Investigating impacts of COVID-19 on urban mobility and emissions

The COVID-19 pandemic has severely impacted human activities in a way never documented in modern history. The prevention policies and measures have abruptly changed well-established urban mobility patterns. In this context, we exploit different sources of urban mobility data to gain insights into the effects of restrictive policies on the daily mobility and exhaust emissions in pandemic and post-pandemic periods. Manhattan, the most densely populated borough in New York City, is chosen as the study area. We collect data generated by taxis, sharing bikes, and road detectors between 2019 and 2021, and estimate exhaust emissions using the COPERT (Computer Programme to calculate Emissions from Road Transport) model. A comparative analysis is conducted to identify important changes in urban mobility and emission patterns, with a particular focus on the lockdown period in 2020 and its counterparts in 2019 and 2021. The results of the paper fuel the discussion on urban resilience and policy-making in a post pandemic world.


Introduction
COVID-19 was declared as a global pandemic by the World Health Organization (WHO) on 11 March 2020 (World Health Organization, 2020). Because of the highly contagious nature of the novel coronavirus, many governments imposed national stay-at-home and lockdown restrictions to reduce transmission of the virus. Cities are at the forefront of this global pandemic (Al-Kindi et al., 2021;Liu, 2020). As a key element of cities, urban transport systems facilitate the rapid flow of people and goods, which makes them the most exposed areas to the risk of infection (Barbarossa, 2020). The outbreak of COVID-19 has largely changed urban mobility patterns. Because of COVID-19 related interruptions, both commute and personal trips decreased dramatically in several parts of the world (Abu-Rayash & Dincer, 2020; Beria & Lunkar, 2021;Bucsky, 2020;De Vos, 2020;Shakibaei et al., 2021). Public transport usage recorded the steepest decrease among all urban transport modes, because it was difficult to guarantee physical distancing on public transport (Hu et al., 2021;Politis et al., 2021;Sun & Zhai, 2020;Zheng et al., 2020). There was also a decrease in walking, especially utilitarian walking, during lockdowns (Hunter et al., 2021), but less sharply when compared to public transport (Shakibaei et al., 2021;Teixeira & Lopes, 2020). Cycling, as a healthy and social-distancingfriendly travel mode (Serafimova, 2020), rose substantially during the pandemic in Europe, North America and Australia, particularly after lockdown periods (Buehler & Pucher, 2021).
Meanwhile, lockdown restrictions have influenced the environment, which resulted in 17 % reduction in global CO 2 emissions (Queré et al., 2020). Data released by NASA (National Aeronautics and Space Administration) and ESA (European Space Agency) indicated that pollution in some of the epicenters of COVID-19, such as Wuhan, Italy, Spain and United States, has reduced by up to 30 % (Gautam, 2020;Muhammad et al., 2020;Wang et al., 2020). There has been an increase of 24 % ozone in southern European cities (Sicard et al., 2020) and 17 % in India (Sharma et al., 2020). In urban areas, the reduction of travel demand resulted in significant noise level reductions (Basu et al., 2021;Rumpler et al., 2020;Sharifi & Khavarian-Garmsir, 2020).
In general, COVID-19 has fundamentally changed social practices and the environment in ways that we are still attempting to understand. COVID-19 restrictive measures have directly impacted urban mobility patterns, which inevitably induced indirect impacts on various aspects, e.g., mode shift, resilience, emission level, etc. It is necessary to investigate both direct and indirect impacts of COVID-19 on urban mobility. Most studies on urban mobility and emissions are based on the biases from the lockdown period compared with the pre-pandemic period (Sicard et al., 2020). However, the current necessity for social distancing might have long-term, or even permanent, effects on urban mobility and emissions (Honey-Roses et al., 2020). Keeping track of these changes after the end of the lockdown is important for future urban transport planning. This paper selects Manhattan Island in New York City (NYC) as a case study and presents a comparative analysis of mobility and emission patterns during the pre-pandemic, pandemic and post-pandemic periods. The results of the paper fuel the discussion on urban resilience and policy-making in a post pandemic world. The main contributions of this study are summarized as follows: • The majority of existing works on COVID-related emission changes focus on either analyzing temporal evolution of emission changes throughout different pandemic periods (e.g., (Putaud et al., 2021) and (Baldasano, 2020)), or quantifying the magnitude of emission changes due to the pandemic at relatively high levels (e.g., city level, county level or country level) (e.g., (Wang et al., 2021) and ). This paper provides a detailed analysis on both spatial and temporal emission pattern changes contributed by the COVID-19. Moreover, this study explores the pandemic-induced shifts in urban mobility, especially trip purpose and network speed, using clustering tools, which are rarely touched by previous studies. • Public Transport, the most important mode of transportation in urban areas, were suspended to prevent people from being infected with COVID-19 during certain pandemic periods. In contrast, active travel, such as walking and cycling, has been promoted for personal wellbeing. It is needed to identify the role that active travel can play during disruptive events. This study provides evidence on the resilience of bike sharing systems during this pandemic.

