How Various Natural Disasters Impact Urban Human Mobility Patterns: A Comparative Analysis Based on Geotagged Photos Taken in Tokyo

Abstract

This study analyzed the impacts of various extreme natural events that affected Tokyo’s human movement patterns between 2008 and 2019 based on geotagged photos shared online. First, we selected six disasters of different types according to severity, damage, and photo count. Next, we delineated three phases representing steady and perturbed conditions (before, during, and after) for each extreme event based on relevant weather measurements. We then analyzed human mobility patterns via two indicators: displacement and mean squared displacement (MSD). A transfer-learning-based convolutional neural network (CNN) model was also developed to classify the photos according to whether they were taken indoors or outdoors. Thus, we analyzed the characteristics of people’s trips within and between the two environments. The results show that, while all extreme events perturbed mobility patterns to different degrees, these patterns mostly followed a truncated power-law distribution during steady and unsteady states.

1 Introduction

Earthquakes, tropical cyclones, and floods pose constant threats to urban communities. Given their growing population densities and interdependent infrastructures, metropolitan areas suffer most from such events in terms of human casualties, infrastructural damage, and economic losses (Gu 2019). For instance, the New York metropolitan area, which accounts for 8% of the US’s economic output, was hit by Hurricane Sandy on Monday, October 29, 2012. 48 lives were lost in New York City; there were 72 deaths across eight states; and economic losses climbed to USD 68 billion (Schelske et al. 2013). Additionally, these disasters have both short- and long-term negative impacts on the population’s daily routines as they affect essential mobility patterns.

Our current understanding of daily human mobility trajectories under steady conditions has improved due to extensive studies on general human movements. During normal conditions, human mobility most likely fits a truncated power-law distribution, with values of the scaling parameter β ranging between 1 and 1.82 (Alessandretti et al. 2017). Studies have also focused on analyzing urban human movements under perturbed states, caused mainly by extreme natural events. Most of these studies have relied on text-based, geotagged records from social networks, generally Twitter because of its freely accessible Application Programming Interface (API) and large amount of data. Fewer studies have used photo-based records collected through photo-sharing platforms, such as Flickr, which provides fewer geotagged records. However, with the advancement of deep-learning techniques, geotagged photos could offer more insight into human mobility patterns during both steady and unsteady conditions.

This study, therefore, investigated and compared urban human mobility patterns before, during, and after six disasters of different types, selected from among the extreme events that impacted Tokyo between July 2008 and December 2019. Urban mobility was evaluated based on geotagged, timestamped Flickr photos by calculating two mobility measurements—displacement and mean squared displacement (MSD)—in addition to determining people’s period of stay for both indoor and outdoor environments and the time spent traveling between them. The research questions were as follows.

• 1. How can human mobility be described during different natural disasters?

• 2. To what extent is human mobility disturbed by different natural disasters?

• 3. Is photo-based, social, geotagged data capable of modeling such mobility?

The originality of this methodology is its use of geotagged photos instead of text-based geodata to explore patterns of urban human movements affected by natural disasters. The development of a convolutional neural network (CNN)-based model to categorize photos according to whether they were taken in indoor or outdoor environments led to investigating variations in stay periods and travel times between the two environments during different periods of each extreme event.

The remainder of this article is organized as follows. Section 9.2 provides a comprehensive literature review. Section 9.3 overviews the study area’s geographical settings, as well as the chronology of the natural disasters that impacted Tokyo between 2008 and 2019. Section 9.4 presents the study’s methodology and describes its key steps. Section 9.5 outlines the obtained results. Section 9.6 discusses the study’s findings, implications, limitations, and suggests directions for future research; and Sect. 9.7 draws conclusions.

2 Literature Review

Social networks present an enormous opportunity as a source of spatial data with rich metadata. Recently, multiple studies analyzing human mobility patterns during unsteady conditions have been carried out with geotagged social data. These studies differ considerably regarding their data sources, approaches, study periods, and spatial wideness of the target area. Text-based geotagged data sources are most popular among researchers, given the large number of posts, but photo-based geodata studies analyzing the impacts of extreme events are relatively rare. This section presents the most relevant previous studies that have employed social data (largely text- and photo-based geotagged data) to analyze human mobility during disastrous events.

