A Study of Intercity Transportation Demand in Fukui Prefecture Under 2018 Heavy Snowfall: Based on Mobile GPS Data
2024 Volume 5 Issue 2 Pages 57-65
Details
2024 Volume 5 Issue 2 Pages 57-65
Understanding intercity transportation demand and trends during snowstorms is crucial for mitigating traffic accidents and congestion. This study analyzed the characteristics of intercity transportation demand using Agoop mobile GPS data, focusing on the 2018 Fukui Heavy Snow Event. The K-dimensional Tree (KDTree) algorithm for neighbor matching was employed to examine the spatiotemporal characteristics of intercity transportation demand, providing insights into the overall trends under the influence of the snow disaster. Additionally, Singular Value Decomposition (SVD) was utilized to decompose and reduce the dimensions of the spatiotemporal OD matrix, facilitating the identification of the composition of intercity transportation flow. The study identified three phases of the event: the stable phase (January 27 to February 2), the snow disaster phase (February 3 to February 11), and the recovery phase (February 12 to February 16). During the snow disaster, intercity transportation demand dropped by 67.86% compared to the stable phase. Intercity transportation demand during the snow disaster included daily demand (M1) and special demand (M2). M1 traffic primarily originated from Fukui City, Sabae City, Awara City, and Sakai City, with Fukui City and Sakai City being key points of departure and arrival. In the M2 model, traffic from southern to northern Fukui Prefecture nearly ceased, shifting from a longitudinal pattern along National Route 8 and the Hokuriku Shinkansen to a horizontal distribution towards Ono and Katsuyama. Departure flows from Sakai City decreased significantly, while Tsuruga City saw a significant increase in traffic. This indicates that the snow disaster mainly affected northern Fukui Prefecture, especially Fukui City and Sakai City.
Heavy snowfall affects 60% of the land and 20% of the population annually in Japan1). Snowstorms exacerbate traffic accidents and congestion, causing casualties and economic losses. During the 2018 Fukui snowstorm, 3 people died, 10 were severely injured, 18 sustained minor injuries, and about 1,500 vehicles were stranded on National Route 813).
Extreme weather induces changes in the spatial and temporal characteristics and micromobility of intercity transportation2), leading to heterogeneity in traffic behavior3). To address the transportation pressures caused by these spatiotemporal variations, government measures often include road closures, snow clearance, or issuing evacuation warnings. However, inappropriate evacuation policies can exacerbate traffic incidents and congestion. For example, in January 2014, the simultaneous rush of many residents to return home following a snowstorm warning in Atlanta, Georgia, in the United States, resulted in massive vehicular gridlock4). Thus, understanding the demands and trends of intercity traffic during snowstorms and assisting governments in issuing timely and effective evacuation alerts can effectively alleviate such issues.
Furthermore, current research on snowstorms primarily relies on ETC (Electronic Toll Collection) data or satellite imagery5-6). However, ETC data is limited to vehicles equipped with ETC systems, while satellite imagery requires ultra-high-resolution images and sophisticated image processing techniques. Consequently, integrating mobile GPS data, which offers wider coverage, real-time accuracy, and greater flexibility, has become a significant approach for studying intercity traffic demands and trends in recent years7-9). Yet, research based on GPS data has primarily focused on disasters such as hurricanes, tsunamis, and COVID-1910-12), with relatively limited studies on intercity traffic demands and trends under snowstorm conditions.
Considering the outlined factors, the utilization of mobile GPS data to understand the overall trends and structural changes in intercity transportation demands during snowstorms presents a promising research opportunity. The study aims to identify the spatiotemporal characteristics of traffic during snowstorms, analyze the overall trends and composition of intercity traffic flows, and uncover the regions and road segments most severely impacted by traffic disruptions. This approach could provide crucial insights into traffic flow patterns and trends during snowstorms, offering significant reference for responding to similar future events.
