key: cord-0830959-4bnn8zu1
authors: Xin, Rui; Ai, Tinghua; Ding, Linfang; Zhu, Ruoxin; Meng, Liqiu
title: Impact of the COVID-19 pandemic on urban human mobility - A multiscale geospatial network analysis using New York bike-sharing data
date: 2022-03-24
journal: Cities
DOI: 10.1016/j.cities.2022.103677
sha: fe9a6cd7038cc23ac25aa50b2d10378b04c40a38
doc_id: 830959
cord_uid: 4bnn8zu1

The COVID-19 pandemic breaking out at the end of 2019 has seriously impacted urban human mobility and poses great challenges for traffic management and urban planning. An understanding of this influence from multiple perspectives is urgently needed. In this study, we propose a multiscale geospatial network framework for the analysis of bike-sharing data, aiming to provide a new perspective for the exploration of the pandemic impact on urban human mobility. More specifically, we organize the bike-sharing data into a network representation, and divide the network into a three-scale structure, ranging from the whole bike system at the macroscale, to the network community at the mesoscale and then to the bicycle station at the microscale. The spatiotemporal analysis of bike-sharing data at each scale is combined with visualization methods for an intuitive understanding of the patterns. We select New York City, one of the most seriously influenced city by the pandemic, as the study area, and used Citi Bike bike-sharing data from January to April in 2019 and 2020 in this area for the investigation. The analysis results show that with the development of the pandemic, the riding flow and its spatiotemporal distribution pattern changed significantly, which had a series of effects on the use and management of bikes in the city. These findings may provide useful references during the pandemic for various stakeholders, e.g., citizens for their travel planning, bike-sharing companies for bicycle dispatching and bicycle disinfection management, and governments for traffic management.

Cities are complex systems where human move around, interact with urban facilities, and produce a variety of flows that reflect their mobility traces. The study of human mobility can help capture the spatiotemporal movement patterns in cities and is crucial for applications such as migratory flow estimation, traffic forecasting, urban planning, and pandemic modeling (Barbosa et al., 2018) . Traditional methods for the detection and analysis of human flow mainly use questionnaire methods. In recent years, the advancement of the positioning and information and communication technologies (ICTs) provides new opportunities with timely collected large-scale movement data for analyzing human mobility (Ding et al., 2015; Yuan et al., 2012) . For instance, in the context of the boom of the sharing economy, biking-sharing has become prevalent over the last years (Si et al., 2019) .

Extensive studies have applied complex network methods to construct geospatial networks from traffic data and explore network structures and characteristics in various transportation domains, e.g., railway transportation (Cats, 2017) , bus transportation , aviation (Wang et al., 2011) , and maritime transport (Ducruet, 2017) .

In this study, we propose a framework to combine geospatial complex network and multi-scale geospatial analysis for the investigation of urban human mobility using bike-sharing data during the pandemic. To the best of our knowledge, there has been no prior study to analyze the impact of the COVID-19 on the bike-sharing system from the perspective of urban human mobility. Using this framework, we aim to: (1) Explore the feasibility of using the bike-sharing data to investigate urban human mobility during the pandemic; (2) Combine the multiscale geospatial analysis with the complex network methods to explore the riding behaviors based on the bike-sharing data at different scales;

and (3) Visually analyze the spatiotemporal factors and the abstract network indicators for the bike-sharing system.

More specifically, in this framework, we organize the bike-sharing data by a network representation. We further divide the network into a three-scale structure, ranging from the whole bike system at the macroscale, to local network communities at the mesoscale and then to the individual bicycle stations at the microscale. At the macroscale, we focus on the analysis of the overall characteristics and distribution trends from the trip data using network indicator statistics and kernel density estimation visualization. At the mesoscale, we first detect network communities and construct the subnetworks based on the detected communities and then visually analyze these subnetwork indicators. At the J o u r n a l P r e -p r o o f microscale, we take the bicycle station which is the basic element of the network as the research object to study the differences of relevant network indicators in different periods.

Using the above methods, we try to answer the following research questions: Overall, does the pandemic have an impact on urban mobility represented by shared bicycle riding?

Specifically, what changes have taken place in the spatiotemporal patterns of riding behavior during the pandemic? Furthermore, how has the characteristics of the bike-sharing network changed in the pandemic, such as community structure, network topology and network flow?

We apply our framework to the origin-destination (OD) bike-sharing data collected from the study area of New York City, USA from January to April in 2019 and 2020. The multi-scale visualization and analysis results reveal a series of spatiotemporal changes of riding behaviors during the pandemic. The network structure also changes significantly.