Regarding text-based geotagged data, researchers have principally used Twitter to analyze how human patterns are affected by disasters. New Yorkers’ tweets were used to compare their movement trajectories during steady conditions and perturbation states caused by Hurricane Sandy (Wang and Taylor 2014). Though human patterns were significantly perturbed during the hurricane, they also showed high resilience, based on the correlated values for the center of mass and the radius of gyration of people’s movements during both steady and perturbed states. Human mobility also better fit truncated power-law distributions during both states. In the same vein, Ahmouda et al. (2019) focused on how hurricanes impact human mobility patterns in the urban and rural areas of different US coastal regions affected by Hurricanes Matthew (2016) and Harvey (2017). Using tweets from the study regions, the authors compared data before, during, and after each hurricane. While displacements and activities were reduced, displacement patterns followed the truncated power-law at all phases of the events. Wang and Taylor (2016) further expanded and confirmed these results for five other types of natural disasters (typhoons, earthquakes, winter storms, thunderstorms, and wildfires), using geotagged tweets to analyze 15 cases across the globe.

To the extent of our knowledge, no study has analyzed human patterns using geotagged photos during disasters; however, some studies have included geotagged photos for other purposes. Most of these used Flickr as their main data source. Preis et al. (2013) analyzed the number of photos taken during Hurricane Sandy by Flickr users living in New Jersey. A strong correlation was found between atmospheric pressure and the count of uploaded photos related to the hurricane between October 20 and November 20, 2012. They suggested that Flickr might be useful for monitoring collective human attention during large-scale disasters in real-time. Yan et al. (2017) employed geotagged Flickr photos to evaluate tourism recovery in the central Philippines after two disasters in 2013: the Bohol Earthquake (October) and Super Typhoon Haiyan (November). The authors gathered insights regarding the spatiotemporal patterns of recovery conditions and trends. Similarly, Li et al. (2018) used geotagged images to quantify the degree of damage from four natural disasters: Typhoon Ruby (2014), the Nepal Earthquake (2015), the Ecuador Earthquake (2016), and Hurricane Matthew (2016). They developed a CNN-based model to classify image content into three damage classes illustrated as heatmaps: severe, mild, and no damage. Geotagged images were also successfully used as ground truth data to improve the accuracy of the May 2011 Mississippi River flood hazard map, extracted using remotely sensed data and a digital elevation model (Schnebele and Cervone 2013).

3 Study Area: Geographic Settings and Past Disasters

Japan is among the top countries with a high vulnerability to natural disasters (Garschagen et al. 2016; Welle and Birkmann 2015). Its location in the Circum-Pacific Mobile Belt, known for its seismic and volcanic activities, causes frequent earthquakes and tsunamis, which can lead to catastrophes, such as the 2011 Tohoku Earthquake. The Japanese archipelago is also hit by tens of typhoons each year, which form over the Northwest Pacific Ocean, causing rainstorms, heavy rain, and landslides. While typhoons are generally predictable (with Japan’s typhoon season usually lasting from July to October) earthquakes are more elusive.

Tokyo’s 23 special wards (hereinafter called Tokyo) were selected as the target area for this study for several reasons. First, according to the 2015 census, the area is home to over nine million citizens and has a population density of over 1,000 persons/km² (Statistics Bureau 2020). Second, Tokyo is the most popular destination for international tourists. According to the Japan National Tourism Organization’s yearly statistics, approximately half of the international visitors to Japan pass through Tokyo. In 2018, for instance, more than 14 million tourists visited the capital. Third, the study area is part of the Tokyo Metropolitan Area (TMA), which is the heart of Japan’s economy (the third largest in the world). From 2000–2014, the TMA generated 37% of Japan’s GDP growth (OECD 2018). Lastly, the area is among the world’s riskiest cities in which to live. Based on a recent report from Lloyd’s City Risk Index, which estimates the economic output threatened by 18 manmade and natural hazards, Tokyo has the highest risk index across the 279 cities surveyed (Lloyd 2018).

