Are footpaths encroached by shared e-scooters? Spatio-temporal Analysis of Micro-mobility Services

Research has studied the spatial distribution of shared e-scooters in different cities. In Hosseinzadeh et al. [3], the influence of factors relating to demographics, diversity, and design on e-scooter trips were evaluated. They implemented a Geographically Weighted Regression model, considering data from 159 traffic analysis zones in Louisville, KY. According to the global coefficient values of the model, 18-29 age population, male percentage, commercial land use, and mixed land use were identified as significant features related to e-scooter trips. A study conducted in Minneapolis [8] analysed the usage of e-scooters and their association with populations of different age groups, lane design, land use, and POI counts. The researchers used street segments as the geographical unit of analysis. They deployed Negative Binomial Regression (NBR) models for four different time frames on weekdays and another model for weekend trips, arguing the impact on related features vary based on the time of e-scooting. According to their findings, e-scooters are more popular among the 18-34 age group. Residential land use affects more morning and night trip generation, while commercial land use shows a higher impact on weekend and weekday-midday models. Bai and Jiao [4] compared e-scooter usage in Austin and Minneapolis, and explored whether the features affecting usage differ in the two cities. By implementing NBR models, they observed some features such as proximity to the city center, land use diversity, and transit accessibility positively correlate with e-scooter rides in both cities, while features like commercial land use and park availability were significant only in one city.

For our study, we selected spatial features from the literature [3, 4, 8] and added family composition as a new feature, since it was identified as an influential factor in travel mode selection [9, 10]. Without being limited to a single regression model, we experimented with four machine learning models and selected the model with the least error metrics in prediction. Then the selected model was used to find most associated features with e-scooter trips. Tokey et al. [8] points out the importance of implementing additional models to different time frames, since the influence on POIs and other features may be dynamic in different temporal dimensions. Seeing the temporal variation of e-scooter usage in Melbourne, we implemented models dividing trip data into 10 (5-weekday and 5-weekend) segments.

When exploring the factors impacting shared e-scooter usage, researchers also looked at dynamically changing features such as weather and gasoline price. Hosseinzadeh et al. [1] investigated the influence of the day of the week, weather, and holidays for shared e-scooter and bike share trips in Louisville. They estimated the daily trip frequency using a Negative Binomial Generalized Additive model. Based on the coefficients of the model, weather features (e.g., rainfall, wind speed, mist) were associated with a decrease in the number of trips, and holiday was a factor that increases the trip count. Another study analysed the sensitivity of weather, gas prices, and holidays on the number of shared e-scooter and station-based bike share rides in Washington, D.C [6]. In that work, the authors implemented NBR models for both micro-mobility systems separately and compared the determinants. According to their findings, warmer weather increases the instances of trips, while humidity and precipitation negatively impact trip counts. However, they observed that e-scooter trips are less sensitive to weather changes compared to bike rides. Lesser physical effort needed and the convenience of parking for e-scooters were the possible reasons given for that observation. Holidays and gas prices showed a positive effect on both e-scooter and bike share trips.

Since we observed sufficient variations in the number of e-scooter rides hourly as seen in Figure 1, we choose to build the model with hourly granular data instead of daily. With the availability of data in Melbourne, we selected a set of weather features from the literature [1, 6]. In addition to the global model to estimate the number of hourly rides in the study area, we implemented three models for selected Statistical Areas. The expectation was to analyse the variation of temporal trip determinants (weather, time, day) in different geographical zones.

Tokey et al. [8] and Heumann et al. [11] highlighted the importance of examining the presence of POIs when analysing e-scooter usage patterns. E-scooter riders participated in interviews of a study expressed that they can reach closer to their destinations with e-scooters because of the free-floating nature of the vehicle [2]. Li et al. [12] states trip data from dockless services such as e-scooters can be much closer to real origin and destination locations of user trips. Based on these reasons, we decided that POIs near trip start and end locations can reveal important information about usage patterns and trip intentions. Therefore, we conducted a buffer analysis to explore the impacts on POIs in different temporal dimensions.

Buffer analysis is a common approach used in Geographic Information Systems to examine proximity [13]. It creates a polygon around a feature of interest that contains an area of a specified width. Jin et al. [14] employed buffer analysis to find the spatio-temportal relationship between Uber pickup locations and public transit coverage in New York City. They analysed the number of pickups within three buffer areas around public transit stops and found that Uber both compete with and compliment public transport. In particular, their analysis on different time frames revealed competition is evident mostly in daytime, in areas with good public transit coverage, while Uber acts as a complement at midnight in areas with less public transit service. Another study used buffer analysis to explore the impact on shared e-scooters on public bus ridership [15]. They implemented regression models to estimate route level bus ridership, using the variables derived from buffer analysis such as the number of e-scooter start-trips, end-trips and population within the buffer area.