Our framework can be adopted or extended to similar topics in other cities in the world to help people understand the impact of the pandemic on urban life, and support relevant companies and governments for their decision-making. There are differences in pandemic severity and pandemic prevention policies in different regions. The comparison of different results around the world is also helpful for a comprehensive understanding of the spatiotemporal similarities and differences of the impact of the pandemic. The remainder of this paper is organized as follows. Section 2 describes the relevant research background. The study area and data are shown in Section 3. Section 4 introduces the research framework and relevant method in detail. Section 5 presents the experiment and analysis. Section 6 introduces the implications of the case study results, and discusses J o u r n a l P r e -p r o o f Journal Pre-proof the limitations of the study and the future work. Finally, Section 7 concludes the paper.

The emergence of multiple new types of data sources in recent years provides a broader perspective for the study of urban mobility. For instance, check-in data , GPS data (Molloy et al., 2021) , sports and health data (Braun & Malizia, 2015) are used to assist the analysis of dynamic human flows. With the popularity of the sharing economy in recent years, bike-sharing data has been increasingly examined (Si et al., 2019; Fishman et al., 2013) . As a type of spatiotemporal data, bike-sharing data contain the bike use information of a large number of users that reflects the riding activities in the city. This rich information helps understand the urban flow patterns from the perspective of social sensing (Liu et al., 2015) . A number of research has employed riding data to study the usage characteristics of shared bicycles and discover different travel patterns (Du and Cheng, 2018; Zhang et al., 2018) . Bike-sharing data have been also widely used for estimation purposes, e.g., travel destination choice analysis (Faghih-Imani and Eluru, 2015) and bike-sharing demand estimation (Faghih-Imani and Eluru, 2016) , which are helpful in improving bicycle scheduling scheme by bicycle companies and enhancing user experiences. In addition, there have been comparative studies conducted to analyze the respective advantages of shared bicycles and other transport modes (e.g., ride-hailing services and taxis) and their interactions to help a deep understanding of the transport connections in the city (Faghih-Imani et al., 2018; McKenzie, 2020) . Moreover, many research work have conducted spatiotemporal analysis by combining bike-sharing data J o u r n a l P r e -p r o o f with other sources of data. For instance, Li et al. (2020) investigated shared bicycle trajectories together with transportation network data to explore the human mobility in the city.

The study of urban human mobility patterns requires dedicated methods from different disciplines. Many research works have applied complex network theory (Newman, 2018) in various spatial fields and constructed geospatial networks by combining complex networks and spatial locations. Barthélemy (2011) conducted a comprehensive survey of geospatial networks and reviewed important spatial network models. The survey also explained how the spatial constraints affect the network structure and properties and discussed various processes taking place on these spatial networks.

Lin and Ban (2013) reviewed research works on transport networks from a complex network perspective and summarized network expression and construction methods of various transportation systems. In terms of specific transportation networks research, Wang et al. (2020) built networks for urban bus data at different scales to analyze the spatial configuration of urban bus networks based on geospatial network analysis methods. In terms of railway network, Cats (2017) made a longitudinal analysis of the topological evolution of multimodal railway network by investigating its topology dynamics using data collected from Stockholm. In the field of aviation, Wang et al. (2011) used geospatial complex network to explore the network structure and node centrality of various cities in China's air transport network, and compared the characteristics of air transport network in China with those in other countries. Dai et al. (2018) investigated the evolving structure of the Southeast Asian air transport network over the period J o u r n a l P r e -p r o o f to captures the main topological and spatial changes. Geospatial complex networks were also applied to maritime transport to allow a new understanding of the factors affecting the development of ports and shipping (Ducruet, 2017) .

For the studies on bike-sharing data, some research models and statistically analyzes the influence of various spatiotemporal variables and factors (e.g., weather) on daily bike-riding and bike demand (Kutela and Teng, 2019; Faghih-Imani and Eluru, 2016) .

Other research focuses on the spatial mining of bike-sharing data, such as using spatial clustering methods to explore the riding purpose (Wang and Lindsey, 2019) or analyze the mobility patterns in urban environments (Keler et al., 2019) . Geospatial complex networks can combine statistical and spatial analysis for network indicators based statistical analysis (Saberi et al., 2017) and spatial visualization of network features (Zhong et al., 2014) . So it is suitable to use the complex networks to study the bike-sharing data.