Figure 9.1 shows the chronology of the significant natural disasters that impacted the TMA from 2008–2019, causing deaths, injuries, and significant economic losses. Based on the official Tokyo Metropolitan Government disaster prevention guidebook (Tokyo Metropolitan Government 2019), we identified 22 extreme events in six categories: earthquake, heavy rain, snowstorm, typhoon, rainstorm, and snow/rainstorm. From these 22, we selected six case studies, each representing one of the six categories. Table 9.1 lists the date of occurrence and the damage caused by the selected disasters.

Fig. 9.1

Chronology of the natural disasters that impacted Tokyo between July 2008 and December 2019. Disasters circled in orange are the case studies for this analysis

Table 9.1 The selected extreme natural events with the approximate date of occurrence and the damage (adapted from the Tokyo Metropolitan Government’s disaster prevention guidebook [Tokyo Metropolitan Government 2019])

4 Data and Methods

Figure 9.2 illustrates the study’s methodology. First, we collected Flickr records associated with photos taken 15 days before and after the six selected natural disasters. These records were cleaned and merged with hourly weather data. Based on this weather data, we defined the study periods “before,” “during,” and “after” each disaster. Second, we developed a transfer-learning-based CNN model to classify photos into two categories (indoor and outdoor). We then compared human mobility based on relevant metrics (displacement, MSD, number of trips, etc.) during the disasters. The following sections detail each key step.

Fig. 9.2

The study’s methodological framework

4.1 Data Collection and Preprocessing

The first step of the analysis involved collecting and preprocessing geotagged photos (Fig. 9.2a). First, we developed a python script to collect records from Flickr through its publicly available API. We set a query to obtain only records from Tokyo taken between July 1, 2008, and December 31, 2019. To minimize the effects of “active users,” defined as those who upload a large number of photos (Hollenstein and Purves 2010), which could bias the results in favor of their behaviors (Hu et al. 2015), we kept only one record if multiple photos were taken within the same minute.

We then downloaded Tokyo’s historical hourly weather data from worldweatheronline.com via its accessible API. Weather parameters included, among other factors, temperature, air pressure, humidity, and cloud cover. These data were used to delineate the start and end dates for each of the three phases in each disaster.

Of the 22 selected disasters, we chose six case studies according to the availability of Flickr samples for each of the disasters’ three phases. Table 9.2 lists the case studies, the start/end dates of each phase, and the corresponding Flickr sample count. It should be noted that, for the Tohoku Earthquake, we considered the major 7.2M_w foreshock from March 9, 2011, as the start of the “during” period.

Table 9.2 The considered disasters with the estimated start/end dates for their three phases and the Flickr records count for each phase

4.2 Indoor/Outdoor Photo Data Classification

The second key step was identifying the environment in which the photos were taken (indoor or outdoor) (Fig. 9.2b) and thereby determine in what environment users were located during the three phases of each case study. The users’ periods of stay before moving to different environments were also estimated. To accomplish this goal, we developed a CNN-based model to classify photos as indoor or outdoor. For the training and testing process, we used two public datasets: Dense Indoor and Outdoor DEpth (DIODE; Vasiljevic et al. 2019) and Indoor Scene Recognition (Quattoni and Torralba 2009). The former contains thousands of RGB photos of indoor and outdoor scenes, while the latter contains 15,620 indoor photos grouped into 67 categories (e.g., store, home, public spaces). Each dataset was split into training and testing/validation sets. Initially, we developed two models, based on the transfer-learning approach via MobileNetV2 (Sandler et al. 2018) and InceptionV3 (Szegedy et al. 2015). Using the training dataset (80%) and the evaluation dataset (20%), we trained and validated the two models considering a binary cross-entropy loss function. We obtained an accuracy of 94% and 84% and a loss of 0.17 and 0.64 for the MobileNetV2 and InceptionV3 models, respectively. Consequently, we selected the MobileNetV2 model to classify the Flickr images as indoor or outdoor. In addition to its high accuracy, the MobileNetV2 was selected because it optimizes memory consumption and execution time while minimizing prediction errors. The total training time was about 14 h (13 h:59 m:2 s) compared to an approximate 29 h (28 h:58 m:46 s) for the InceptionV3 model. The model training was conducted on a 10-Intel-core i9-10900X with 120 GB RAM and a 3.70 GHz CPU.

