, the loading and peaking characteristics of traffic, the total number of vehicles involved in the evacuation, and the effective start and end time of the auto-based reentry. Among the key findings of this work were that the hurricane Irma and Michael evacuations began several days before landfall, peaking two to three days before the storm. It is expected that state departments of transportation and emergency management officials can apply similar methods to assess and better plan future evacuations.

Keywords

sustainability and resilience transportation systems resilience disaster response recovery and business continuity emergency evacuation

Mass evacuations, particularly those at a statewide level, represent the largest single-event traffic movements that can occur. These complex transportation events can last several days, span thousands of miles of roadway, and include hundreds of thousands of people and vehicles traveling with vital urgency. Often, evacuations are plagued by enormous travel delay and congestion and are nearly always criticized for their inefficiency and lack of management. However, few studies quantitatively examine such events to assess objectively what travel conditions were actually like. Typically, opinions are based on media reports that tend to sensationalize poor operations and focus strictly on areas that are performing poorly.

There are many reasons why mass evacuations have tended not to be comprehensively studied. Obviously, they are large and complex, but another reason is the lack of standardized methods by which to quantify traffic characteristics systematically at the proper scale. Few indicators, apart from a lack of fatalities and the amount of vehicles moved, determine if any evacuation was “effective” or not. Instead, outside of media reports, emergency managers and transportation professionals often work under a general assumption that an evacuation was effective if people were able to get out of danger and no one drowned in their homes.

To provide a basis of measurement and comparison, this paper describes research to examine and assess evacuation characteristics. More importantly, the paper attempts to create and apply a method to measure and quantify evacuations in an unbiased, practical, and repeatable fashion that is both intuitive and beneficial to state officials. The research methods are based on simple, yet widely available, and easily understood traffic count datasets.

Traffic volume counts serve as a fundamental parameter of traffic measurement, but they can yield enormous insights into the ebb-and-flow of daily commutes and when, where, how much, and how fast people are able to move during an evacuation. These wide-area, long-term vehicle counts can also be used to illustrate the movement of evacuees after the hurricane to understand better how many vehicles were affected, when the recovery began, and even how long it took based on when the traffic patterns returned to normal.

Building on these concepts, the objectives of the research were to quantify spatially and temporally key aspects of the evacuation and reentry process in Florida during the record-setting 2017 and 2018 hurricane seasons, specifically:

When did the auto-based evacuation make a measurable impact on traffic (when did it noticeably start)?

What were the loading characteristics of the evacuating traffic on the network?

What was the peak evacuation volume and when did it occur?

When did the auto-based evacuation conclude?

How many vehicles were used in the evacuation?

When did reentry begin?

When did reentry effectively conclude?

These objectives were achieved through the observation and analysis of roadway volumes collected from ground based sensors (predominantly, magnetic-loop detectors) during the evacuations and reentries from hurricanes Irma (2017) and Michael (2018) in the State of Florida. These two events provide a unique opportunity to study the evacuation phenomenon because they are among the largest in the history of the United States; they affected nearly all of the major metropolitan population centers of the state; and traffic volumes are recorded on a geographic scale and at levels of fidelity rarely achieved in prior evacuation studies.

The scientific contribution of this work is its demonstration of a straightforward and reproducible methodology to measure the auto-based evacuation response and reentry of an area. The methods demonstrated in this paper also have a significant practical value for state transportation and/or emergency management agencies seeking quickly and accurately to assess evacuation characteristics. This research also expands the literature by providing insights into the less-often-studied topic of evacuation reentry timing and participation. Finally, it creates a set of aggregate evacuation parameters that can be used to calibrate evacuation planning and simulation models, making the paper a valued reference for future research studies.

Background

The 2017 evacuation from Hurricane Irma has been referred to as the largest evacuation in the history of the United States. Approximately 6.5 million Floridians were placed under either mandatory or voluntary evacuation orders ( 1 ). The overwhelming response to Hurricane Irma was fueled by several factors that were unique to the storm: (i) Hurricane Irma had already devastated several Caribbean islands, including the U.S. Virgin Islands and Puerto Rico, resulting in several known deaths at the time ( 2 ). (ii) At one point, Hurricane Irma was the fifth strongest hurricane ever recorded in the Atlantic Ocean. (iii) The storm’s path and “cone-of-uncertainty” threatened nearly the entire State of Florida. (iv) Fluctuations in the storm’s path indicated possible devastating storm surge to nearly all of Florida’s coastal areas, where the majority of residents live.

The National Hurricane Center’s (NHC) storm path prediction for Hurricane Irma 67 h before landfall suggested a Saffir-Simpson scale Category 4 hurricane making landfall in Southeast Florida and continuing up the eastern coast. However, 21 h later the NHC’s revised storm path predicted a landfall on the Florida Keys and a northern approach along the western coast ( 3 ). It can be surmised that the storm’s path generated evacuees from both the eastern and western portions of the state as well as coastal regions in the south from Key West north to Jacksonville, FL.

Ultimately, Hurricane Irma made two landfalls within the State of Florida. The first was near Cudjoe Key in the lower Florida Keys, on September 10, 2017 at approximately 9:10 a.m. Eastern Time as a Category 4 hurricane with sustained winds of 130 mph (209 kph). The second landfall was at approximately 3:35 p.m. Eastern Time near Marco Island, just south of Naples, FL as a Category 3 hurricane with winds of 115 mph (161 kph) ( 4 ). The storm left approximately 6.7 million homes (65% of the state), without power ( 5 ). Hurricane Irma was responsible for taking the lives of 75 Floridians and costing an estimated $49 billion ( 6 ). The lower Florida Keys remained closed to non-residents for approximately three weeks following the storm ( 7 ).