Given that e-scooter trip start/end locations are closer to riders’ actual origin/destination, we explored the difference between the number of trips started and ended having POIs within the buffer for multiple time frames on weekends and weekdays. Further, the number of trips started and ended within three buffer zones near paths (footpath, cycle lane, and shared path) were compared to explore the effect of e-scooter parking on different types of paths.

The city of Melbourne, Australia started an e-scooter trial partnered with Lime and Neuron micro-mobility service providers on 1st February 2022. Although the trial started with a fleet of 750 e-scooters for each provider, due to the growing demand, providers increased the number of e-scooters deployed in the city [16].

We accessed e-scooter trip details using the General Bikeshare Feed Specification API¹¹1A standardized data feed for shared mobility system availability: https://github.com/MobilityData/gbfs provided by Lime. The free bike status API endpoint provides a real-time snapshot of available e-scooters in Melbourne. The API response consists of a unique ID for each e-scooter, the geographical coordinates, and the battery level of the vehicle. We were able to derive trip-starts, and trip-ends by continuously collecting these API responses in regular time intervals as proposed by McKenzie [17]. For example, an e-scooter that appears in the available vehicle list at $t=0$, does not appear in the available list from $t=1$ to $t=3$ and reappears at $t=4$ indicates that the particular e-scooter was in use from $t=1$ to $t=3$. Hence, $t=1$ can be identified as a trip-start and $t=3$ as a trip-end.

We implemented a scheduled python script to retrieve data from the API every minute. One limitation we observed in the dataset is, the unique ID assigned to each e-scooter was changed every 15 minutes for data security purposes. Due to this limitation, we could not identify complete trips (i.e., match each trip-start with a unique trip-end) from the dataset. As a solution, we aggregated trips into 15-minute time intervals and analysed trip-starts and trip-ends separately. Hence, the aforementioned limitation does not impact the analysis presented in this paper.

This study used shared e-scooter trip data collected from 1st August 2022 to 30th October 2022. The city of Melbourne trial operating zone consists of 350 Statistical Areas (SA) [18]. Figure 2 shows the distribution of e-scooter trip density over the study area. We presented the trip density (trips/sq. km.) instead of the trip count, since the area of different SAs are not uniform. Trip density $TD$ was calculated for each SA as follows:

Where $T(r)$ is the number of trips that occurred in region $r$ and $A(r)$ is the area of that region. We used trip-start locations, and trip-start timestamps to conduct our spatial analysis.

To find the spatial factors most associated with e-scooter trips, we modeled a set of demographic and land use variables representing each SA. These variables were selected based on the literature [3, 4, 8]. Population, gender, age distribution, vehicle ownership, and family composition were collected from the platform of the Australian Bureau of Statistics [18]. These demographic-related data for the year 2021 were available for Statistical Area 2 zones (SA2 is an aggregation of SAs). We converted these values to SA level by dividing the values by the number of SAs in each SA2 zone.

To represent the land use distribution, we collected POI data from the CLUE (Census of Land Use and Employment) project of the City of Melbourne Open Data Portal [19]. It includes information about land use from 2002 and is updated annually. We selected records in census year 2020, as it was the most updated dataset. Each record contained the coordinates of POIs. Locations of the tram stops in the study area were acquired from the Open Street Map [20]. We calculated two design indices used in the previous studies: Shanon’s entropy, which quantifies the diversity of a region, and the Mixed used Index ($MXI$), which captures the variance between residential and other POIs [3, 4]. The Shanon’s entropy $E$ and the $MXI$ of a region are calculated as equations (2) and (3), respectively.

$Ni(r)$ is the POI count of type $i$ in region $r$ and $N(r)$ is the total POI count in region $r$.

$P(r)$ is the residential percentage in region $r$.

To explore how temporal and weather factors influence trip frequency in Melbourne, we collected hourly weather data from the Australian Government Bureau of Meteorology platform [21]. The day of the week and the time of the day were also modeled as variables. We selected these variables according to the literature.

Figure 1 illustrates the number of e-scooter trips started hourly on different days of the week. There is a clear variation between the number of trips that occurred on weekdays and weekends. Two noticeable peaks are visible on weekday trips, one around 8 am and the other around 5 pm. The number of trips started in the afternoon is higher than the number of trips started in the morning irrespective of the day. A similar observation of higher afternoon trips was common in a few other cities such as Minneapolis, Washington D.C. and Berlin [8, 6, 11].