4.3 Calculating Human Mobility Parameters and Trip Characteristics

In the third step, we computed parameters describing human mobility for all users before, during, and after each disaster (Fig. 9.2c). Two parameters were considered: displacements and MSD.

Displacement denoted the distance between two consecutive photos taken by an individual. The Haversine formula (Robusto 1957) was used to calculate displacements based on the geographic coordinates of the origin and destination points as follows:

$$ d = 2r \times \arcsin \left( {\sqrt {\sin^{2} \left( {\frac{{\phi_{2} - \phi_{1} }}{2}} \right) + \cos \phi_{1} \cos \phi_{2} \sin^{2} \left( {\frac{{\varphi_{2} - \varphi_{1} }}{2}} \right)} } \right) $$

(9.1)

where $r$ referred to the earth’s radius, which equals approximately 6,371 km; $\phi$ was the latitude; and $\varphi$ was the longitude.

MSD is a concept adapted from statistical physics and Brownian motion (Barbosa et al. 2018), which has been employed across disciplines, such as for hurricane track forecasting (Meuel et al. 2012). MSD was used to calculate the spatial extent of an individual’s displacements in an area, and it was defined by the following formula:

$$ MSD\left( t \right) = \left[ {\left( {r\left( t \right) - r_{0} } \right)^{2} } \right] $$

(9.2)

where $r_{0}$ was the location of the first photo taken at the beginning of the observations; and $r\left( t \right)$ was the location of a photo taken at time $t$.

The two mobility parameters were calculated for a 24-h period beginning at 3:00 AM and ending at 2:59 AM the next day. For the displacements, we fitted the data for the three event phases into four distributions: exponential (EXP), lognormal (LGN), power-law (PL), and truncated power-law (TPL). The goodness of fit was tested and compared for the four distributions. The Maximum Likelihood Estimation (MLE) method was employed to identify the model with the highest goodness of fit using the python package powerlaw (Alstott et al. 2014).

We also investigated the characteristics of trips between or within indoor and outdoor environments. Using the MobileNetV2 model, we predicted the nature of the environment where all sampled images were taken. Consequently, using the timestamp information from the photos, we measured the users’ length of stay in each environment by calculating the period between the last and first photos among consecutive photos taken in a similar environment. We categorized these stay periods into three classes: short (≤10 min), medium (>10 min and <2 h), and long (≥2 h). We also determined the number of trips from indoor to outdoor environments and vice versa.

5 Results

5.1 Temporal Analysis of the Photos’ Spatial Distribution and the Nature of the Environment

Figure 9.3 illustrates the spatial distribution of the collected photos and their density for each phase of the analyzed natural disasters. The figure also shows the percentages of indoor and outdoor images. In all cases, the photos were usually taken near railway stations connected by the Yamanote loop line—mainly the Tokyo, Akihabara, Ueno, Ikebukuro, Shinjuku, and Shibuya stations—as most of Tokyo’s tourist attractions are located near these stations.