Hurricane Michael was a Category 5 hurricane that made landfall near Mexico Beach, FL on October 10, 2018 at approximately 12:30 p.m. With sustained wind speeds of 155 mph (250 kpm), Hurricane Michael was the strongest storm by wind speed to strike the mainland U.S. since Hurricane Andrew in 1992 ( 8 ). However, initial reports suggested Hurricane Michael would make landfall as a Category 3 hurricane, which may have had an impact on evacuation participation rates leading up to landfall ( 9 ). Hurricane Michael’s intensity projections 54 h before landfall forecast a Category 1 or Category 2 storm ( 10 ). Ultimately, Hurricane Michael was directly responsible for 16 deaths and approximately $25 billion in damage. In total, 21 counties issued evacuation orders, of which 12 held mandatory orders in place ( 11 ).

Literature Review

Available data sources by which to examine evacuations have been rapidly increasing. Examples include geotagged Tweets ( 12 , 13 ), travel time predictions ( 14 ), and mobile phone location data ( 15 , 16 ). While these sources help address gaps, such as under-representation of the younger population and low participation rates in surveys ( 17 ), they have their own limitations, such as the need for geo-locations and use of the social media platform. The advantages of traditional detectors remain, including low cost ( 18 ), real-time data access for departments of transportation ( 19 ), and lack of need for active use of social media platforms by evacuees, indicating the value of this data source alone or in conjunction with emerging data sources.

Detectors are most prevalent on high volume, high capacity roads. This corresponds well with prior survey-based research which found that evacuees have a strong preference for Interstates (e.g., Haddad [ 11 ], Dow and Cutter [ 20 ]) and highways (e.g., Wu et al. [ 21 ]), although familiarity and experience with a roadway also play a role in evacuation route selection (e.g., Wu et al. [ 21 ], Murray-Tuite et al. [ 22 ], Lindell and Prater [ 23 ], Vogt and Sorensen [ 24 ]). A few studies have used detector data to investigate different aspects of the evacuation. Wolshon ( 25 ) used detector data from Louisiana collected during Hurricane Katrina to assess how well the maximum capacities suggested by the Highway Capacity Manual matched the detector reported flows for different types of roadways. These roadway types included freeways operating in the normal direction, contraflow freeway segments, four-lane arterial roadways, and two-lane arterials. On all of these roadways, the maximum flows were lower than the theoretical values ( 25 ).

Li et al. ( 26 ) used automatic traffic count data from tollbooths to develop empirical response curves for Hurricane Irene for a single county in New Jersey. They also identified evacuation volumes and compared these with the volumes from the previous week. They identified the evacuation traffic as starting 6 h before the mandatory evacuation order for the barrier island and an overall quick response to a mandatory evacuation order, which they suggested could be explained by the large tourist population ( 26 ). They later expanded their spatial coverage and data to include weigh-in-motion stations and historical travel time data ( 27 ). This study reported that the evacuation took approximately 36 h and the evacuation traffic was more obvious near the shore, tending to move west instead of north along the shore. Similar spatial patterns were observed for Hurricane Sandy, although volumes were lower than for Hurricane Irene ( 28 ). Investigations into the evacuation of Hurricane Irma have attempted to illustrate and quantify congestion caused by the evacuation ( 29 , 30 ) and identify roadway bottlenecks ( 31 ) as well as road network accessibility ( 32 ). Irma has also served as a case study to investigate evacuation decision making ( 13 , 33 ) and to analyze the operational and safety impacts of emergency shoulder use ( 34 ).

These prior studies focused on the evacuation preceding the event. Compared with research into the evacuation process, fewer studies have investigated the length of time people remain away from home ( 21 ) or reentry traffic ( 35 ). However, from survey data, for Hurricane Lili the average duration of time away from home was 2.33 days ( 36 ) and for Hurricane Katrina the average was 13.8 days ( 21 ). This large range suggests that the amount of time evacuees stay away from home can vary substantially, depending on the hurricane and its effects. Furthermore, some people may permanently migrate (never return).

Managing reentry can be challenging. In contrast to evacuation where destinations are dispersed, in reentry, traffic converges to the area(s) that were evacuated ( 37 ). These areas may have suffered damage and have debris issues and utility outages. Several studies have reported low compliance with official reentry plans: 38% for Hurricane Ike ( 38 ) and 46.4% returning on or after the scheduled return date for Hurricane Rita ( 39 ). Considering this relatively low compliance, it is important for researchers to investigate and agencies to understand when evacuees will return and the volumes in which they do so. This study uses aggregate data to improve this understanding.

This paper advances knowledge beyond what is currently understood in the literature, by demonstrating how key parameters of the evacuation process can be identified using simple and widely available traffic count data. At present, there is no systematic method to identify when an evacuation began, how long it lasted, or to identify the start and end of reentry. It is anticipated that practitioners will be able to use these metrics to evaluate the effectiveness of evacuation messaging, while researchers may utilize these metrics to validate more sophisticated procedures or to better model evacuation choice behavior.

Methodology

Broadly, the research methodology utilized traffic count data taken from across the State of Florida to investigate the auto-based evacuation response and reentry of coastal communities from both hurricanes Irma (2017) and Michael (2018). The first part of the methodology was to process traffic count data used in the analysis. The second part of the methodology discussion demonstrates how this data was used to estimate the start and end of the auto-based evacuation, the traffic loading and peaking characteristics, and the total number of vehicles used in the evacuation process, as well as the effective start and end of the auto-based reentry.

Data Collection and Processing

The Florida Department of Transportation (FDOT) Transportation Data and Analytics Office gathers roadway data from across the State of Florida. Real-time traffic information is provided during emergencies such as hurricanes and wildfires. Traffic information, namely volume, speed, and vehicle classification are collected hourly from telemetric monitoring stations located throughout the state. There are 255 data collection sites on Florida roadways at the time of this study; each provides bidirectional hourly counts and speeds. For the analysis of the Hurricane Irma evacuation, data was collected, cataloged, and processed for a 36-day period beginning August 27, 2017 and ending October 1, 2017. The analysis of Hurricane Michael encompasses the same locations and included a 14-day period that began on October 1, 2018 and concluded on October 14, 2018.