(1) Research area
Fukui Prefecture is located in central Japan on Honshu Island, near the Sea of Japan, as shown in Fig.1. Fukui Prefecture, with a population of 762,679, accounts for approximately 6/1000 of Japan’s total population and covers about 1/100 of Japan’s land area. It includes both urbanized and rural areas. Its traffic demand and flow are shaped by a warm, humid climate and heavy snowfall in winter due to the Sea of Japan cold air mass, which significantly impacts traffic during snowstorms. The 2018 Fukui Heavy Snow Event brought severe snowfall to Fukui Prefecture and surrounding areas. This event caused traffic disruptions, school closures, and affected residents. As a result of the snow disaster, around 1,500 vehicles became temporarily stranded on National Route 813).
(2) Research data
This study primarily utilizes mobile GPS data collected from Agoop during the snow disaster in Fukui Prefecture in 2018, along with road data and administrative district data. Agoop is a Japanese information service business company. The road data were downloaded from OpenStreetMap (OSM), and the administrative district data were obtained from the Geospatial Information Authority of Japan. The Agoop mobile GPS data provide valuable insights into population mobility, covering a period of 21 days from January 27 to February 16, 2018. The data structure is presented in Table I.
(3) Data processing
Mobile data contains a significant amount of noisy data. In addition to standard data cleaning methods, we focused on cleaning redundant data, ping-pong data, and drift data.
Redundant Data: This occurs when the same user has multiple records at the same timestamp in the mobile data. This may be caused by a short sampling interval or insufficient precision in the timestamp field. For consecutive records (n ≥ 3) with the same location, only the first and the nth records are retained. The intermediate records are considered redundant and can be discarded.
Ping-Pong Data: These are location data where the mobile device frequently switches between nearby base stations within a short period of time. This often occurs due to unstable signals, dense base stations, or the device being in a specific environment (e.g., high-speed travel or underground metro). This data is cleaned using the same method as drift data.
Drift Data: Due to environmental interference, insufficient satellite coverage, or signal obstructions, there may be discrepancies between the collected data and the actual situation, leading to large deviations between the recorded and actual locations, manifesting as sudden jumps in the data. We first determined the geographic boundaries of the study area (135.449339, 35.343678, 136.832362, 36.451457) and removed data outside this region. Then, within the study area, we applied the following cleaning methods:
Speed Threshold: If the speed between the current trajectory point and the previous/next points is excessively high, the data is considered drift.
Distance Threshold: If the distance between the current trajectory point and the previous/next points is too large, the data is considered drift.
Angle Threshold: If the angle formed by the previous, current, and next trajectory points is too small, the data is considered drift.
The main research methodology used in this study includes includes K-dimensional Tree (KDTree) to identify road segments severely affected by the snow disaster and Singular Value Decomposition (SVD) to analyze the spatiotemporal characteristics of the OD matrix. The research workflow is illustrated in Fig.2.
(1) K-dimensional Tree (KDTree)
KDTree (K-Dimensional Tree) is a data structure that efficiently finds the nearest neighbor in high- dimensional spaces. It divides the space into multiple regions by recursively constructing a tree to organize the data.
Assume there are N data points, with dimensions D = 0,1,2...,k. At each node, the tree selects a dimension D to split the data. In the first layer (root node), the data is sorted by the first dimension, and the median point becomes the root. Points smaller than the median go to the left subtree, and larger points go to the right. In the next layer, the second dimension is used for splitting, and the process repeats. This recursive construction continues until a leaf node is reached, or a stopping condition is met. At the i-th layer, if we are at the d-th dimension (where d=i%k), the splitting boundary is the median value md along this dimension. The formula is as follows14):
In practice, we convert the Agoop data into a GeoDataFrame of point data and the OSM road network data into a GeoDataFrame of line data. By utilizing KDTree, we build a spatial index based on OSM road nodes to match the point dataset to the nearest road segments in the line dataset. This method returns a GeoDataFrame containing all columns from the point dataset as well as all columns from the line dataset. Additionally, a new column is added to the point dataset, which records the distance from each point to the nearest line in the line dataset. Each time an Agoop GPS point is encountered, KDTree quickly finds the nearest OSM road node, and the GPS point is “matched” to the closest road.