Although there are many transportation network researches in the field of traditional transportation, there is still insufficient research on bike-sharing data analysis using complex network at present. Austwick et al. (2013) were among the few who applied complex networks to bike-sharing data, and pointed out that spatial analysis of complex networks has not been fully explored. The main reason is that traditional complex network analysis methods, such as indicator calculation, need to be extended for bike-sharing data with inherent spatial characteristics. Some researchers have used complex networks to analyze the influence of interference factors on the bike-sharing system. For instance, Saberi et al. (2018) studied the impact of the London Metro strike on the bike-sharing system based on complex network indicators. Yang et al. (2019) studied the changes in the bike-sharing system caused by the operation of the new metro line. Bike-sharing data can strongly support the study of urban human mobility by combining with methods from complex network, spatial analysis and other domains. However, many of these studies mainly focus on statistical analysis at an aggregated level and their results are mostly statistical values without sufficient spatial-related visual representations. Comprehensive analysis at multi-scales and intuitive visualizations are still needed.

The outbreak and spread of the COVID-19 pandemic reflect the vulnerability of the urban system and pose challenges for urban disaster resistance (Shamsuddin, 2020) .

However, in urban research field, there is still little research on the impact of COVID-19 on human mobility (Liu, 2020; Ghosh et al., 2020) , especially based on bike-sharing data.

More works with this regard are needed to help people understand the impact of the pandemic on the urban system and assist in the prevention and management of public health events in future cities.

The analysis in this research utilizes bike-sharing datasets from New York City, which is the most populous city in the United States, and has been seriously affected by the COVID-19 pandemic. There are five administrative regions in New York, including Manhattan, Brooklyn, Queens, Bronx, and Staten Island. Since Bronx and State Island has not bicycle stations, we selected the three administrative regions of Manhattan, Brooklyn and Queens with bicycle stations as the study area. Figure 1 shows the study area and the distribution of the bicycle stations. The data is collected from the official website of Citi Bike 1 , the largest public bike system in the United States (Tedeschi, 2016) . Citi Bike was launched in May 2013 and is a kind of dock-based bike-sharing system. On its website, Citi Bike provides free available trip data from July 2013 in the format of CSV files with each file containing the amount of data items for one month. To provide the context knowledge, we visualized the timeline of the pandemic in New York City in the first half of 2020 in Figure 2 , it shows that the impact of the pandemic at the initial stage is very serious with the number of infections reaching the peak in March and April. Some important policies related to pandemic protection were announced during this period. To make a comparative analysis of the pandemic influence, we select the bike-sharing data of New York City from January to April in 2020, which conclude the initial stage of the pandemic, and the same months in 2019. There are in total of 9,143,730 records in this dataset. The original data contains data of Jersey City which does not belong to the study area of this research. Thus, we removed this part of the data.

The data are typical OD data that records the origin and the destination of every single riding. Table 1 shows the attributes and example values of the data. analysis and the comparison of differences between the pandemic and the normal periods allow us to investigate the impact of the pandemic on the bike-sharing system and on urban human mobility. For the network construction and analysis in this study, we mainly use two libraries: Python-Igraph package (Csardi and Nepusz, 2006) and

Python-NetworkX package (Hagberg et al., 2008) . To construct the network, we first extract the bicycle stations from the raw data as the nodes of the network. Then we extract the trip connection relationships between each pair of origin and destination stations. We further aggregate them based on the pair of origins and the destinations to obtain the edges of the network. The direction of an edge is the riding direction between stations. The weight of an edge is reflected by the bike flow between the OD pair. Finally, we obtain a weighted directed network constructed from riding data. An example of the constructed bike-sharing network is shown in Figure 4 where red dots represent bicycle stations, gray lines represent directed edges between J o u r n a l P r e -p r o o f the stations, and the widths of these edges represent bike flows. In this network, there are more bike flows from Station A to Station B than vice versa. A large number of flows is from Station C to Station B, while only a few from Station A to Station C.

At the macro scale, we first analyze the statistical indicators of the riding data. Then we analyze the network statistical indicators related to the network topology and network flow. Finally, we apply line charts, bar charts and kernel density estimation for the visualization and analysis of the statistics and the spatial distributions.

Trip volume derived from riding data is an intuitive indicator of urban human mobility.

A large trip volume reflects high urban human mobility of to a certain extent. The trip volume usually has time-varying characteristics. For instance, riding flows are significantly affected by varying temperatures, and by human activity patterns such as daily commuting patterns.

Riding duration is another indicator that can be used to reflect different riding purposes. For instance, short-duration riding is often used to connect the last/first mile with subway stations and other transport hubs, while long-duration riding is often used for leisure purposes. Due to different purposes of riding, the riding behaviors will follow distinct spatial and temporal distribution patterns (Xu et al., 2019) . For instance, bike ridings for commuting purpose often connect accommodation or organizations with J o u r n a l P r e -p r o o f transportation hubs in the morning and evening rush hours, while leisure ridings are mostly distributed in scenic spots. In addition, different riding durations lead to different losses to bikes, which will also affect the maintenance cost and riding pricing of the bike-sharing company. In terms of bike usage, indicators such as average use times and average use duration can not only be used to analyze users' riding behavior, but also provide assistance for bike companies' daily operations such as renewal and maintenance strategies.