The evacuation analysis focuses on four general regions: Naples, the Florida Keys, and Southeast Florida were analyzed during Hurricane Irma and regions of North Florida were investigated during Hurricane Michael. Naples and the Florida Keys were included in the analysis because Hurricane Irma made landfall in both regions. Southeast Florida was included in the analysis because this region of Florida is the most heavily populated and was directly in the path of Hurricane Irma. Unlike Irma, Hurricane Michael showed a consistent and ultimately accurate storm path projection, leading to the evacuation being focused on the Panhandle region. For this reason, only one analysis area was investigated for Hurricane Michael.

The data collection sites were selected to encompass each of the four regions, similar to the way a cordon line identifies the inner and outer limits of an area. A cordon line is an imaginary line drawn around a study area. Traffic data is collected at roads which cross the cordon line. These locations and analysis regions were provided in Figure 1. Given the relative location of each count station, directional counts were classified as “inbound,” into the region, or “outbound,” out of the region. Drawing a cordon line around a major city, a net increase in the number of inbound vehicles would be expected in the morning, while the opposite would be expected in the afternoon, for a typical commute. As such, it should also be expected that the number of vehicles entering the region in the morning should be approximately equal to the number exiting in the evening. A failure to maintain this equilibrium would result in an overall net increase or decrease of vehicles within the cordoned area. However, during an evacuation, this pattern is broken resulting in the number of vehicle exits significantly outnumbering vehicle entries.

Figure 1.

Florida Department of Transportation data collection sites and analysis regions.

Evacuation Analysis

Fundamentally, the change in the number of vehicles within a defined cordon boundary can be measured by adding the number of vehicles crossing a cordon line into the area and subtracting the number of vehicles exiting. This simple method can determine the change in the number of vehicles within the boundary area. By establishing a cordon line around an evacuating city or region, it is possible to count the number of vehicles entering and exiting the study area. The number of evacuating vehicles can be estimated by calculating the net change in vehicles crossing the cordon boundary over a period of time. Let the number of vehicles entering an evacuation area A from location i along the cordon line for area A, over time interval t, be represented by ${IN}_{it}^{A}$ . Likewise, let the number of vehicles exiting A at i, during t, be represented by the variable ${OUT}_{it}^{A}$ . The start of the evacuation is noted as τ and the recovery time, after the evacuation and reentry of A, as T. The net change in vehicles can be calculated at any time t, as $Δ_{t}^{A}$ in Equation 1:

Δ_{t}^{A} = \sum_{i = 1}^{I} ({IN}_{it}^{A} - {OUT}_{it}^{A})

(1)

In general, daily commuting patterns tend to result in approximately the same number of vehicles entering and exiting a region during a 24 h period $\sum_{t = 1}^{24} Δ_{t}^{A} ≅ 0$ . While seasonal variations or special circumstances often occur that violate this assumption, the daily equilibrium tends to remain relatively in balance. Determining the approximate time an evacuation begins (τ) and recovery ends (T) has been a significant challenge for emergency managers. As the traffic pattern changes over time, however, the imbalance caused by the evacuation in favor of outbound vehicles becomes evident, that is, $\sum_{t = 1}^{24} Δ_{t}^{A} < 0$ . While it remains difficult to estimate the precise time at which the evacuation begins and recovery ends, because of the stochastic nature of driving patterns and behaviors, this research shows, to the hour, when the traffic pattern deviated from a typical commuting regimen. For example, Figure 2 represents the cumulative net change in vehicles crossing the cordon line with a thick, dashed line. In a typical situation, this dashed line should return to near zero, daily. That would represent approximately the same number of entering vehicles as exiting vehicles in any 24 h period. The thick dashed line in Figure 2, however, does not return to zero. It instead takes a persistent negative value beginning at approximately 09:00 on September 5, 2017 until landfall. The return to a near zero net change in traffic does not occur until September 17, 2017 at 16:00, several days after landfall. Therefore this research defines the start of the auto-based evacuation τ and the recovery time (the end of the reentry) T as the start and end times corresponding to a net loss in vehicles that is inclusive of the hurricane’s landfall time, $t_{l}$ .

Figure 2.

Evacuation and reentry traffic analysis for Naples, FL.

The total number of evacuating vehicles for area A is calculated as the minimum value of the cumulative $Δ_{t}^{A}$ . The clearance point of the auto-based evacuation ( $t_{cp}$ ) is the time at which the cumulative $Δ_{t}^{A}$ reaches its minimum value (i.e., when the maximum number of evacuating vehicles have exited the cordoned area). For a hurricane evacuation, the clearance point typically occurs before or at landfall ( $τ < t_{cp} \leq t_{l}$ ). The clearance time ( $t_{ct}$ ) is the duration between the start of the evacuation and the clearance point ( $t_{ct} = t_{cp} - τ$ ). The peak evacuation traffic is seen when $Δ_{t}^{A}$ reaches a minimum value. The peak evacuation hour, $t_{p}$ , is the hour that sees $Δ_{t}^{A}$ reach a minimum value. This minimum can then be considered the peak evacuation exit volume of the area. Evacuation peak demand flow rate and evacuation peak hour factor can also be calculated, if detectors report 15 min count intervals or shorter. Likewise, the same is true for reentry vehicles. The reentry reaches its peak hour ( $t_{r}$ ) when $Δ_{t}^{A}$ reaches a maximum value. A similar calculation can also be commuted to determine a reentry peak hour factor, as.

By considering the maximum value of the cumulative $Δ_{t}^{A}$ as 100% of the auto-base evacuation demand, then $t_{ct}$ represents the clearance time for 100% of the auto-based evacuees. It is therefore possible to estimate the clearance time for any proportion of the auto-based evacuation. For example, the clearance time corresponding to 90% of the auto-based evacuation $t_{ct 90}$ is the time at which 90% of the cumulative $Δ_{t}^{A}$ minimum is achieved. In this fashion, it is possible to estimate vehicle exit rates and if travel time (TT) data is available, these exit rates could be adjusted to estimate vehicle-loading rates (LR) by subtracting out the average travel time ( $t_{ct} - T T_{ct} = L R_{ct})$ .