Literature review
As a response to COVID-19 restriction measures, residents in different countries have changed their daily travel patterns (Beck & Hensher, 2020;Eisenmann et al., 2021;Oum & Wang, 2020). For example, it was reported that the total number of daily trips reduced by 57 % in Budapest, Hungary because of COVID-19 restriction measures, where car usage increased from 43 % to 65 %, bicycle usage increased from 2 % to 4 %, and public transport decreased from 43 % to 18 % (Bucsky, 2020). Eisenmann et al. (Eisenmann et al., 2021) stated that public transport lost ground during the lockdown period whereas individual modes of transport, especially the private car, became more important in Germany. The study based on data collected from 11 cities in the United States recorded different effects of COVID-19 on nonmotorized activities (Zhang & Fricker, 2021): the non-motorized activities increased in less densely populated cities, but decreased in densely populated cities. In Brazil, the use of private vehicles grew as the main mode of transport to the principal activity (Oestreich et al., 2023). Various methods have been applied to study the COVID-19 impacts on activity and travel behavior patterns, such as online surveys (De Haas et al., 2020), stated preference surveys (Shamshiripour et al., 2020) and objective data measures via GPS Logger and Travel Diary App (Axhausen, 2020).
A number of studies compared the emission levels before and after lockdown, witnessing the significant reduction in vehicular emissions. The study (Wang et al., 2021) employed a quantitative analysis on air pollution changes due to restriction measures across 325 cities in China, showing a reduction of 3.3 %, 3.3 %, 4 %, 15.3 % and 13.1 % in CO, NO 2 , SO 2 , PM 10 and PM 2.5 respectively. Five cities in India also witnessed improved air quality levels during lockdown periods, especially for NO 2 concentrations (Nigam et al., 2021;Singh & Chauhan, 2020). These results are in agreement with other studies conducted in South Korea (Seo et al., 2020), Saudi Arabia (Anil & Alagha, 2021), Turkey (Celik & Gul, 2022), Iran (Broomandi et al., 2020) and other Asian counties (Roy et al., 2021). Chen et al. (Chen et al., 2020) analyzed the effects of lockdown measures on air pollutants over 28 cities in the United States. The authors observed consistent reduction in CO and NO 2 which was mainly contributed by the mitigation of transporting activities. Similarly, two studies (Tian et al., 2021) (Mashayekhi et al., 2021) from Canada recorded significant reduction in CO and NO2 during lockdown. 20 %-40 % emission reductions resulted from decreased transportation and industrial activities were observed in several European countries (Skiriene & Stasiškieně, 2021). The studies (Collivignarelli et al., 2020;Putaud et al., 2021) based on data from Italy showed a remarkable decrease in NO 2 , while decreasing rates of particulate matters were not that significant. Similar findings were drawn from other European countries, such as Spain (Baldasano, 2020), Poland (Filonchyk et al., 2021) and United Kingdom (Jephcote et al., 2021).
According to (Bohler et al., 2021), "city resilience focuses on increasing-or at least securing-the performance of urban systems in the face of multiple hazards in crises, rather than preventing or solely mitigating the loss of assets due to a specific event". In the field of transportation, resilience is seen as the ability to absorb, reduce and resist the effects of disturbance, maintaining an acceptable level of service and restoring the regular and balanced operation within a reasonable cost and time (Gonçalves & Ribeiro, 2020). Changes in people's daily travel patterns during the COVID pandemic encouraged a set of actions by governments to reset mobility so as to achieve its resilience, such as promoting active travel (e.g., walking and cycling) and mobility devices (e.g., e-scooters and e-bikes) (Mouratidis & Papagiannakis, 2021;Putaud et al., 2021;Serafimova, 2020;Sultan et al., 2021;Valenzuela-Levi et al., 2021). A positive impact of shared mobility services on urban transportation resilience was recorded during the COVID pandemic, although several cities have included them as part of micro-mobility strategies (Dias et al., 2021). The utilization of nonshared individual transport modes, e.g., bikes, has also risen during the pandemic period (Rahman et al., 2021;Shortall et al., 2022), where the resilience and reorganization capacity of public spaces are essential to boost individual mobility and micro-mobility.
By reviewing the COVID-19 related studies, it is found that the majority of works investigating the pandemic impacts on travel behavior reported a significant drop in the willingness to travel during the pandemic. Of all transport modes, the COVID-19 had the most direct impact on public transport. A modal transfer from public transport to sharing bike system was observed (Teixeira & Lopes, 2020). It is necessary to further investigate the resilience of sharing bike system during the pandemic. Moreover, researchers found a significant reduction in vehicular emissions which was mainly contributed by the mitigation of transporting activities. However, most of existing studies focus on emission changes at relatively high levels (e.g., city level or country level). More detailed analysis on spatial and temporal emission pattern changes is needed.