The network statistical indicators in this study mainly cover two aspects: network topology and network flow. The former contributes to the analysis of the structure and the exploration of the connectivity and connection characteristics between bicycle stations.

The latter focuses on the understanding of the characteristics and distribution of the flows on the network. Besides, we use some computational indicators to explore the distribution of the above indicators. Below we introduce each indicator in detail.

Network topology indicators We use three indictors, i.e., degree, connectivity, and aggregation coefficient, to reflect the topology characteristics of the whole network. In a network, the degree K of a node is the number of edges directly connected to the node.

For the bike-sharing network, it is the number of ride connections at a station. In a directed network, depending on the direction of the edge, it can be divided into out-degree and in-degree. A high degree of a node means a bicycle station has extensive riding connections with other stations.

The connectivity of the whole network δ is quantitatively calculated by Equation 1, in which N is the number of nodes and L is the number of edges. A large δ indicates that the riding connections between stations in the bike-sharing network is relatively denser and thus the overall connectivity of the network is better. When calculating the node average flow <F>, d is the number of days in the current month.

When calculating the node average degree <K>, d is set to 1. For the calculation of edge average flow <W>, n represents the number of edges and d is the number of days in the current month.

To further investigate the discrete characteristics of network indicator distribution, we

is the standard deviation and <x> is the average value. In particular, the coefficient of variation is not affected by measurement scale and dimension.

Bicycle stations can be regarded as discrete points in space. Density visualization can present the spatial distribution of these points in an intuitive way and help an immediate perception and understanding of the macro characteristics. Compared to the J o u r n a l P r e -p r o o f commonly used point density calculation methods, e.g., quadrat method and Voronoi diagram method, that ignores the heterogeneity of spatial distribution or the continuity of spatial phenomena, kernel density estimation method (Parzen, 1962) has the attenuation effect that takes both spatial heterogeneity and continuity into account. As a field representation of spatial phenomena, kernel density estimation visualization based on spatial smoothing and spatial interpolation technologies can be applied in our study to explore the bike flow intensity and its temporal change from a macro spatial perspective.

The kernel density estimation f(x) is calculated using Equation 7, where h is the bandwidth, n is the number of discrete points in the bandwidth range, and K(x) is the kernel function. Previous studies have shown that the selection of kernel function barely affects the results while it is necessary to pay attention to the selection of bandwidth . The selection of h is related to the scale of analysis. Generally, larger values of h correspond to the analysis at the macro scale that reflects the trend distribution,

whereas smaller values of h are helpful to find local characteristics. In addition, the selection of h is generally positively related to the dispersion degree of points. Therefore, in practice it is necessary to adjust h according to the analysis needs and actual effects to achieve optimal results.

At the meso-scale, we focus on the analysis of network communities. We first detect the communities using the Infomap algorithm (Rosvall and Bergstrom, 2008) , and then J o u r n a l P r e -p r o o f use multiple small maps to visualize the dynamic patterns of the detected communities.

Nodes closely connected in the network can form a node subset, or network community (Leskovec et al., 2008) . Community detection aims to divide these nodes into local communities such that nodes in the same community have stronger connections than nodes in different communities. In this study, the node set in one community corresponds to the bicycle station set with more frequent riding connections in the bike-sharing network. The community detection relies not on the proximity of spatial locations but the high strength connections between the nodes. Based on the detected communities, we can analyze their spatial distributions in combination with spatial visual analysis. Dynamic changes of community structure can be further observed by integrating temporal information. In the context of COVID-19, this is especially important for providing a reference for exploring the impact of the pandemic on the bike-sharing network structure.

In addition, as the intermediate scale between the bike-sharing system and station, the network community analysis provides a meso-scale perspective for understanding the characteristics of the sub-networks.

Community detection is a basic research problem of complex networks. For the weighted directed networks in this study, Infomap algorithm is one of the few algorithms suitable for community detection of small and medium-sized networks (Rosvall and Bergstrom, 2008) . The comparative analysis of many studies shows that Infomap performs well for the above-mentioned tasks (Lancichinetti & Fortunato, 2009; J o u r n a l P r e -p r o o f Journal Pre-proof Algesheimer, & Tessone, 2016) . Therefore, this study selects Infomap algorithm for the community detection.