Methodological Limitations

There are several limitations to the proposed methodology, which warrant discussion before presenting the research findings. The first is that the proposed method is based on the assumption that in any 24 h period the total number of inbound trips should approximately equal the total number of outbound trips. While the total number of inbound and outbound trips can generally be assumed equal, this does not always occur within a 24 h cycle. For example, a tourist may travel into a region for several days before exiting. This would result in a net increase in vehicles for several days before canceling out. This is particularly problematic in Florida, which has many desirable tourist destinations. This leads to the next limitation of this study, which is that the methodology cannot classify trips by purpose. The current method assumes that any vehicle that exits the cordon and does not return is an evacuee. For example, a tourist who entered the cordon line before the data collection period and later leaves without returning is counted as an evacuee. This deficiency within the methodology is somewhat mitigated by extending the data collection period, which was done for this study. Another limitation to the methodology results from the intermingling of evacuation origins and destinations. This occurs when a vehicle evacuates from one cordoned area and travels into another cordoned area. The cordoned area from which the vehicle departed would correctly count this vehicle as an evacuee (−1). However, the area to which the vehicle evacuated would count this vehicle as a net gain (+1) and that would detract from the total number of evacuees departing this area. Furthermore, evacuees who select destinations within their origin cordon area are not counted. The methodology is also limited by the availability of data collection sites. Three out of the four cordon lines used in this study are not true cordon lines, as several low volume roads were not included in the analysis. However, as the data collection sites were provided by FDOT, the major highways entering and exiting the cordoned regions were included and with them the vast majority of vehicles.

Results

The results focused on the development and analysis of figures that show the hourly net change in vehicles ( $Δ_{t}^{A}$ ) and the cumulative net change in vehicles $(\sum_{t = 1}^{T} Δ_{t}^{A})$ for the Florida communities affected by hurricanes Irma and Michael. These figures were used to determine the total number of evacuating vehicles $(\sum_{t}^{t_{cp}} Δ_{t}^{A})$ , start of the auto-based evacuation (τ) and end of the recovery period (T), clearance point ( $t_{cp})$ , peak evacuation volume $(Δ_{t}^{A}$ minimum) and hour $(t_{p})$ , as well as the peak reentry volume $(Δ_{t}^{A}$ maximum) and hour $(t_{r})$ . The figures also show evacuation orders and time of landfall. The results further discussed the development of evacuation time estimate curves, which show the cumulative percent evacuating each region over time. From these curves, it was possible to estimate the 90% clearance time $t_{ct 90},$ 50% clearance time $t_{ct 50}$ , and so forth. This figure was also used for the calculation of the 50th percentile displacement time—the number of hours that at least 50% of the evacuated vehicles remained outside the cordoned area.

Evacuation Figures

Figure 2 shows the evacuation and reentry traffic resulting from the Hurricane Irma evacuation of the Naples, FL region. The primary y-axis displays $Δ_{t}^{A}$ , the number of evacuated vehicles hourly with a solid line. The secondary y-axis displays the cumulative number of evacuating vehicles for all periods $, \sum_{t = 1}^{T} Δ_{t}^{A}$ , with a dashed line. When this value turns negative because of evacuation, it can be seen dropping below the x-axis in the figure. The x-axis is time, in hours. Landfall, $t_{l}$ , is called out in the figure as well as the first mandatory evacuation order given for the region. The figure shows a typical example week of traffic to demonstrate the disparity between the evacuation and routine conditions. In general, the daily traffic shows a morning peak of traffic entering the region ( ${IN}_{it}^{A} > {OUT}_{it}^{A}$ ) and an afternoon peak where vehicles are leaving the region ( ${IN}_{it}^{A} < {OUT}_{it}^{A}$ ). Traffic shows net losses in the number of vehicles before landfall and net increases, after landfall, representing the evacuation and reentry, respectively. The maximum traffic demand periods during both the evacuation and reentry are shown on the figure as the peaks and valleys of the evacuation traffic line. It is important to note that the cordon line, which encircled the Naples region, did not constitute a true cordon, as data for many smaller roads were not available. However, the cordon likely captures the majority of evacuees.

Naples saw a net decrease of 123,202 vehicles in the days leading up to the storm. According to the traffic data, the evacuation of Naples began around 9:00 a.m. on Tuesday September 5, 2017. This suggests that the vehicular evacuation of the Naples region began approximately 75 h before the first mandatory evacuation orders were given and 126 h before landfall on Marco Island. The figures suggest that over 48,000 vehicles had exited the Naples region before mandatory evacuation orders were given. That is to say, 38% of the vehicles that would eventually leave the region did so before governmental directives. The analysis also suggests that the evacuation took 122 h to complete, concluding just 4 h before landfall. Reentry into the Naples region began 2 h after landfall, peaking the following day at 1:00 p.m. (22 h after landfall). Ultimately, traffic did not return to pre-storm levels until September 17, 2017 at 4:00 p.m., over one week after landfall.

The figures for the Florida Keys are shown in Figure 3. Unlike the other regions, this portion of the Florida Keys has only one primary evacuation route and therefore the analysis represents data collected from only one detector location. The analysis found that 40,731 vehicles crossed the cordon line, not to return until after the storm. The evacuation began at approximately 1:00 p.m. on September 5, 2017, 2 h before the mandatory evacuation order was given for regions of Monroe County and 116 h before landfall (on Cudjoe Key, September 10, 2017 at 09:00). The evacuation peaked 25 h after the mandatory orders were announced and concluded 102 h later (12 h before landfall on Cudjoe Key). The reentry began 2 h after landfall, peaking on September 18, 2017 at noon (195 h after landfall), and took 503 h (nearly 3 weeks) to complete.

Figure 3.

Evacuation and reentry traffic analysis for Florida Keys.