Data description
In this study, taxi trip data, bike sharing trip data and traffic speed data were used. As there is no detailed emission data and motorized vehicle data available in the study area, taxi trip information was employed to estimate emission levels as well as to reflect motorized vehicle activities instead. Similarly, as it is hard to collect trip information of all the non-motorized vehicles in the study area, sharing bike trip data was selected to represent non-motorized vehicle activities.

Study area
Manhattan, one of five boroughs of NYC, was selected as the study area. Manhattan's daytime population is approximately 3.94 million in an area of just 23 mile 2 , including 41 % commuting workers, 37 % local residents, 10 % out-of-town visitors, 9 % local day-trip visitors, and 3 % hospital patients and commuting students (Moss & Qing, 2012). This makes Manhattan the most densely populated of the five boroughs of NYC.

Taxi trip data
We extracted taxi trip records generated by both medallion taxis (yellow taxis) and street hail liveries (green taxis) in 2019 and 2020 from the NYC Taxi and Limousine Commission (TLC) database. 1 Note that the taxi data in 2021 was not available while the research was being conducted. The collected yellow and green taxi trip records include pickup and drop-off dates/times, pick-up and drop-off locations, and trip distances reported by taximeters.

Bike sharing trip data
The second dataset comes from a bike-sharing service in NYC -Citi Bike. Bike-sharing trip records were collected from the Citi Bike database 2 during the lockdown period in 2020 as well as the same periods in 2019 and 2021. The collected bike trip records consist of start and end dates/times, and start and end station positions.

Point of interest data
In order to investigate the influence of different trip purposes on bike usage, we obtained the Point of Interest (POI) information from the NYC open data website. 3 In this study, seven categories of POI points were extracted, namely, 1) residential, 2) education facility, 3) recreational facility, 4) transportation facility, 5) commercial, 6) government facility and 7) health services. Bike stations were then labeled by these seven categories on the basis of their closest POI points.

Traffic speed data
Lastly, we gathered traffic speed records during the lockdown period in 2020 as well as the same periods in 2019 and 2021. The original speed data was collected by the Traffic Management Center (TMC) of NYC Department of Transportation (NYCDOT). 4 The dataset contains speed information generated in every five-minute period from major arterial roads. It should be noted that the dataset only provides average speed between end points on a link in every interval without traffic volume or occupancy information.