Fig. 9.3

Spatial distribution of photos and their density before/during/after the six studied natural disasters (green = before; red = during; blue = after). The numbers and colored, camera-like icons correspond to an example of a path followed by a random user during the three phases. Black dots in the study-area maps represent the main stations located along the Yamanote line. Percentages at the bottom of the maps correspond to the percentages of indoor (I) and outdoor (O) photos taken during each phase

Typhoon Wipha (Fig. 9.3A3) and 2015’s heavy rain (Fig. 9.3A5) had the greatest impact on human movement, evidenced by the small spatial densities and large indoor percentages (>80%) of the photos taken during these events. Photos taken during both events were related to food or taken inside a room or restaurant. This could suggest that the population was well prepared for the events. After Typhoon Wipha, human mobility slowly returned to normal, mainly around the Tokyo and Ikebukuro stations, which can be concluded from the slightly wider photo density in those areas and the increase in the number of photos taken outdoors (+20%). The impacts of the four-day 2015’s heavy rain lasted longer, supported by the limited number of photos taken after the event and the slight increase in outdoor photos (+2%). The photos taken after these incidents showed indoor environments (e.g., restaurants, rooms, etc.), natural scenes, and cultural events. In contrast to Typhoon Wipha and 2015’s heavy rain, the 2010 rainstorm (Fig. 9.3A1) did not heavily impact human movement, which can be concluded from the large densities of photos taken at the time. Only 32% of those photos were taken indoors. This may be explained by the fact that the 2010 rainstorm was neither as severe nor as long as Typhoon Wipha and 2015’s heavy rain.

During snow-related events, a wider spatial extent of photos was observed across the study area, especially during the snowstorm of January 22–23, 2018 (Fig. 9.3A6), while the indoor/outdoor photo percentages show that the event in February 2014 (Fig. 9.3A4) slightly pushed people to stay indoors. However, despite snowy weather, people tended to visit attractions, such as zoos and parks. The snowstorm of 2018 encouraged people to go outside and see Tokyo in such rare, snowy conditions, evidenced by the increased photo density and number of outdoor photos at the time.

The Tohoku Earthquake (Fig. 9.3A2) impacted people’s movement differently than the rain- and snow-related events. Post-earthquake photo clusters were observed across the study area beyond the usual hotspots near the Yamanote line stations. The percentages of indoor and outdoor photos were relatively equal (52% and 48% before; 48% and 52% after, respectively). However, the number of indoor photos increased (63%) during the event. People took photos to document unusual circumstances, such as empty store shelves and water stored in bathtubs.

The insights gathered from this temporal analysis of photos’ spatial distribution and density, coupled with the nature of the environment, show how people reacted to different types of natural disasters. While rainy events generally led to fewer activities, their impacts on human mobility depended on their duration and severity. Snow-related events led to more outdoor activities, which might be attributed to the events’ rareness and lack of severity. During most of the sampled events, the human activities were mainly located near the Yamanote line. However, the Tohoku Earthquake caused a wide number of people to disperse beyond this area.

5.2 Displacements

Figure 9.4 shows the heatmaps for comparing the displacement fit of the four distributions—EXP, LGN, PL, and TPL—for each phase of the analyzed disasters. The loglikelihood ratio $R$ values suggest that, although not statistically significant in most cases at levels below 10%, TPL fit the data more appropriately than PL, LGN, and EXP. This partially coincides with results reported by Wang and Taylor (2014). LGN showed a better fit than other distributions during and after the heavy rain of September 8–11, 2015, and after the snowstorm of January 22–23, 2018. EXP showed the worst fit for almost all case studies.

Fig. 9.4

Heatmaps comparing the displacement fit of the four considered distributions (EXP, LGN, PL, and TPL) for the three phases (during, before, and after) of each analyzed natural disaster. The first value, R, in each cell, corresponds to the loglikelihood ratio between two distributions. A positive number indicates that the displacement data are more likely in the y-axis distribution, while negative numbers indicate the contrary. The second value, labeled p, denotes the significance value. The last value depicts the statistical significance level of the comparison, in which five levels are considered: (*****) = 0.1%; (****) = 0.5%; (***) = 1%; (**) = 5%; (*) = 10%. () is used when a p-value is not significant at all considered levels

The plots in Fig. 9.5 present the complementary cumulative distribution function for the TPL distribution fit of the displacements calculated for each phase of the analyzed events. The equation for the TPL distribution was:

Fig. 9.5

Complementary cumulative distribution functions for the displacements and fitted TPL distribution for the three phases of each analyzed natural disaster: a rainstorm, b the Tohoku Earthquake, c Typhoon Wipha, d snow/rainstorm, e heavy rain, and f snowstorm

$$ P\left( {\Delta d} \right) \propto \Delta d^{ - \beta } e^{ - \lambda \Delta d} $$

(9.3)

where $\Delta d$ was the displacement; $\beta$ was the scaling parameter; and $\lambda$ was the exponential cutoff value. The $\lambda$ values increased during wet-weather-related disasters, which suggests limited trips, while values recorded during and after the Tohoku Earthquake increased. The values of $\beta$ ranged from 1 to 1.92, with a few exceptions returning values above 2. The differences between these values and those reported in the literature could be due to the small sample sizes used in this comparative analysis, as other research has reported that sample size may affect the goodness of fit results (e.g., Wang and Taylor (2016)). Additionally, concerning Twitter, Devkota et al. (2019) pointed out that all activities are not necessarily recorded. This is even more true for Flickr, which is less popular than Twitter.

Overall, the results suggest that all sampled extreme events affected human mobility patterns. These patterns, during different disaster phases, most often followed a TPL distribution, though this finding was not always statistically significant.

5.3 Spatial Extent of Human Activity

For all disasters, the spatial extents of Flickr users’ displacements in the study area are presented through the sum of MSD values in 80–140 min periods (Fig. 9.6). The spatial extent of human activity during the analyzed events is limited, as MSD values remained small.

Fig. 9.6

The MSD values computed for the three phases of each analyzed natural disaster: a rainstorm, b the Tohoku Earthquake, c Typhoon Wipha, d snow/rainstorm, e heavy rain, and f snowstorm

Apart from the 2011 Tohoku Earthquake, the MSD values before, during, and after each disaster followed similar patterns. Before the occurrence, people tended to move across the study area freely; however, once a disaster occurred, their spatial activities tended to diminish. The MSD values generally increased slowly over time—a pattern previously reported in several studies (Ahmouda et al. 2019; Barbosa et al. 2018). For the analyzed events, the pace of returning to normal conditions, and the time required to do so, most likely depended on the severity of each disaster. For instance, the MSD values recorded after the 2010 rainstorm showed a quick return to approximately normal values. Conversely, for the 2014 snow/rainstorm and the 2015 heavy rain, the MSD values showed a slight growth after the two events but did not change thereafter. Concerning Typhoon Wipha and the 2018 snowstorm, the computed MSD values remained nearly zero throughout the covered 120-minute period. Other studies (e.g., Ahmouda et al. 2019) reported a period of 240 min before human activities resumed in Houston, Texas, and in North and South Carolina after Hurricanes Harvey and Matthew, respectively.

However, the observed MSD values before, during, and after the Tohoku Earthquake showed different patterns. Only for this event was the spatial extent recorded after the disaster (40 km) wider than that recorded before the disaster (20 km). This might be attributed to the fact that people tend to flee to safer outdoor spaces after earthquakes—a postulation confirmed by analyzing the trips between indoor and outdoor environments, as shown in Sect. 9.5.4.

5.4 Trip Characteristics Between Indoor and Outdoor Environments

Figure 9.7 illustrates the number of trips between indoor and outdoor environments recorded during the different phases of each sampled event. We adopted the following notation (X, Y, T), where X and Y referred to the nature of the environment (I = indoor or O = outdoor) of the origin and the destination, respectively, and T referred to the category of time spent on each trip (S = Short [≤10 min]; M = Medium [>10 min and <2 h]; L = Long [≥ 2 h]).