Figure 4 shows the evacuation of Southeast Florida. This cordon line included nine detector locations along the major highways and freeways exiting the region. Again, it was not possible to conduct a true cordon, as many lower capacity streets were not available for analysis. Southeast Florida saw 276,052 vehicles leave the area along the observed routes in the days leading up to the storm. The first mandatory evacuation orders were given for regions of Miami-Dade County on September 7, 2017 at noon. The evacuation began just 3 h later (93 h before landfall on Cudjoe Key) and peaked 66 h before landfall. The net egress of the South Florida region concluded on September 8, 2017 at 5:00 p.m. (40 h before landfall). The figure shows that at this point 276,052 more vehicles had exited the region than entered. However, in the period between this clearance and landfall, 20,282 more vehicles (7.35%) entered the region than exited. That is to say, after the cumulative change in volume reached its minimum value before landfall, over 20,000 more vehicles traveled into and stayed in Southeast Florida, than exited during this 40 h period. This was likely because the storm path projection changed from the east coast of Florida to the west coast, causing this region to serve as an evacuation destination for some travelers. After landfall, vehicles almost immediately began to flow into the region, with reentry traffic peaking 31 h after landfall. Traffic in the region did not return to pre-storm levels for nearly eight days (223 h after landfall).

Figure 4.

Evacuation and reentry traffic analysis for Southeast Florida.

Figure 5 shows the evacuation from Hurricane Michael in the North Florida region. The cordon line used in the analysis consisted of seven detector locations on the major exit routes of the area. Severe damage to the power grid resulted in the loss of service to many of the data collection sites, leading up to and after the storm’s landfall. Detector failure began at midnight on October 10, 2018 and continued (off and on) until the data collection period ended. This period is shown in the figure as a yellow overlay depicting times of poor data quality. The detector failure began 13 h before landfall, at which time 16,370 vehicles had exited the region. The remaining detectors indicated that 18,302 vehicles had exited before landfall. The evacuation began on October 7, 2018 at 7:00 a.m., 27 h before the first mandatory evacuation orders went into effect. The evacuation traffic peaked 9 h after the mandatory order and 41 h before landfall. Because of the detector error, it was not possible to determine the exact time of the clearance point. However, the figures suggest this may have occurred just 2 h before landfall. No estimate for the evacuation reentry time could be determined, as the data collection period of 14 days concluded before the evacuating vehicles could return.

Figure 5.

Traffic analysis of evacuation from North Florida region caused by Hurricane Michael.

Table 1 provides a summary of the number of vehicles evacuated along the study routes, as well as the dates and times the evacuation began, peaked, and concluded. The table also shows the dates and times corresponding to reentry (beginning, peaking, and concluding), landfall, and first evacuation orders within the respective study regions. The table shows that Southeast Florida experienced the largest net loss in vehicles. This was expected as this region has the highest population and was likely to see the greatest number of evacuees. In general, the evacuations began several days before the storm made landfall. The Florida Keys, Southeast Florida and the Panhandle saw the peak evacuation hour two to three days in advance of the landfall. This is a significant finding because it suggests that hurricane warnings and evacuation notification were taken seriously. However, Naples did not experience peak demand until 28 h before the storm arrived. Again, this was likely because of the shifting storm track. Naples and Southeast Florida had similar recovery times of just over a week. The Florida Keys required more than 20 days for the traffic patterns to recover. This was likely because the keys were the hardest hit and access was restricted to the lower keys for nearly three weeks.

Table 1.

Summary of Evacuation (Evac.) Time, Orders, and Reentry

Region	Total vehicles	Evac. begins (τ)	Evac. ordered	Evac. peak (t_p)	Evac. ends (t_ct)	Landfall	Reentry begins (τ)	Reentry peak (t_p)	Reentry ends (T)
Florida Keys	39,804	9/5/17 13:00	9/5/17 15:00	9/6/17 16:00	9/9/17 21:00	9/10/17 9:00	9/10/17 11:00	9/18/17 12:00	10/1/17 10:00
Southeast Florida	269,646	9/6/17 12:00	9/6/17 9:00	9/7/17 15:00	9/8/17 17:00	9/10/17 9:00	9/10/17 9:00	9/11/17 18:00	9/18/17 7:00
Naples, FL	123,389	9/5/17 9:00	9/8/17 12:00	9/9/17 11:00	9/10/17 11:00	9/10/17 15:00	9/10/17 17:00	9/11/17 13:00	9/17/17 16:00
Panhandle (Michael)	16,802^*	10/7/18 7:00	10/8/18 10:00	10/8/18 19:00	N/A^*	10/10/18 12:30	N/A^*	10/11/2018 15:00^*	N/A^*

Note: N/A* = Based on incomplete data.

Evacuation and Reentry Time Estimates

An evacuation time estimate provides the approximate time required to evacuate a proportion of the population. The prior analysis measured the time to evacuate 100% of the observed vehicles. Based on this, it was possible to estimate the time required to evacuate 50%, 90%, and 99% of the vehicles, as well. Furthermore, the data provided by the evacuation reentry allows for a similar analysis of the proportion of vehicles returning to the region. From there, it is possible to estimate the how long proportions of the evacuees were displaced. Figure 6 shows the evacuation time estimates for the four study regions. The y-axis shows the cumulative percent of vehicles exiting the cordoned area as a proportion of observed vehicles. The x-axis shows the number of hours in reference to landfall. Negative values indicate times before landfall, whereas positive numbers are post landfall within the respective region. The period leading up to landfall shows the evacuatio; the period after landfall shows the reentry of vehicles into the region.

Figure 6.

Estimate of evacuation and reentry times for the four study regions.