(2) Singular Value Decomposition (SVD)
The SVD algorithm is primarily used for dimensionality reduction in feature extraction, as well as in recommendation systems, natural language processing, and computer vision. Suppose M is an mx n matrix. The resulting matrix U is an mxm orthogonal matrix, where the vectors in U are known as the left singular vectors. S is an mxn diagonal matrix with its diagonal elements representing the singular values and all off-diagonal elements being zero. The decomposition process is shown in Fig.3. The right singular matrix VT is an nxn orthogonal matrix, where the vectors within V are the right singular vectors. The δh is the h -th diagonal element of S, uh is the h-th column of U, and vh is the h -th column of VT , which is the transpose of V.
The formula for SVD can be expressed as follows15):
(3) Application of SVD in spatio-temporal OD matrix for intercity transportation
SVD can be applied to decompose spatiotemporal matrices, breaking them down into simpler, additive spatiotemporal variation patterns. The practical significance of performing SVD on the OD matrix of intercity transportation is that M represents a collection of r types of intercity transportation flow patterns. uh represents the temporal distribution of the h-th type of intercity transportation flow, ;%0A%09%09%09newWindow.document.open();%0A%09%09%09newWindow.document.write('<img src=%22./Graphics/5.2_57_M001.jpg%22>');%0A%09%09%09newWindow.document.close();%0A%09%09)
represents the spatial distribution of the h-th type of transportation flow patterns, and δh indicates the importance of the h-th type of intercity transportation flow patterns. When mapping a high-dimensional spatio-temporal OD matrix to a lower-dimensional space, assuming that the top k components with the largest singular values are retained (where 1 ≤ k ≤ r), and the components with smaller singular values are discarded, a lower-dimensional spatio-temporal OD matrix ;%0A%09%09%09newWindow.document.open();%0A%09%09%09newWindow.document.write('<img src=%22./Graphics/5.2_57_M002.jpg%22>');%0A%09%09%09newWindow.document.close();%0A%09%09)
is obtained.
(1) Road network matching
To observe the traffic flow trends during snowstorms, it is necessary to visualize the traffic distribution across various routes on the entire road network. This process aids in subsequent intercity travel statistics and streamlines the application of SVD. We downloaded the major and minor road data of Fukui Prefecture from OpenStreetMap (OSM) and combined it with the cleaned move point data from Agoop mobile data. Using the KDTree algorithm, we performed nearest neighbor matching. The matched results are shown in Fig.4.
The trajectory visualization analysis reveals that areas with high traffic volume are mainly concentrated in the northern region of Fukui Prefecture, radiating from Fukui City to surrounding cities such as Awara City, Sakai City, and Sabae City. Starting from February 3rd, the traffic pressure in Awara City and Sakai City significantly increased. From February 5th to February 7th, the traffic volume in Awara City, Sakai City, Fukui City, Katsuyama City, and Ono City noticeably decreased. By February 7th, the traffic pressure on certain sections of National Route 8 increased, which is consistent with network information13). Meanwhile, the traffic volume in Katsuyama City and Ono City almost disappeared. After February 13th, traffic patterns gradually began to return to their stable pre-snow disaster state.