Table I and Table II summarise all the variables used in the spatial and temporal analysis respectively.

This section discusses the method employed to analyse associations between spatial features and e-scooter usage. We explored the relationship between the independent and dependent variables shown in Table I, using regression models. Then the analysis was extended on different time frames. Through this, we expected to investigate the temporal dynamics of the spatial features associated with e-scooter trips.

This section describes the method of analysing relationships between temporal determinants and e-scooter usage. First, hourly e-scooter trip count in the entire study area was estimated, and the most affected temporal features were derived. Then we extended the analysis on different SAs (geographical unit of analysis in this work), to explore spatial dynamics of the temporal features associated with e-scooter trips.

Regression analysis helps us recognise the variables that have an impact on the number of e-scooter trips in terms of space and time. However, it does not provide insights into the association between e-scooter usage and POIs. To better understand the relationship between e-scooter parking and locations of POIs, we conducted a buffer analysis taking different time frames into account. The purpose of the buffer analysis is two-fold;

One is to explore whether there are POIs around trip start/end locations during different time frames. Figure 4 shows the placement of 60m buffer around trip-start locations and the availability of POIs. In [8] a 60m buffer was used for the same purpose of capturing adjacent buildings to e-scooter parked locations. We used POI location data collected from CLUE project from the City of Melbourne. The percentage of trip start/end locations having a given type of POI within the buffer area was calculated using equation (4).

Where $T(x,t)$ is the number of trip start/end locations having at least one item of POI type $x$ inside the buffer area at time $t$ and $T(t)$ is the total number of trip start/end locations at time $t$. The value of $T(t)$ and the summary of trip start/end counts during the study period is given in Table III.

Our second objective of buffer analysis is to analyse the effect of e-scooter parking on different path types. For this, we collected footpath, shared path, and cycle lane shapefile datasets from the City of Melbourne open data portal. We created a 10m buffer around each path type and calculated the percentage of trips started/ended within the buffer area at different time frames using equation (4). Here, $T(x,t)$ is the number of trip start/end locations within the buffer area of path $x$ at time $t$ and $T(t)$ is the total number of trip start/end locations at time $t$. Since we observed different patterns of e-scooter parking around each path type, we extended the analysis by breaking the 10m buffer into three smaller buffers as shown in Figure 5; on the path, within 5m buffer, and 5m-10m buffer. We used the QGIS tool for buffer creation.

Each type of path is dedicated for a certain purpose. For example, a footpath is designated for pedestrians, a cycle lane is for cyclists and a shared path can be used by both. The aim of the extended buffer analysis around paths, is to explore how each of these path types were affected by e-scooter parking.

To select the best spatial regression model, we assessed the performance of the implemented models in terms of trip density prediction. Table IV depicts the regression model results with respect to the standard error metrics; Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE) and Mean Absolute Percentage Error (MAPE). We compared the standard errors for the SVR, NN, RFR models and (non-standardized) NBR, (non-standardized) RFR separately. Results show that the RFR has the least MAE, MSE and RMSE among all four machine learning models. Therefore, we selected RFR to derive the most influential features among independent variables and for the further spatial analysis of building regression models for different time frames.

In a regression model, we can understand the importance of features based on how they contribute to the model output prediction. We used the Scikit-learn built in feature_importances_ method to calculate an importance score for each independent variable. Even though this score quantifies how important each feature is for the prediction, the direction of correlation with the output is not visible. Therefore we calculated the Pearson correlation between e-scooter trip density and each feature. The feature importance scores of the spatial variables according to the RFR model is illustrated in Figure 6. Features represented in blue colour bars are positively correlated with e-scooter trip density and the red ones are negatively correlated.

As per the spatial regression model for total trips, population density in a SA is the most important feature in predicting e-scooter trip density. SAs with higher population tend to have higher e-scooter usage. Percentage of population owning motor vehicles is the second important feature according to the model. It is negatively correlated with the e-scooter trip density. The percentage of families without children is a new feature that we added in this study. Being the third important feature, family composition is an influencing factor for e-scooter trip prediction. According to our analysis, the most important age group for predicting e-scooter trip density is 30-39 years old.

To understand the temporal dynamics of spatial features associated with e-scooter trips, we compared feature importance scores of the models implemented for different time frames. Figure 7 shows the feature importance scores of the top most variables of models developed for weekday and weekend different time frames. The importance of car ownership and population of young age groups increases in weekend model compared to weekday model. Population density is the most important feature in the weekday morning models and car ownership % becomes more important in midday, evening and at night weekday models. The importance of office% is more evident in the weekday 6-10 model. Tram stop density is an important feature in all the weekend models except 19-23. Cafe% is important for e-scooter trip density prediction in 0-5 weekday model and three weekend models.