From this figure, the evacuation clearance time may be estimated for any cumulative percent evacuated. For example, the time needed to evacuate 50% of the residents of the Florida Keys was 33 h. Likewise, 99% of evacuees in the Naples region were able to clear the area within 107 h, as compared with the last 1%, which required an additional 18 h. The time of the official evacuation orders is also called out in the figure, along with the proportion of the vehicles which had exited the region before this order. For example, for the Panhandle, evacuation orders were issued approximately 50 h before landfall. At this time, over 8% of the evacuating vehicles had already left the region. Naples also issued evacuation orders approximately 50 h before landfall. However, over half of the vehicles used in the evacuation had exited the region before this order. While it cannot be known for certain, it is likely that many of these earlier evacuees were tourists or other transient populations. Figure 6 also presents a comparison of the exiting rate and by extension the loading rate for each region. The figure suggests that Southeast Florida mobilized quickly as compared with the Florida Keys and Naples regions. However, regions showing slower mobilization began comparatively earlier. The figure also allows for an analysis of the 50th percentile displacement time. Looking at the time lapse between when 50% of vehicles evacuated and when 50% of vehicles returned, provides insight into how long the typical evacuee was displaced. The figure clearly shows extended reentry times for the Florida Keys, which experienced severe damage resulting from the storm requiring curfews and travel restrictions in the lower keys ( 7 ).

Table 2 provides the clearance and reentry times for 50%, 90%, 99%, and 100% of evacuees for each region. The table also provides the 50th percentile displacement time. The table shows that Naples and the Florida Keys had the longest clearance times. The Keys also experienced the longest reentry time. The evacuation tail, generally considered the last 10% of evacuees to exit the region, was also found to be the longest for the Keys and Naples at 32 h and 25 h, respectively. It is not likely coincidental that these two regions were also directly hit by the storm. Southeast Florida had significantly shorter clearance times and evacuation tail, despite evacuating more vehicles. This was likely because these areas have more high capacity roads and freeways and their evacuations started much later when compared with the other regions. The evacuation of the Panhandle in response to Hurricane Michael was also comparatively shorter, and saw an 11 h evacuation tail.

Table 2.

Summary of Evacuation Analysis Results

Region	Clearance time (hours)				Reentry time (hours)				50th percentile displacement time (days, hours)
Region	50%	90%	99%	100%	50%	90%	99%	100%	50th percentile displacement time (days, hours)
Florida Keys	33	72	91	104	193	424	501	503	278 (11d, 14h)
Southeast Florida	27	47	54	59	64	173	189	190	130 (5d, 10h)
Naples, FL	71	100	107	125	63	140	167	167	120 (5d, 0h)
Panhandle (Michael)	49	66	74	77	NA	NA	NA	NA	NA

Half of the population of Southeast Florida that evacuated by vehicle did so within 31 h after the evacuation began. However, it was not for another 130 h that half of the population reentered. Therefore, the average evacuee from Southeast Florida was displaced for over five days. Using this same approach, 50% of the Naples auto-based evacuees were displaced for 120 h as well. The displacement time for the 50th percentile of the auto-based evacuees for Florida Keys residents was 278 h, over 11.5 days. It should also be noted that the evacuation reentry was more gradual than the evacuation itself.

Conclusion

The perceived success of an evacuation, or lack thereof, often is based on media reports, anecdotal observation or, worse, rumors and social media discussion. In reality, a highly effective evacuation could be assumed a failure because of a few limited but highly visible areas of congestion. This has suggested the need for a better way to describe and assess large statewide evacuations in more systematic and objective ways. Unfortunately, this is not easy to accomplish because there are few, if any, data records or performance measures generated that accurately and effectively describe the conditions of these events. In fact, there is no standardized methodology to quantify the characteristics of an evacuation that is transferable and repeatable between state departments of transportation.

Fortunately, there are many commonly used data measures for analyzing routine transportation conditions. The intent of this work was to adapt and apply them to develop a method capable of describing mass evacuations. In fact, these methods can also be applied to describe evacuation reentry traffic patterns; a historically lightly studied area in practice and research. Results of the application of the research methodology showed that the evacuations from hurricanes Irma and Michael began several days before landfall. Vulnerable residents in the Florida Keys started their evacuations five days before Hurricane Irma’s landfall, with nearly 20% departing before the mandatory evacuation order. This observation was unexpected because prior survey results suggested that a two-day loading period was most likely ( 40 ). In general, the evacuations peaked two to three days before landfall and between the hours of 8:00 a.m. and 3:00 p.m., confirming prior research that suggested a preference for morning departures ( 41 ). From an emergency preparedness standpoint, these trends are positive and suggest an increased civic awareness of hazard risk perception. The research also found that half of the auto-based evacuees from Southeast Florida and the Naples region were displaced for up to five days. The 50th percentile displacement time for Florida Keys residents who evacuated by car saw significantly longer displacement times of over 11 days.

The results of this effort show that the proposed methods could be quite effective to identify key parameters of the evacuation in a systematic and repeatable manner. The metrics identified in this paper could be used to evaluate the effectiveness of evacuation messaging and infer behavioral responses of evacuees and interpretation of threat. These methods are also beneficial to researchers seeking direct observation values for developing and validating evacuation decision models or more sophisticated evacuation performance parameters. The long-term contribution of this work is likely to be in the analysis of future and past evacuation events, as these parameters may be used to compare evacuation responses using an empirical method. This research provides a system for state departments of transportation and emergency management officials to analyze future auto-based evacuations. The method also facilitates parametric comparisons between evacuation events, an area needed to continue to evolve and improve evacuation practice. Standardized measures for hurricane evacuations are needed to facilitate systematic evaluations of performance. Future researchers could build on methods presented here to develop a level-of-service (LOS) analysis for emergency evacuations. This would be similar to the way the Highway Capacity Manual uses the standardized collection and processing of freeway densities for its LOS evaluations. With additional research, the methods laid out in this paper could also lead to a more comprehensive understanding of evacuation traffic processes and behavioral responses to improve their planning and management.

The proposed analysis procedure did not attempt to investigate shadow evacuees. However, it would be possible to do so, if additional cordon lines could be drawn around evacuation zones. The difference between the number of evacuating vehicles between the “inner” and “outer” cordon lines would be the number of shadow evacuation vehicles within the area between the cordon lines. Also, the proposed approach classifies any vehicle exiting the region without returning before landfall to be an evacuee. This classification approach is blurred by the persistence of background traffic both before and after landfall. The presence of tourist or other transient populations, who may evacuate earlier than residents, also affects evacuation estimates.