Emission estimation
As there is no available emission data for the study area, we estimated traffic emissions based on the collected taxi trip records. In general, there are two types of emission estimation models, namely, microscopic and macroscopic ones (Boulter & McCrae, 2007;Samaras et al., 2019). Microscopic emission models can estimate the instantaneous emissions by relating emission rates to vehicle operation on a second-by-second basis. Models of this type, such as PHEM (Passenger car and Heavy duty Emission Model) (Hausberger et al., 2015) and MOVES (Motor Vehicle Emissions Simulator) (United States Environmental Protection Agency, n.d.), require detailed speed profiles of vehicles as an input. While macroscopic models, such as COPERT (Ntziachristos et al., 2009) and ARTEMIS (Assessment and Reliability of Transport Emission Models and Inventory Systems) (Boulter & McCrae, 2007), have been developed to estimate fleet emissions over a region using average speed as an input. As the NYC taxi dataset lacks the detailed vehicle kinematics information (e.g., instant speed) that is required by microscopic models. We adopted the COPERT macroscopic model to estimate exhaust emissions. Aforementioned ARTEMIS and COPERT models share many similarities. More specifically, COPERT has been developed in the framework of ARTEMIS project. Both of them use the average speed to predict emissions, with the produced factors being expressed in mass of pollutant per unit of distance travelled. They share several speed dependent emission factors (Saharidis & Konstantzos, 2018). In the previous study (Sjodin & Jerksjö, 2008), COPERT and ARTEMIS models were evaluated against the same set of on-road optical remote sensing emission data. The results showed that the performance of both models has been proven for several vehicle categories, traffic situations and pollutants.
The COPERT model utilizes mean driving speed and travel distance to estimate the related exhaust emissions. Unitary emission factors consist of speed continuous functions that are representative of the traffic conditions encountered (Andre & Rapone, 2009). The emissions are computed as the product of the total travel distance and the unitary emission factors, given by: where, E i is the emissions (g) of pollution i; F i,j is the emission factor (g/km) of pollutant i for vehicle category j, V j is the mean driving speed (km/h) for category j; N j is the number of vehicles of category j; and D j is the average distance (km/veh) driven per vehicle of category j.
NYC taxi data provides travel distance and duration information for each trip. Thus, we modified formula (1) to calculate taxi exhaust emissions, which can be expressed as: where, D k and T k are the travel distance (km) and duration (h) of trip k, respectively.
During 2019-2021, over 20 vehicle models from 7 automotive manufacturers were approved for use as a yellow taxi in NYC. 5 In order to simplify the calculation process, all the taxis are assumed to be EURO 5 passenger cars. In Fig. 1, the Carbon Monoxide (CO) and Nitrogen Oxides (NO x ) emission factors with corresponding driving speed categories for the EURO 5 passenger car are depicted. Note that the model handles only mean speeds higher than 5 km/h. We used the default emission factor values, as there is no detailed emission data available to validate the model.

Density-based clustering
In order to identify cycling usage behaviors across different docking stations, we segmented stations into clusters considering of their usage patterns and trip purposes. Although clustering is a powerful unsupervised knowledge discovery tool, each algorithm is sensitive to certain parameters. In partitioning-based clustering algorithms (e.g., k-means and its variants (Garg, 2006;Hartigan & Wong, 1979)), the number of clusters to be obtained is required to be given a priori; while hierarchical-based clustering methods (Guha et al., 1998;Tian et al., 1996) usually require manual intervention to decide when to assign finished clusters. The properties of density-based clustering algorithms include detection of arbitrary shape and different sized clusters, and therefore does not need the number of clusters as an input parameter. Nevertheless, the most common density-based approach, DBSCAN (Density-Based Spatial Clustering of Applications with Noise) (Ester et al., 1996), has a problem in detecting meaningful clusters in data of varying density. Therefore, we employed OPTICS (Ordering Points To Identify the Clustering Structure) (Ankerst et al., 1999), which was proposed to address the weaknesses of DBSCAN.
In this study, docking station demands and their POI features were chosen as clustering inputs. Station demands were normalized to [0,1] using min-max normalization. Categorical POI features were one-hot encoded with a single high (1) bit and all the others low (0). In total, 8 input features were generated for each station (1 for station demand and 7 for POI features).