(2) Spatio-temporal OD matrix decomposition
For the period from January 27 to February 16, 2018, the spatio-temporal origin-destination matrix for intercity transportation in Fukui Prefecture was analyzed and dimensionally reduced by employing SVD. Suppose the spatio-temporal OD matrix of intercity transportation on a specified date s is Ds, which contains n OD pairs. Transform Ds into a 1 × n row vector ds. The intercity transportation flow over m days can be represented as an m × n matrix M . Through singular value decomposition, the spatio-temporal OD matrix of transportation flow can be expressed as the product of three matrices. To construct the OD matrix, it is first necessary to identify the move and stay data from the Agoop data. Typically, this is done using a time threshold method: after dividing Fukui Prefecture into 500- meter grid cells, if a user stays in a particular grid for more than 30 minutes, it is considered a stay. The time period between two stays is identified as a trip. Next, we focus solely on the move data, using the grid at the start of the first trip as the origin and the grid at the end of the last trip as the destination. Finally, all trips are aggregated by intercity travel, with each unique OD pair forming the rows and columns, and the dates forming the columns, resulting in the construction of the OD matrix. This matrix consists of 215 rows and 21 columns.
Singular values are indicators of the importance of each demand pattern in the spatiotemporal OD matrix. A larger singular value indicates that more of the original information in the matrix is preserved. As shown in Fig.5, the singular value distribution graph demonstrates that the first two singular values are significantly higher than the others, with minimal changes observed after the third singular value. By selecting the first two singular values, we can represent 88% of the original information contained in the spatiotemporal OD matrix through the corresponding demand patterns. These two singular values can be associated with two typical demand patterns and passenger flow types, denoted as M1 and M2. Additionally, M1 and M2 can be further decomposed into u1, v1 and u2, v2, where ui represents temporal variations and vi represents spatial variations.
In Fig.6 and Fig.7, u1 and u2 represent the temporal unit vectors of transportation demand across different categories. The sign of the temporal unit vectors corresponds to the direction of fluctuations in intercity transportation demand over time, while their absolute values indicate the magnitude of these fluctuations. Similarly, v1 and v2 represent the spatial unit vectors of transportation demand for different categories, where positive values indicate an increase in traffic flow and negative values indicate a decrease.
In Fig.8, the departure and arrival volumes of intercity transportation demand for different cities are aggregated separately. The intensity of the colors reflects the magnitude of fluctuations in the corresponding directions.
By integrating the temporal and spatial unit vectors, it is possible to observe the fluctuations in various types of intercity transportation demand. If significant fluctuations are observed in the temporal unit vectors during a specific period, and corresponding significant fluctuations are noted in the departure and arrival flows of a particular city or between specific OD pairs, this indicates a substantial increase in that type of demand within the respective temporal and spatial scope. Conversely, if the temporal and spatial unit vectors exhibit opposite fluctuations, it suggests a significant decrease in that type of demand within the respective temporal and spatial range.
Transportation Flow Type I (M1) represents the daily intercity transportation demand within Fukui Prefecture and can be divided into three distinct stages: the stable stage from January 27 to February 2, the snowstorm stage from February 3 to February 11, and the recovery stage from February 12 to February 16. Starting on February 3, Transportation Flow Type I, represented by u1 , showed a marked decline, with its value dropping from 0.28 to a low of 0.09. This indicates that, due to the combined effects of the snowstorm and traffic control measures in some areas, the daily intercity transportation demand decreased by 67.86% within a short period. A brief rebound occurred from February 7 to February 9, coinciding with the implementation of traffic control and government policies, thereby reflecting the effectiveness of these measures. Spatially, Passenger Flow Type I primarily originates from four major urban areas: Fukui City, Sabae City, Awara City, and Sakai City.
Compared to u1, u2 exhibits greater volatility, suggesting it captures a different temporal variation pattern from the main trend, specifically representing the dynamic transportation demand during the snowstorm period, i.e., Transportation Flow Type II (M2). Beginning on February 1, the value of u2 initially dropped sharply from -0.10 to -0.39, before rapidly rising to a peak of 0.41. This indicates that the snowstorm caused significant dynamic changes, such as substantial disruptions in transportation during the initial stage followed by recovery or emergency transportation flow. Spatially, most OD pairs showed minimal impact (values close to zero), but several exhibited significant positive and negative fluctuations. These fluctuations may indicate relatively small but crucial spatial changes in the time pattern corresponding to u2. Positive peaks represent abnormal increases in transportation flow during specific periods, whereas negative peaks indicate decreases in transportation flow. Traffic volume from cities in the lower part of Fukui Prefecture to the upper part generally decreased, likely due to the concentration of the snowstorm in the upper part of Fukui. In contrast, in the upper part of Fukui, the distribution pattern shifted from a vertical distribution in the original M1 mode to a horizontal distribution, with increased traffic flow from Fukui and its surrounding cities towards Ono and Katsuyama.