The basic idea of Infomap algorithm is as follows. Based on Shannon's information theory (Shannon, 1948) , Infomap uses random walks on the network to obtain a probability path as the proxy of information flow in the real system. An appropriate encoding scheme is selected to encode the random walk path. The essential goal is to find the optimal community division scheme by minimizing the coding length here.

For the quantitative description of above random walk path codes, suppose that there is a community division scheme M, which divides nodes into m communities, then the average coding length per step describing a random walk can be measured by Equation 8. It consists of two parts: the coding length H(Q) of the movement between communities, and the coding length H(P i ) of the movement within communities. Each part is weighted according to the frequency of occurrence, pout is the probability of random walk switching community, is the probability of moving within community i.

The implementation of Infomap algorithm is mainly divided into the following steps:

1. Each node in the network is regarded as an independent community to obtain the initial community division scheme M1. At this stage, the number of communities is equal to the number of nodes in the network.

2. Randomly sample a node sequence from the network, and take the following actions for each node in the sequence in order:

2.1. Try to assign the node to the communities where its neighbor nodes are The above process can combine with optimization algorithms such as simulated annealing to improve optimization efficiency. Based on the final output from the above algorithm, we can determine the community ownership of each bicycle station in the bike-sharing network. The Infomap algorithm in this study is implemented by Python-Igraph package (Csardi and Nepusz, 2006) .

Both the locations of stations and the flow intensity between stations will affect the community detection results. It is necessary to carry out spatiotemporal analysis of the community results, for example, whether the pandemic will affect the trip volume, trip location, etc., and whether these will further affect the network structure, such as community detection results.

Journal Pre-proof Figure 6 The three-tier structure of network, community and node

Each community can be regarded as a new network constructed by the nodes within the community and the edges between these nodes. As shown in Figure 6 , the community is at the middle scale of the entire network and bicycle stations. Current network statistical analysis in general only focuses on the entire network, and various indicators are described using numerical values, e.g., listed in a table. In this study, we combine the statistical analysis with spatial visualization to visually display various network statistical indicators of the community networks. Different from the previous studies that only focused on the analysis of community division results, this study takes each community as an analysis unit to further carry out a visual comparative analysis between them, observe the temporal character of various indicators, and analyze the network structure more deeply.

The analysis unit at the micro-scale is station, which is the basic unit of bike-sharing data statistics. As the smallest component of the network, the analysis of station mainly J o u r n a l P r e -p r o o f focuses on network topology and network flow in Section 4.2.2. For the network topology, we will investigate the connection relationship between directly connected nodes and indirectly connected nodes in the network respectively. For the network flow, we will investigate the flow of each station.

For the directly connected nodes, node degree is intuitive to reflect the connection relationship between the target node and its directly linked nodes. We first calculate the degree of each node. To investigate the aggregation between the node and its connected nodes, we calculate the local aggregation coefficient. For any node i in the network, its local aggregation coefficient LCi can be calculated by Equation 9 (Fagiolo, 2007) . In

Equation 9, ki is the node degree and the calculation of i k  is shown in Equation 10, if and only if there is an edge connection between node i and node j, aij=1, otherwise aij= 0.

In the bike-sharing network in this study, a high LC value indicates that there is good aggregation between the target station and the stations directly connected to it.

For the indirectly connected nodes, closeness centrality CC can be used to reflect the degree of network proximity between them. If the path from one node to every other node on the network is very short, the node has good connectivity to other nodes and has the We summarize of the complex network indicators of bike-sharing network analysis introduced in this study in Table 2 . In this section, we first analyze the general distribution of trips from the bike-sharing data. We calculate the monthly statistics of four indicators from January to April 2019 and 2020. They are the total number of trips, the average trip duration, the average daily service times of each bike, and the average daily service duration of each bike. We visualize their monthly distributions using Line charts in Figure 7 .

The trip volume (Figure 7 had turned better. On these two days the riding volumes increased dramatically reaching local peaks, which reflects to some extent that the above-mentioned public reports may affect people's travel decisions. In addition, these two days are weekends and the weather was fine 2 which may also stimulate more bicycle travels. There were more ridings in the above two days, making the corresponding curve nodes in the peak positions. Table 3 . During March and April 2020, the proportion of short-term ridings decreased significantly, while the proportion of long-duration ridings (e.g., over 20 minutes) increased significantly. These facts show that during the pandemic period, people reduced the number of rides, but they rode longer. This may because of the following reasons: (1) due to the low safety factor of tight space in subway or bus, bike riding undertook some main commuting tasks during the pandemic period;

(2) riding has become an alternative way of sports and leisure during the closure of entertainment facilities. In addition, the time distribution of riding behavior also changed significantly. In terms of the start time of riding, as shown in Figure 9 , the morning peak during the J o u r n a l P r e -p r o o f Journal Pre-proof pandemic period (in March and April) has been weakened and delayed. This may be related to the start of the home-office or the shifting of the commuting time to avoid social contact clustering. In this section, we calculate the statistics of the complex network indicators based on the equations introduced in Section 4. In order to take into account the impact of seasonal change factors, we calculate and analyze the indicators respectively from January to April of the two years. The computing results are shown in Table 4 The heterogeneity of edge flow in 2020 has experienced an obvious pattern of first decline and then rise, and the valley is in March when the pandemic broke out. At the beginning of the pandemic, the edge flow on the network is reduced and its distribution is more uniform.