Time series clustering
In the NYCDOT dataset, the speed profile of each link is represented as a time series. In order to extract underlying patterns of time series, time series clustering was performed. Usually, time series clustering involves three parts -similarity measure, prototype computation and clustering algorithm (Li et al., 2019). The clustering algorithm is used to cluster time series based on series similarities. The similarity between time series is calculated using the selected similarity measure. During clustering, several prototypes are generated to summarize the key characteristics of all series in given clusters.
The fundamental issue in time series clustering is to measure the similarity among time series. The Euclidean distance (Faloutsos et al., 1994), the Dynamic Time Warping(DTW) (Sakoe & Chiba, 1978), the Hausdorff distance, the hidden Markov model (HMM) (Smyth, 1996) and the longest common subsequence (LCSS) (Oestreich et al., 2023) are the most popular similarity measures for time-series data (Aghabozorgi et al., 2015). Due to its sensitivity to small distortions in the time axis, the Euclidean distance may fail to yield an intuitive measure of similarity. The Hausdorff distance and HMM are time-consuming and have poor performance (Morris & Trivedi, 2009). LCSS focuses more on shape similarity and has a high time cost. It is easier for DTW to find the shape similarity between speed series. DTW aims to find an optimal alignment between two given time series by warping them in a nonlinear fashion to match each other, which makes it possible to produce a more intuitively correct measure of similarity. Therefore, DTW was selected as the similarity measure in this study. Although the mean and PAM (Partition Around Medoids) are commonly used prototype computation functions, they do not consider the unique feature of DTW. Therefore, we employed DBA (DTW Barycenter Averaging) (Petitjean et al., 2011) that was designed for DTW. K-means was selected as the clustering algorithm, which has been proven to be effective in time series clustering (Gullo et al., 2012;Hautamaki et al., 2008).

Results
This section presents a comprehensive analysis of the pandemic impact on urban mobility and exhaust emissions using the datasets and methods described in the previous sections. A NYC 2020 historical timeline 6 of COVID-19 is listed as follows: • On March 1, the first case of COVID-19 was confirmed in NYC.
• On March 16, all public schools were closed, and bars and restaurants were closed the following day. • On March 20, New York State on PAUSE executive order was announced. • On June 8, NYC entered Phase 1 of reopening.
• On June 22, NYC entered Phase 2 of reopening. • On July 6, NYC entered Phase 3 of reopening.
• On July 19, NYC entered Phase 4 of reopening.
• On September 29, elementary students returned to public schools across NYC.