Fig. 9.7

Trip characteristics illustrating the number of trips, based on the nature of the origins and destinations, for the three phases of each analyzed natural disaster: a rainstorm, b the Tohoku Earthquake, c Typhoon Wipha, d snow/rainstorm, e heavy rain, and f snowstorm. The following notation was adopted to represent the origin, destination, and length of the trips: (X, Y, T), where X and Y referred to the nature of the environment (I = indoor or O = outdoor) of the origin and the destination, respectively, whereas T referred to the category of the time length spent for each trip (S = Short [≤10 min]; M = Medium [>10 min and <2 h]; L = Long [≥2 h])

Overall, the number of trips from/to different environments tended to be lower during all events, compared to the corresponding numbers before and after the events, except for the 2018 snowstorm (Fig. 9.7f). This was most likely because people were excited about experiencing the snow as it is a rare occurrence in Tokyo. The number of trips generally began to increase once the natural event finished, and other patterns began to appear, especially longer trips from indoor to outdoor areas—e.g., (I, O, M), (I, O, L), (O, O, M), and (O, O, L).

Regarding the “before” phase, during all periods of wet-weather-related events, particularly before Typhoon Wipha (Fig. 9.7c) and the 2015 heavy rain (Fig. 9.7e), and with the exception of the 2010 rainstorm, people made shorter trips within or to indoor environments. Few trips to the outdoors were observed. While shorter trips indoors can be reasonably justified by the fact that people are taking shelter to protect themselves from bad weather, trips to the indoors from outdoors can be explained as shopping or commuting trips, in accordance with Ahmouda et al. (2019). Heavy rain and typhoons had the most considerable impact on people’s trips, even before the events, which is why more indoor and fewer outdoor trips were observed. This indicates that people were generally well-informed and prepared for the natural disasters. Once the perturbed states ended, people gradually began to return to their normal lives, as indicated by the increasing number of trips to the outdoors and the decreasing number of trips indoors.

For the 2011 Tohoku Earthquake (Fig. 9.7b), the “before” period showed roughly similar characteristics between the different trip classes, with more people tending to go outside for short and long durations. After the earthquake, the number of short and long trips to the outside increased, while the number of trips indoors decreased. The earthquake’s abruptness and magnitude caused people to leave covered places, which supports the previous results related to the wide spatial extent of human activities observed via the computed MSD values.

6 Discussion

The results of this study demonstrate that natural disasters disturb human mobility patterns in Tokyo, and the degree of this impact depends on the nature, length, and severity of the events. This section discusses how this analysis may inform our understanding of human mobility during disasters and sheds light on some prospective applications, which may help improve disaster preparedness and management. It also clarifies the study’s limitations and gives directions for future extensions of the research.

To the extent of our knowledge, this study is the first to analyze human mobility during perturbed states based on geotagged photos. The findings complement those reported by most studies that have used text-based social records, generally extracted from Twitter (Ahmouda et al. 2019; Wang and Taylor 2016). The urban human movement is, in most cases, governed by the TPL distribution. Contrary to other studies, this research compared the impacts of different natural disasters on human movements in the same area, instead of across several distanced study regions. Taking advantage of the rich, visual content from the sampled photos, we examined the nature of the environments in which they were taken (i.e., indoors or outdoors) using deep-learning techniques. Subsequently, we extracted the human trip characteristics across these environments, including duration of stay.

These insights are helpful for planners in two ways. First, they will help planners understand people’s responses to different extreme events by analyzing the length of their displacements, the spatial extent of their activities, and how long they stay indoors and outdoors. Second, they will help planners spatially locate popular areas where people tend to go before/during/after a natural disaster. Thus, planners, in cooperation with authorities, may prepare event-dependent guidelines designed to help citizens know how to behave properly during and after such events. Risky areas, such as the vicinity of rivers or building sites in case of typhoons or heavy rains, may be marked as a warning to citizens. Suitable potential evacuation shelters may also be pinpointed near locations of dense social activity. The media can be used to advise residents about these measurements and teach them how to safely prepare themselves, depending on the nature of the projected event. For example, staying indoors would be preferable during wet weather, as would heading to higher ground during earthquakes. Preparing water, nonperishable food supplies, and evacuation routes are crucial to any event.