In the M1 model, the origin departure flows and destination arrival flows are mainly concentrated in Fukui City and Sakai City. However, in the M2 model, the origin departure flows are primarily in Sabae City and Echizen City, with a significant decrease in departure flows from Sakai City. Additionally, the destination arrival flows in Fukui City have greatly decreased, while the traffic volume in Tsuruga City has significantly increased. This indicates that the snow disaster primarily affected the northern part of Fukui Prefecture, particularly Fukui City and Sakai City.
In addition, using Google Trends data to identify the cities with the highest search volumes for snowstorm-related keywords in Fukui Prefecture over the past 10 years, as shown in Fig.9, reveals that the cities with the highest search volumes largely align with those identified in this study as being affected by snowstorms. The scores range from 0 to 100: a score of 100 indicates the highest search volume, 50 represents half that, and 0 indicates insufficient data. This consistency in search patterns suggests that the spatiotemporal characteristics of snowstorms in Fukui Prefecture have remained similar over the past decade. Consequently, the traffic patterns observed during this particular snowstorm may reoccur in future events.
This study utilized the KDTree algorithm and Singular Value Decomposition (SVD) to decompose and analyze the traffic patterns and demand in Fukui Prefecture during the 2018 snow disaster. The findings are as follows:
(1) Identification of Event Phases: The study identified three phases of the event: the stable phase (January 27 to February 2), the snow disaster phase (February 3 to February 11), and the recovery phase (February 12 to February 16). During the snow disaster, intercity transportation demand dropped by 67.86% compared to the stable phase.
(2) Spatial and temporal characteristics of transportation under different modes: Intercity transportation demand during the snow disaster included daily demand (M1) and special demand (M2). In the M1 model, traffic primarily originated from Fukui City, Sabae City, Awara City, and Sakai City, with Fukui City and Sakai City being key points of departure and arrival. In the M2 model, traffic from southern to northern Fukui Prefecture nearly ceased, shifting from a longitudinal pattern along National Route 8 and the Hokuriku Shinkansen to a horizontal distribution towards Ono and Katsuyama. Departure flows from Sakai City decreased significantly, while Tsuruga City saw a significant increase in traffic.
(3) Impact Scope: Snow disaster mainly affected northern Fukui Prefecture, especially Fukui City and Sakai City. The traffic pressure is mainly concentrated in Fukui City, Sabae City, Awara City, Echizen City, Sakai City, Katsuyama City, and Tsuruga City. Once a snow disaster occurs, these areas are more prone to traffic congestion or accidents. The government should strengthen the allocation of evacuation resources in these cities.
(4) An analysis using Google Trends data for snowstorm-related keyword searches over the past 10 years in Fukui Prefecture revealed that the highest-ranked cities largely correspond to those identified in this study as being affected by snowstorms. This indicates that the spatiotemporal characteristics of snowstorms in Fukui Prefecture have shown similar patterns over the past decade. Consequently, the traffic patterns observed during this particular snowstorm are likely to reoccur in future snowstorms. However, due to the lack of precise snowfall data for Fukui Prefecture over the past decade, and given that each snowstorm exhibits distinct spatiotemporal characteristics, the impact on the road network and traffic flow may vary. Therefore, the present study can only provide a general identification of areas and roads significantly affected by snowstorms. Accurately predicting the road segments most impacted in future snowstorms remains a key direction for further research.
This study employed a Python library named ‘transbigdata’. The authors gratefully acknowledge the developers for granting permission to use this library.