The authors confirm contribution to the paper as follows: study conception and design: S. Parr, P. Murray-Tuite, B. Wolshon; data collection: L. M. Acevedo; analysis and interpretation of results: S. Parr, L. M. Acevedo, P. Murray-Tuite, B. Wolshon; draft manuscript preparation: S. Parr, L. M. Acevedo, P. Murray-Tuite, B. Wolshon. All authors reviewed the results and approved the final version of the manuscript.

Footnotes

Acknowledgements

The authors acknowledge the technical support provided by the Florida Department of Transportation, namely Joey Gordon and Steve Bentz. Portions of this paper were adapted from Embry-Riddle Aeronautical University graduate student and co-author Lorraine Acevedo’s Master’s thesis.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was supported by a grant from by the United States Department of Transportation through the Gulf Coast Center for Evacuation and Transportation resiliency, a member of the University of Texas University Transportation Center; Cooperative Mobility for Competitive Megaregions (CM2) Grant Number: USDOT 69A3551747135.

ORCID iDs

Scott A. Parr

Lorraine M. Acevedo

Pamela Murray-Tuite

References

Marshall

. 4 Maps That Show the Gigantic Hurricane Irma Evacuation. Wired. Septmeber 9, 2017. Accessed August 15, 2018. https://www.wired.com/2017/09/4-maps-show-gigantic-hurricane-irma-evacuation/.

Savaransky

. Florida Governor Declares State of Emergency over Hurricane Irma. The Hill. 2017. https://thehill.com/homenews/news/349143-florida-governor-declares-state-of-emergency-over-hurricane-irma. Accessed August 14, 2018.

IRMA Graphics Archive: 5-Day Forecast Track, Initial Wind Field and Watch/Warning Graphic . U.S. National Oceanic and Atmospheric Administration (NOAA), National Hurricane Center, 2017. https://www.nhc.noaa.gov/archive/2018/IRMA_graphics.php?product=5day_cone_with_line_and_wind.

Jansen

. Timeline: Hurricane Irma’s Progress to Monster Storm. Gannett Satellite Information Network. USA Today, 2017. www.usatoday.com/story/news/2017/09/10/timeline-hurricane-irma-fluctuating-strgrowing-stronger-weaker-crashed-into-caribbean-islands-florid/651421001/.

O’Connor

. Florida’s Hurricane Irma Recovery: The Cost, the Challenges, the Lessons. Insurance Journal, November 30, 2017. www.insurancejournal.com/news/southeast/2017/11/30/472582.htm.

Wile

. The Estimated Costs of Hurricanes Irma and Harvey are Already Higher Than Katrina. Mondy. 2017. https://money.com/hurricane-irma-harvey-economic-cost/. Accessed August 15, 2018.

Associated Press. Curfew Lifted in Florida Keys 3 Weeks After Hurricane Irma. U.S. News & World Report, 2017. www.usnews.com/news/best-states/florida/articles/2017-10-02/curfew-lifted-in-florida-keys-3-weeks-after-hurricane-irma.

Beven

J. L.

II Berg

Hagen

. National Hurricane Center: Tropical Cyclone Report: Hurricane Michael (AL142018). U.S. National Oceanic and Atmospheric Administration (NOAA), National Hurricane Center, Miami, FL, 2019.

Roberson

. Hurricane Michael: First Florida Evacuations Ordered in Gulf County, Others in Panhandle. 2018. https://www.tallahassee.com/story/news/2018/10/08/hurricane-michael-florida-evacuation-gulf-county-panhandle-wakulla-bay-mandatory/1567904002/.

10.

U.S. National Oceanic and Atmospheric Administration (NOAA), National Hurricane Center. Michael. Graphics Archive: 5-Day Forecast Track, Initial Wind Field and Watch/Warning Graphic. 2018. https://www.nhc.noaa.gov/archive/2018/MICHAEL_graphics.php?product=5day_cone_with_line_and_wind.

11.

Haddad

. List of Mandatory Evacuation Zones in Florida Ahead of Hurricane Michael. October 10, 2018. https://www.clickondetroit.com/weather/hurricane/list-of-mandatory-evacuation-zones-in-florida-ahead-of-hurricane-michael.

12.

Kumar

Ukkusuri

S. V

. Utilizing Geo-Tagged Tweets to Understand Evacuation Dynamics During Emergencies: A Case Study of Hurricane Sandy. Paper presented at the Companion Proceedings of the The Web Conference, Lyon, France, 2018.

13.

Roy

K. C.

Hasan

. Modeling the Dynamics of Hurricane Evacuation Decisions from Twitter Data: An Input Output Hidden Markov Modeling Approach. Transportation Research Part C: Emerging Technologies, Vol. 123, 2021, p. 102976.

14.

Marasco

Murray-Tuite

Guikema

Logan

Time to Leave: An Analysis of Travel Times During the Approach and Landfall of Hurricane Irma. Natural Hazards, Vol. 103, No. 2, 2020, pp. 2459–2487.

15.

Yin

Chen

Zhang

Yang

Wan

Ning

. Improving Emergency Evacuation Planning with Mobile Phone Location Data. Environment and Planning B: Urban Analytics and City Science Vol. 47, No. 6, 2020, pp. 964–980.

16.

Yabe

Sekimoto

Tsubouchi

Ikemoto

. Cross-Comparative Analysis of Evacuation Behavior after Earthquakes Using Mobile Phone Data. PLoS One Vol. 14, No. 2, 2019, p. e0211375.

17.

Martín

Cutter

S. L.

. Bridging Twitter and Survey Data for Evacuation Assessment of Hurricane Matthew and Hurricane Irma. Natural Hazards Review, Vol. 21, No. 2, 2020, p. 04020003.

18.