However, this trend did not continue with the continuous decrease of edge flow.

To investigate the spatiotemporal patterns of station flows, we divide each day into five time slots: 0:00-6:00, 6:00-10:00, 10:00-16:00, 16:00-20:00 and 20:00-24:00 J o u r n a l P r e -p r o o f Journal Pre-proof (Faghih-Imani and Eluru, 2015) . For each time slots in March and April in 2020 and 2019, we interpolate the station flows using the kernel density estimation method. Recall that in Table 3 there are more short-duration ridings in normal months. The short-duration riding will limit the range of riding activities, making the area connected by riding more local, thus forming a larger number of communities with a smaller range. In the pandemic months, bike-sharing is no longer limited to the short-distance connection, and the mid-or long-distance tasks undertaken by public transportation systems in the past may be partially replaced by shared bicycles. With the increase of the proportion of long-duration riding, the area of riding connection becomes larger and the community coverage expands. In previous research work, the network indicators were oriented to the whole network.

Our approach extends the work to allow a detailed exploration at community scale. We achieve this by first calculating the network connectivity δ, the global aggregation indicator C, the coefficient of variation of station flow CV(F) and the coefficient of variation of edge flow CV(W) of these community networks. The results are shown in Table 5 and visualized in Figure 12 . When calculating a community, F and W are the station flow and edge flow in the community network respectively. Substituting F and W into Equation 6, we can obtain CV(F) describing the station flow distribution of this community network and CV(W) describing the edge flow distribution of this community network. Figure 12 visualizes these statistical indicators of the community network.

In terms of network topology, the network connectivity δ of most communities increased in the normal period and decreased in the pandemic period which is consistent with the change of the overall network indicator. The Long Island region where community 1 is located has low network connectivity in normal times. With the development of the pandemic, it became the community with the best network connectivity in April in 2020.

This region is isolated by natural rivers and less connected with other regions, so the network structure is relatively stable and its connectivity is relatively less affected by the pandemic. For global aggregation indicator C, we can see from Table 5 and Figure 12 In terms of network flow, some regions tend to show phenomena different from the overall trend. For instance, different from the overall downward trend of CV(F) in the two months of 2020 shown in Table 4 , the CV(F) of Long Island region where community 1 is located and northwest Brooklyn region where community 3 is located increased during this period. That is, with the development of the pandemic, the flow distribution of bicycle stations in these two regions is more uneven. For CV(W) of the communities, it can be seen from the data in March 2020 that the arrival of the pandemic has reduced the CV(W)

of most communities, that is, the heterogeneity of edge flow distribution is weaker. This is consistent with the overall trend of the whole network in Table 4 . However, as the development of pandemic, the heterogeneity of edge flow distribution within communities is still declining in April 2020 which is different from the overall trend of the whole network. Overall, the pandemic has a strong negative impact on the stability of the bike-sharing system. It has led to a reduced number of urban riding trips and changed people's riding habits to some extent. Besides, it has changed the connection structure and connection character of the bike-sharing network. All of these changes will further affect the structure of the network community and the characteristic of stations.

These research results can also support the decision-making for bike users, bike companies and government agencies. Firstly, in terms of bike users, on the one hand, the temporal pattern of the riding behavior and the spatial distribution of riding flow can provide useful reference for travel. Users can formulate appropriate travel plans in combination with their location and target region, so as to avoid travel peaks and high-flow areas. This can help reduce aggregation and thus reduce the risk of infection. On the other hand, the decrease of riding volume undoubtedly indicates travel risk, but from the generally increased riding duration, it proves the feasibility of long-distance riding for commuting and leisure purposes which may bring beneficial inspiration to citizens.