Taxi trips and estimated emissions
A comparison of taxi pollutant emissions in 2019 and 2020 is shown in Fig. 2. During the first two months of 2020, we can observe a reduction in taxi emissions with a trend that intensifies gradually. This might be because safety concerns started weighing on the collective conscience. After imposing the PAUSE order on March 20, both NO x and CO emissions dropped to their lowest levels within 2019-2020. It was only by Phase 1 of reopening that emissions started to rise. This trend continued over the four phases of reopening. Then, the emissions reached relatively stable levels after public schools were reopened. Table 1 summarizes changes in taxi demands and emissions for two specific periods: lockdown period (from March 20 to June 8) and afterreopening period (from July 19 to December 31). During the lockdown period, taxi demands and associated emissions decreased by around 96 % compared with the same period of the previous year. After a series of reopening phases, there were still 78 %, 77 % and 77 % reductions in daily taxi demands, NO x and CO emissions respectively, when compared to the levels of 2019.
Although non-essential workers were required to stay at home during the lockdown, weekly fluctuations were still observed in Fig. 2. Thus, we plot the weekday and weekend hourly emissions during the lockdown period as well as the same period of 2019 (see Fig. 3). As NO x and CO emissions shared similar patterns, only estimated NO x results are  Fig. 3 and later in Fig. 4. The plot shows two peak periods for weekdays in 2019, where the evening peak was wider and higher than the morning peak. In 2020, taxi emissions during weekends were generally less than working days, which might be because people avoided unnecessary trips during weekends. During the lockdown, weekday peaks were not noticeable. This probably resulted from reduced commute trips.
Spatial distributions of total NO x emissions during the lockdown and the same period in 2019 are plotted in Fig. 4. Estimated emissions were mapped to taxi zones 7 across Manhattan. Although in 2019 taxis produced an order of magnitude more emissions than 2020, spatial distributions of emissions in these two years were similar. Midtown Manhattan experienced a higher level of emissions, while less densely populated areas had less emissions.
We computed the histograms of taxi travel speeds during the lockdown period and the same period of the previous year, as shown in Fig. 5. The mean travel speed of a taxi trip during the lockdown period was 25.15 km/h with a deviation of 9.11 km/h, while during the same period of the previous year, the mean travel speed was 15.95 km/h with a deviation of 7.05 km/h. Higher and more variable travel speeds observed during the lockdown can be mainly attributed to less congested traffic. Fig. 6 shows the average daily bike-sharing usage and number of active docking stations during the lockdown as well as the same periods of 2019 and 2021. In 2019, the average daily bike-sharing usage was 47,573 trip/day, which dropped to 25,825 trip/day during the lockdown. However, the reduction in bike usage (45.7 %) caused by the lockdown was not as significant as the reduction in taxi usage (96 %). The highest daily bike usage, 57,012 trip/day, was recorded in 2021, indicating 19.8 % and 120.8 % increases compared to the 2019 and 2020 cases, respectively. These findings suggested that bike-sharing, as a transport mode that minimizes close contact with others, has become more and more popular in the pandemic and post-pandemic periods. Moreover, a growth of bike-sharing services was recorded: the number  of stations generating trips increased from 420 in 2019 to 743 in 2021, which might also be one contributor to the bike usage increase. In Fig. 7, hourly bike-sharing usage for weekdays and weekends during the lockdown and its counterparts in 2019 and 2021 are plotted. For weekdays in 2019 and 2021, bike usage followed a bimodal pattern, peaking at rush hours in the morning and the late afternoon; while during the lockdown, only an afternoon peak was seen. The impact of the lockdown on bike usage during weekends was limited. Weekend bike usage in 2019 and 2020 followed similar patterns, whereas the usage level in 2021 was generally higher than the previous two years.

Bike-sharing usage
In order to analyze the impact of trip purpose on bike-sharing usage in the pandemic and post-pandemic periods, docking stations were clustered based on their demands and POI features. The clustering results for the lockdown period and corresponding periods in 2019 and 2021 are summarized in Table 2. Under each POI feature category, the proportion of stations grouped into each cluster was shown. For example, in the 2019 case, 32 %, 0 %, 0 % and 68 % "Residential" stations were grouped into cluster 1, cluster 2, cluster 3, and cluster 4,      respectively. Key POI features for each cluster are highlighted and average bike usage for each cluster is listed. It can be easily seen that stations linked to transportation facilities, government facilities and health services were always the busiest ones regardless of the lockdown. Trips generated by these facilities were unavoidable to some extent, even during the lockdown. Stations with education facilities nearby maintained a similar usage level before and after the lockdown, while during the lockdown this type of trip reduced because of school closures. Remarkable changes were observed in trips generated by recreational and commercial facilities. Bike-sharing usage related to commercial facilities was at its highest level in 2019, dropped sharply during the lockdown, and then grew again in 2021 but was still below the level of 2019. The changes in recreational trips were complex: in 2019, 100 % stations were included in the first two highest-usage clusters (cluster 1 and cluster 2); whereas, in 2021 only 66 % stations remained in the highest-usage cluster, and the other 34 % of stations experienced the lowest usage level among all the five clusters. Bike-sharing usage in residential areas was relatively stable: the majority of these stations had similar bike demands before, during and after the lockdown.