Hong

Frias-Martinez

. Modeling and Predicting Evacuation Flows During Hurricane Irma. EPJ Data Science, Vol. 9, No. 1, 2020, p. 29.

19.

Archibald

McNeil

. Learning from Traffic Data Collected Before, During and After a Hurricane. IATSS Research, Vol. 36, No. 1, 2012, pp. 1–10.

20.

Dow

Cutter

S. L

. Emerging Hurricane Evacuation Issues: Hurricane Floyd and South Carolina. Natural Hazards Review, Vol. 3, No 1, 2002, pp. 12–18.

21.

H.-C.

Lindell

M. K.

Prater

C. S

. Logistics of Hurricane Evacuation in Hurricanes Katrina and Rita. Transportation Research Part F: Traffic Psychology and Behaviour, Vol. 15, No. 5, 2012, pp. 445–461.

22.

Murray-Tuite

Yin

Ukkusuri

Gladwin

Changes in Evacuation Decisions between Hurricanes Ivan and Katrina. Transportation Research Record: Journal of the Transportation Research Board, 2012. 2312: 98–107.

23.

Lindell

M. K.

Prater

C. S

. Critical Behavioral Assumptions in Evacuation Time Estimate Analysis for Private Vehicles: Examples from Hurricane Research and Planning. Journal of Urban Planning and Development, Vol. 133, No. 1, 2007, pp. 18–29.

24.

Vogt

B. M.

Sorensen

J. H

. Evacuation Research: A Reassessment. Oak Ridge National Laboratory, U.S. Department of Energy, TN, 1992.

25.

Wolshon

. Empirical Characterization of Mass Evacuation Traffic Flow. Transportation Research Record: Journal of the Transportation Research Board, 2008. 2041: 38–48.

26.

Ozbay

Bartin

Iyer

Carnegie

J. A

. Empirical Evacuation Response Curve During Hurricane Irene in Cape May County, New Jersey. Transportation Research Record: Journal of the Transportation Research Board, 2013. 2376: 1–10.

27.

Ozbay

. Hurricane Irene Evacuation Traffic Patterns in New Jersey. Natural Hazards Review, Vol. 16, No. 2, 2014, p. 05014006.

28.

Ozbay

Bartin

. Effects of Hurricanes Irene and Sandy in New Jersey: Traffic Patterns and Highway Disruptions During Evacuations. Natural Hazards, Vol. 78, No. 3, 2015, pp. 2081–2107.

29.

Ghorbanzadeh

Burns

Rugminiamma

L. V. N.

Ozguven Erman

Huang

. Spatiotemporal Analysis of Highway Traffic Patterns in Hurricane Irma Evacuation. Transportation Research Record: Journal of the Transportation Research Board, 2021. https://doi.org/10.1177/03611981211001870.

30.

Menon

Staes

Bertini

R. L

. Measuring Transportation Network Performance During Emergency Evacuations: A Case Study of Hurricane Irma and Woolsey Fire. University of South Florida, Tampa; Center for Transportation, Equity, Decisions & Dollars (CTEDD), Office of the Assistant Secretary for Research and Technology, Arlington, TX, 2020, p. 117.

31.

Staes

B. M.

Bertini

R. L.

Menon

Yuksel

. Examining Freeway Bottleneck Features During a Mass Evacuation. Transportation Research Record: Journal of the Transportation Research Board, 2021. 036119812110035.

32.

Zhu

Y.-J.

Collins

J. M

. Estimating Road Network Accessibility During a Hurricane Evacuation: A Case Study of Hurricane Irma in Florida. Transportation Research Part D: Transport and Environment, Vol. 83, 2020, p. 102334. https://trid.trb.org/view/1696912.

33.

Wong

S. D.

Pel

A. J.

Shaheen

S. A.

Chorus

C. G

. Fleeing from Hurricane Irma: Empirical Analysis of Evacuation Behavior Using Discrete Choice Theory. Transportation Research Part D: Transport and Environment, Vol. 79, 2020, p. 102227. https://trid.trb.org/view/1681411

34.

Sharma

Faruk

M. O.

El-Urfali

. Operational and Safety Impact Analysis of Implementing Emergency Shoulder Use (ESU) for Hurricane Evacuation. Transportation Research Record: Journal of the Transportation Research Board, 2020. 2674: 282–293.

35.

Wolshon

. The Role of Transportation in Evacuation and Reentry: A Survey of Practice. Journal of Transportation Safety & Security, Vol. 1, No. 3, 2009, pp. 224–240.

36.

Lindell

M. K.

Kang

J. E.

Prater

C. S

. The Logistics of Household Hurricane Evacuation. Natural Hazards, Vol. 58, No. 3, 2011, pp. 1093–1109.

37.

Zhang

Wolshon

Herrera

Parr

. Assessment of Post-Disaster Reentry Traffic in Megaregions Using Agent-Based Simulation. Transportation Research Part D: Transport and Environment, Vol. 73, 2019, pp. 307–317.

38.

Siebeneck

L. K.

Lindell

M. K.

Prater

C. S.

H. C.

Huang

S. K

. Evacuees’ Reentry Concerns and Experiences in the Aftermath of Hurricane Ike. Natural Hazards, Vol. 65, 2013, pp. 2267–2286.

39.

Siebeneck

L. K.

Cova

T. J

. An Assessment of the Return Entry Process for Hurricane Rita 2005. International Journal of Mass Emergencies and Disasters, Vol. 26, 2008, pp. 91–111.

40.

Baker

E. J

. Florida Statewide Regional Evacuation Study Program: Volume 2-11 South Florida Region Regional Behavioral Analysis. Florida Division of Emergency Management, South Florida Regional Planning Council, Tallahassee, FL, 2010. http://www.sfrpc.com/SRESPWeb/Vol2-11.pdf. Accessed July 19, 2019.

41.

Lindell

M. K.

Murray-Tuite

Wolshon

Baker

E. J

. Large-Scale Evacuation: The Analysis, Modeling, and Management of Emergency Relocation from Hazardous Areas. Taylor and Francis, New York, 2019.