Compared with the closed space of public transportation, the riding space is more open and is thus easy for bike riders to maintain a safe distance. Under the premise of adequate personal protection, it serves as an alternative way for commuting and leisure purposes. Secondly, from the perspective of bike-sharing companies, they should pay attention to the loss caused by the decrease of trip volumes during the pandemic and come up with corresponding schemes to reduce economic losses. However, the pandemic may bring new tasks and potentials for bike-sharing, which will also provide a reference for the company's future business strategy. Besides, our work suggests that the companies should pay attention to the hot spot areas of bike flow to dispatch the bikes timely and to disinfect bike, especially the frequently used bikes, to ensure the safe use of bikes during the pandemic. Finally, for the government, the urban public transport system was severely impacted during the pandemic, and many lines may have to be temporally closed to ensure safety. To some extent, bike-sharing provides an alternative possibility to promote urban traffic during the pandemic. In this special period of reducing travel clustering, promoting healthy travels, and making pandemic recovery plans, the government should assess the advantages and risks of bike-sharing riding during the pandemic through rigorous and neutral investigations, and make appropriate decisions.

In terms of local practice, this study provides empirical evidence for more long-duration bicycle riding during the pandemic, which indicates shared bicycles may undertake more tasks for leisure and commuting. It is recommended to add temporary J o u r n a l P r e -p r o o f bicycle stations in leisure areas distributed in the city. In particular, it is necessary to pay attention to avoid riding aggregation when setting up stations. This would be beneficial to increasing the accessibility and convenience of citizens' bike-sharing use and enrich their leisure activities during the pandemic.

Since shared bicycles is a relatively safe way of travel during the pandemic, temporary bicycle lanes could be appropriately added. Although there were a lot of home office work, people such as doctors and nurses still need to go to the workplace during the pandemic and there is a large demand for medical treatment. This measure can promote the convenience of citizens' riding travel during the pandemic. In particular, the location of temporary lanes should be more inclined to hospitals and other places that are critical to resist the pandemic and protect people's life and health.

Further recommendations are on guaranteeing safety for riding. For instance, publishing real-time bike-sharing flow distribution will be helpful for citizens' travel decision-making, and the kernel density visualization in this study serves as a good reference and a starting point. In addition, disinfection should be strengthened at flow-intensive stations considering the spatiotemporal distribution patterns of riding flow found in this study.

The implementation of relevant policies requires financial support. In particular, the subsidy policies for bike-sharing companies need to be formulated as the stable operation of these companies is conducive to urban traffic during the pandemic.

In terms of international practice, for other cities with bike-sharing, out findings could not be overgeneralized but can provide some references in combination with their specific J o u r n a l P r e -p r o o f circumstances. For cities where bike-sharing is not yet popular, we recommend them to increase investment in bike-sharing as a green and safe transportation mode. In addition, bike-sharing not only meets the needs of daily commuting connection and leisure, but also a substitute for public transport during the pandemic. This is beneficial for maintaining the normal operation of the city in critical time.

Our analysis confirmed that bike-sharing data can be used to explore and reveal the impact of the pandemic on urban human mobility. However, the use of bike-sharing data does have some limitations in understanding urban human mobility. The spatial coverage of bike-sharing system in cities is limited and bike-sharing data does not represent all mobility activities. On the one hand, it only records the bicycle usage of bike-sharing companies, but other private bicycle trips cannot be counted. On the other hand, the bike-sharing system is only an integral part of the transportation system, other travel modes, such as buses, subways and taxis, should be also included for a comprehensive understanding of urban mobility.

The research methods proposed in this study has great potential to be applied to other cities worldwide and for the understanding of the impact of the pandemic on urban mobility if there is bike-sharing data available. As a global outbreak, COVID-19 has a huge impact scope. Therefore, we see a wide scope of applications. In fact, these J o u r n a l P r e -p r o o f methods can be applied to typical OD data including bike-sharing data from other cities or other companies and other types of OD data suitable for studying the human mobility.

In terms of research results, some results of this study may be universal, which could provide reference for different cities to understand the impact of the COVID-19 on bike-sharing. However, due to diverse influencing factors of urban human mobility such as the severity levels of the pandemic and local government policies in different cities, the research results of other cities will show personalized characteristics. Dedicated research needs to be conducted for the exploration of urban mobility patterns in specific cities by applying our methods to their data, which also constitutes the motivation for the extension of our methods to other places.

As the COVID-19 pandemic is still on-going, future research will collect latest data from New York and continue to track the impact of the COVID-19 on New York bike-sharing. Meanwhile, we will collect more data from other cities around the world to carry out comparative analysis of the pandemic impact on the use of bike-sharing from an international perspective.

For other aspects, more types of human mobility data and big crowdsourced data like taxi data, flight data, mobile phone data and geotagged social media data will be collected 7 Conclusion

The outbreak of COVID-19 has a huge impact on our societies and greatly reduced the social and economic activities of many cities. In this study, we have used geospatial network methodology to study the bike-sharing system of New York which confirms the feasibility of studying urban human mobility during the pandemic using bike-sharing data.