Average link speed
Since the original traffic speed dataset provided by NYCDOT only contains average speed between end points on every link without associated traffic volume or occupancy information, it is unable to calculate a system-wide average speed of all the links. Thus, in order to give a synthetic view of the information, we performed time series clustering and prototype computation on speed series. After preliminary analysis, 22 links that can provide complete information covering the whole study period were selected. We first applied the clustering method illustrated in Section 3.3 on speed data collected between March 20 and June 08 in 2019. 4 clusters were obtained. Then, for each cluster, we computed the prototypes of corresponding speed series from the same periods of 2020 and 2021 in order to conduct a comparative analysis. Fig. 8 illustrates the prototypes computed for different clusters and periods. We observe that due to the reduced traffic demand, average link speed during the lockdown was much higher than its counterparts in 2019 and 2021. In both of the 2019 and 2021 cases, a clear afternoon peak was observed. However, the pandemic slightly impacted the afternoon peak pattern. In 2019, the afternoon peak mainly occurred around 5 pm, which shifted to about 4 pm in the pandemic and postpandemic periods. This was probably because of the COVID-caused flexible working trends. When compared to the 2019 case, traffic speed between 8 pm and 12 am in 2021 increased in two clusters (cluster 2 and 3), indicating improved traffic conditions on the links belonging to these clusters.

Resilience of sharing bike versus taxi
Although the results presented in the Sections 5.1 and 5.2 show that the pandemic negatively affected both taxi and bike sharing usage, we still witnessed the resilience of the sharing bike system (BSS) over the taxi system: despite the remarkable decrease in the BSS usage, it was still much smaller than the taxi system (− 46 % versus − 96 %). This is also supported by the modal transfer from the taxi to the BSS presented in the following paragraphs.
In order to further study the competitiveness and possible modal transfers from taxis to sharing bikes, we computed the ridership ratio of the BSS daily ridership and the taxi daily ridership during the lockdown: ridership ratio = BSS daily ridership /taxi daily ridership (3).

Fig. 8.
Clustering results for link speed series. 4 clusters was first obtained from speed series in 2019. Then, for each cluster, we computed the prototypes of corresponding speed series from the same periods of 2020 and 2021 in order to conduct a comparative analysis.
The bike/taxi ridership ratio could help us further identify the potential role of BSS in improving urban transport resilience during COVID-19 by analyzing how the substitution between taxi and sharing bike ridership changes throughout the pandemic, i.e., how BSS can fill the gap in the taxi system to some extent when its normal function is being interrupted. Table 3 summarizes the sharing bike/taxi ridership ratios computed for the lockdown in 2020 and the same period in 2019. It is shown that the ridership ratio increased from 0.23 in 2019 to 3.04 in 2020, indicating a significant modal shift from the taxi to the sharing bike during the pandemic. It is interesting to see that the weekend ridership ratio experienced much higher increase than the weekday ridership ratio. All of these results provide evidence on the importance and resilience of bike sharing systems during the pandemic.