The methods and results presented in this study could also be applied to analyze tourists’ mobility behaviors, focusing on several key aspects: movement trajectories, periods of stay indoors and outdoors, trip characteristics across these environments, and most frequented attractions during each phase of an extreme event. Tourists are usually unfamiliar with the places they visit, and planners and researchers often overlook them when studying disaster management. Previous research has demonstrated that Flickr records reflect the number of international arrivals in many countries. Barchiesi et al. (2015), for instance, concluded that the number of visitors to the UK, estimated based on Flickr data, correlates with official estimates. In another study, Preis et al. (2020) covered Group of Seven (G7) countries, including Japan, and reached the same conclusion. However, generally, few studies have analyzed the mobility behavior of tourists during natural disasters. These insights are of vital importance for Tokyo city planners, given that Tokyo is the most popular destination for international tourists visiting Japan. Tokyo is also prone to recurring natural disasters, which might create overwhelming experiences for unlucky, newly arrived visitors. Based on the results that could be produced from research concerning tourists, planners could make advance preparations or update measurements specifically designed to help tourists know how to prepare for or react during a disaster, as well as what locations to avoid.

Although the results obtained in this research provide insights about human mobility patterns during different extreme events, the analysis suffers from several limitations. The first relates to the limited number of geotagged records. Like any other social data shared online, people care about their privacy, so they tend to hide their geolocations and other sensitive information. This leads to fewer geotagged records. One way to remedy this issue is to obtain geotagged photos from other platforms, such as Instagram and Snapchat, or tweets with images. Another possible solution is to predict an approximate location based on visual clues or textual tags. A second limitation of this analysis is that Flickr does not cover the activities of the general population because not all internet users take photos, let alone share them on social networks. Therefore, combining data from different sources may be useful to cover as much of the population as possible.

This study could be extended in numerous ways. First, in addition to Flickr, geotagged records from other social platforms (e.g., Instagram) might be considered, either to complement the Flickr data or to compare the analysis outcomes based on different data sources and examine biases. Another promising avenue is to investigate the environments in which photos were taken by exploring the textual and visual tags extracted from the photos’ content. While the nature of the environment (indoor/outdoor) extracted in this research gives an idea of the most frequented place types during different phases of the sampled events, it is vitally important to examine these frequented locations in more detail. Based on the photos collected for this study, we propose the following general categories: Nature, Culture, Entertainment (festivals, crowded areas, etc.), Shopping, Infrastructure (transportation), and Residence (hotels, homes, etc.). A third extension path for this analysis would be to examine the links between disturbed human mobility and disaster severity using parameters other than weather. This would require analyzing extreme events of the same type and collecting additional spatial datasets regarding, for instance, economic loss indicators and remote sensing data to assess each disaster’s severity.

7 Conclusions

We have explored and compared human movement patterns before, during, and after different extreme natural events based on geotagged, timestamped photos shared online. Given Tokyo’s importance as one of the largest and most densely populated cities in the world, and considering the frequent dangers it faces from recurrent natural disasters, we selected its 23 special wards as the target area. Among 22 extreme natural events that impacted Tokyo from 2008–2019, six disasters of different types were selected: earthquake, typhoon, snowstorm, rainstorm, heavy rain, and snow/rainstorm.

Regarding the research questions, the results show that (1) human mobility can be described, in most cases, via TPL models and (2) that natural disasters perturb human displacements to different extents depending on the disaster’s severity and type. Wet-weather-related disasters, particularly Typhoon Wipha and the 2015 heavy rain, disturbed human movements the most in terms of displacements, traveled distances, and trip characteristics. However, the Tohoku Earthquake, given its abruptness, showed different human mobility patterns; people tended to flee to safer outdoor places, as is evidenced by the long distances traveled after the earthquake and the high number of trips outdoors. For the third research question, compared to text-based, social geodata, geotagged photos provide the possibility of extracting the nature of the locations (indoors/outdoors) at which photos were taken using deep-learning techniques; however, one major disadvantage of this data is the lack of sufficient samples to conduct the required analyses for some events.

In summary, using geotagged photos for disaster management is promising and might complement the existing knowledge acquired using other, traditional data. However, many challenges related to the sample size and heterogeneity of geotagged photos remain and must be addressed in future studies