In the design of research methods, we combine spatial and temporal factors, abstract network indicators and concrete spatial visualization methods to obtain more comprehensive analysis results. Aiming at solving the problems existing in the current research and providing a new perspective for understanding the pandemic impact, a multi-scale geospatial complex network analysis framework is proposed for the exploration of shared bicycle riding. The spatiotemporal variations of the topology and flow of bike-sharing network is explored comprehensively. Many interesting results have been found regarding the changes of the connectivity and aggregation etc., of bike-sharing networks during pandemic. In addition, we combine spatial visualization and complex network analysis to present the spatiotemporal distribution of abstract indicators in an intuitive way.

To answer the questions raised at the beginning of this paper, our analysis showed that the pandemic had a considerable impact on the urban human mobility represented by the shared bicycle riding. Specifically, we discussed that the spatiotemporal patterns of shared bicycle riding changed significantly during the pandemic, reflected in aspects such as riding temporal distribution and the spatial flow distribution. Furthermore, the J o u r n a l P r e -p r o o f multi-scale analysis showed that the pandemic has comprehensively affected the riding network, changed the network community structure, and had a negative impact on the network topology and network flow at different scales. Our research results can serve as a reference for other researchers worldwide conducting similar work and for a world effort to combat the pandemic.

The structure of spatial networks and communities in bicycle sharing systems

Human mobility: Models and applications

Spatial networks

Downtown vibrancy influences public health and safety outcomes in urban counties

Topological evolution of a metropolitan rail transport network: The case of Stockholm

The igraph software package for complex network research. InterJournal, complex systems

The evolving structure of the Southeast Asian air transport network through the lens of complex networks

Understanding taxi driving behaviors from movement data

Better understanding the characteristics and influential factors of different travel patterns in free-floating bike sharing: Evidence from Nanjing

Multilayer dynamics of complex spatial networks: The case of global maritime flows (1977-2008)

A review on bike-sharing: The factors affecting bike-sharing demand

Hail a cab or ride a bike? A travel time comparison of taxi and bicycle-sharing systems in New York City

Analysing bicycle-sharing system user destination choice preferences: chicago's divvy system

Incorporating the impact of spatio-temporal interactions J o u r n a l P r e -p r o o f Journal Pre-proof on bicycle sharing system demand: A case study of New York CitiBike system

Clustering in complex directed networks

Bike share: a synthesis of the literature

Study of covid-19 pandemic in london (uk) from urban context

Exploring network structure, dynamics, and function using NetworkX

Los Alamos National Lab.(LANL)

Extracting commuter-specific destination hotspots from trip destination data-comparing the boro taxi service with Citi Bike in NYC

The influence of campus characteristics, temporal factors, and weather events on campuses-related daily bike-share trips

Community detection algorithms: a comparative analysis

Statistical properties of community structure in large social and information networks

Understanding intra-urban human

Emerging study on the transmission of the Novel Coronavirus (COVID-19) from urban perspective: Evidence from China

Social sensing: A new approach to understanding our socioeconomic environments

Urban mobility in the sharing economy: A spatiotemporal comparison of shared mobility services. Computers, Environment and Urban Systems

Observed impacts of the Covid-19 first wave on travel behaviour in Switzerland based on a large GPS panel

On estimation of a probability density function and mode. The annals of

Maps of random walks on complex networks reveal community structure

Understanding the impacts of a public transit disruption on bicycle sharing mobility patterns: A case of Tube strike in London

A complex network perspective for characterizing urban travel demand patterns: graph theoretical analysis of large-scale origin-destination demand networks

Resilience resistance: The challenges and implications of urban resilience implementation

A mathematical theory of communication

Mapping the bike sharing research published from

Rebalancing citi bike : a geospatial analysis of bike share redistribution

Exploring the network structure and nodal centrality of China's air transport network: A complex network approach

Neighborhood socio-demographic characteristics and bike share member patterns of use

Analysing the spatial configuration of urban bus networks based on the geospatial network analysis method

Check-in behaviour and spatio-temporal vibrancy: An exploratory analysis in Shenzhen

Unravel the landscape and pulses of cycling activities from a dockless bike-sharing system

A spatiotemporal and graph-based analysis of dockless bike sharing patterns to understand urban flows over the last mile. Computers, Environment and Urban Systems

A comparative analysis of community detection algorithms on artificial networks

Evaluating the effect of city lock-down on controlling covid-19 propagation through deep learning and network science models

Mining bike-sharing travel behavior data: An investigation into trip chains and transition activities

Detecting the dynamics of urban structure through spatial network analysis