Conclusions
Well-established urban mobility patterns have been abruptly changed due to due to COVID-19 restrictive measures. In this context, we exploited different sources of urban mobility data to gain insights into the effects of COIVD-19 on the daily mobility as well as emissions in the pandemic and post-pandemic periods. Manhattan, the most densely populated borough in NYC, was chosen as the study area. We collected data generated by taxis, sharing bikes, and road detectors between 2019 and 2021, and estimated exhaust emissions using the COPERT emission model and taxi trip records. A comparative analysis was conducted to identify important changes in urban mobility and emission patterns, with a particular focus on the lockdown period in 2020 and its counterparts in 2019 and 2021. Important findings drawn from the results are summarized in the following paragraphs.
Remarkable reductions in taxi trips and exhaust emissions were witnessed during the lockdown period: taxi trips and emissions reduced by 96 % compared to the same period of the previous year. During the lockdown, weekend taxi demands were much lower than weekdays, while such a scenario only happened during rush hours in 2019. Although there was a huge difference between 2019 and 2020 emission levels, their spatial distributions were similar.
The need for social distancing has encouraged the utilization of bikesharing services. Compared to the same period in 2019, there was a 19.8 % increase in bike-sharing usage in 2020; even during the lockdown, the reduction in bike usage (45.7 %) was not as significant as that of taxi usage (96 %). The bike usage reduction during the lockdown was mainly caused by weekday commute pattern changes: in 2019 and 2021, bike usage followed a bimodal pattern, while during the lockdown, daytime bike usage dropped dramatically and only an afternoon peak was observed. The clustering results suggest that stations close to transportation facilities, government facilities and health services experienced the highest usage levels regardless of the lockdown. Compared to the pre-pandemic period, stations with recreational and commercial facilities nearby generated fewer bike trips during the pandemic and post-pandemic periods.
The clustering results computed using traffic speed series revealed certain changes in traffic conditions: the afternoon peaks have shifted to 4 pm since the lockdown, which occurred around 5 pm in the prepandemic period. Moreover, nighttime traffic conditions on certain roads have improved since the lockdown.
The results clearly show the resilience of the bike sharing system over the taxi system regarding the loss of ridership. Although both systems recorded their ridership plummeting, the bike sharing system experienced a less remarkable ridership decrease than the taxi system (46 % drop in sharing bike ridership versus 96 % drop in taxi ridership). This is supported by the increase in ridership ratio between the sharing bike and taxi. The sharing bike/taxi ridership ratio experienced over 12 times increase because of the pandemic, rising from 0.23 in 2019 to 3.04 during the lockdown in 2020. Such modal transfer suggests the potential of bike sharing systems for the resilience of urban transport systems during disruptive events, because it can quickly provide alternative transport options to urban residents.
COVID-19 has severely impacted human activities in a way never documented in modern history. It does, however, provides a once-in-alife-time opportunity to reconfigure future urban transport. Our study demonstrates an increasing demand for bike-sharing services due to disease transmission concerns. Local governments can leverage the increasing interests in individualistic forms of mobility through more non-motorized-mobility-friendly urban planning. Special attention should be given to investments in physical mobility infrastructure, including extensive cycling-network extension to promote active travel (such as walking and cycling) and mobility devices (such e-scooters and e-bikes). It is worth noting that cycling also has its own limitations. Firstly, cycling is more user-friendly to people in good health, while people in poor health or with disabilities are less likely to utilize bicycles. Secondly, bicycle usage is greatly impacted by road conditions, weather conditions, safety issues, etc. For example, adverse weather conditions may prevent people from using bicycles. Bicyclists are more vulnerable and usually associated with higher accident injury severity level than motorized road users. Lastly, cycling, by its nature, is a slow mode of transport, which is mainly used for short distance trips. In most cities, modal share for cycling is much smaller than transport modes that are faster and more suitable for relatively long-distance trips.
The shifted peak hours observed in our study show the potential of relieving traffic congestion during peak hours by promoting flexible working. A global survey (Citrix, n.d.) concluded that 70 % of respondents believed their productivity at home to be the same or higher than at the office. The study (Hensher et al., 2021) reported a 54.02 % reduction in the Pre-COVID-19 total time costs in Sydney which has important implications for road investment linked to congestion. Thus, it is reasonable to support the appeal of working from home as a policy lever to reduce levels of congestion on the roads and crowding in public transport.

Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability
Data will be made available on request.

Acknowledgements
This study was sponsored by the Engineering and Physical Sciences Research Council (EPSRC) (Project No. EP/R035199/1).

Appendix A. Clustering models
The OPTICS model requires two parameters: r, which describes the maximum distance (radius) to consider, and MinP, describing the minimum number of points required to form a cluster. The minimum number of stations in each cluster MinP was set to 20 which equals to 2.5*number of features (8 features in total: 1 for station demand and 7 for POI feature). And the maximum distance r was set to infinity so as to allow clustering across all scales. Fig. A1 depicts the reachability plots for clustered stations. A reachability plot displays the points whose y value is the reachability distance between two consecutive points. Points belonging to a cluster have a low reachability distance to their nearest neighbor, the clusters show up as valleys in the reachability plot. The deeper the valley, the denser the cluster. It can be observed that the OPTICS model adopted in this study can effectively segment bike stations into clusters.
In the speed series clustering, the performance of k-means is sensitive to two pre-specified parameters: the cluster initialization and number of clusters. Since k-means is a greedy algorithm, certain initial partitions may lead to local minima. Therefore, we ran the algorithm with 10 random initial partitions and chose the one with the best performance. The number of clusters is the most critical parameter in k-means. We used the DB (Davies-Bouldin) index to determine the appropriate number of clusters. Fig. A2 shows the DB values computed for different numbers of clusters. The results showed that 4 clusters yielded the lowest DB value, indicating the best performance. Thus, the optimal cluster number = 4 was used in this study.