The performance of temporal regression models implemented were evaluated using the metrics in section V-A. Based on the model results shown in Table V, we selected the RFR model to explore the effectiveness of explanatory variables in trip count prediction. The same model was chosen for further temporal analysis of building regression models on different spatial zones. Figure 8 depicts the feature importance scores of the temporal variables in the RFR model. Positively correlated continuous variables, negatively correlated continuous variables and discrete variables are represented in blue, red, and gray respectively. According to the temporal regression model for total trips, humidity is the most important feature that affects the output prediction. High humidity levels are associated with a lower number of e-scooter trips. While precipitation is the least important weather feature as per the model, though it also reduces the trip count. Wind speed is a feature that positive correlates with e-scooter trips. In our model, the 19-23 time frame is the most important of all time frames and among days of the week, weekend is more important than weekdays for e-scooter trip count prediction.

To understand the disparity of temporal determinants in different spatial zones, we calculated feature importance scores of models implemented on three SAs. Figure 9 shows the results. In the first model (SA1: with most offices), weather features show more importance than time and day features. In the second model (SA2: with most cafes), time 19-23, 6-10 are the most important for output prediction. The importance of weather features are reduced in the second model. In the third model (SA3: with most recreation), weather features hold more importance than day and time. Under the assumption that SA1, SA2, and SA3 consist of most work-related, leisure-related and recreational trips, this results demonstrate the impact of weather on shared micro-mobility usage vary based on the trip intention. In model 2 and 3, the important scores for Saturday and Sunday are higher than weekdays, while in model 1, some weekdays have higher important scores than weekend. It indicates that weekends are more important in the areas with leisure-related trips. Overall, the importance of temporal variables for shared e-scooter trip count prediction is diverse across geographical areas.

In this section, results of the buffer analysis is presented. We selected residences and offices as the POIs of concern, since those were the most important according to the total trip regression model. Figure 10 shows percentage of trips with offices and residences within the buffer area at different time frames on weekdays. Comparing the percentage of trip-starts having offices within the buffer area against relevant trip-stops shows, 6-10 is the only time frame having more stops (mean = 41.3%) than starts (mean = 35.7%). The difference between percentage of trip starts and ends having offices in the buffer area were statistically significant, confirmed by paired t-test at 0.05 significance level; t=-11.78, p$<$0.05. In contrast, in the 6-10 time frame, more trip-starts (mean = 64.2%) happened near residences than trip-ends (mean = 57.3%). A paired t-test verified the difference between percentage of trip starts and ends with residences in the buffer area to be statistically significant at 0.05 significance level; t=10.45, p$<$0.05. One possible reason for these observations can be more riders start e-scooter trips around residences and reach offices in 6-10.

Next, we discuss the results of the buffer analysis around paths. We first looked at the percentage of trips started/ended within a 10m buffer area from each type of path. To understand e-scooter parking patterns near different path types, we compared the percentage of trips originated and ended near a selected path type on different time frames. One important thing to note here is, we did not compare the percentage of trips across path types, because the land coverage of footpaths is higher than the land coverage of cycle lanes and shared paths. The percentages of e-scooters trips originated and ended near footpaths, bike lanes and shared paths at different time frames in weekdays and weekends are shown in Figure 11. In all time frames, the percentage of e-scooters stopped near footpaths (mean = 74.4%) are higher than the trips started (mean = 71.4%) around a footpath. A paired t-test confirmed that difference between percentage of stopped and started e-scooter trips around footpaths were statistically significant at 0.05 significance level; t=-5.26, p$<$0.05. When comparing e-scooters parked near bike lanes, higher percentages of trip starts and stops can be seen at daytime (6-18_mean = 15.3%, 19-5_mean = 13%). Most of the shared paths in Melbourne are situated around recreation locations such as parks. It can be the reason for having more trip-starts and trip-ends near shared paths on weekends compared to weekdays in every time frame (weekday_mean = 3.8%, weekend_mean = 4.6%).

We extended the buffer analysis around paths by breaking the 10m buffer into smaller areas. Through this, we wanted to explore whether e-scooters are parked ’on the path’ or around the path. We calculated the percentage of trips started and ended within three buffer areas in five time frames (similar time frames used in the spatial regression). Table VI shows the mean percentage values relevant to trip-starts and trip-stops of weekdays and weekends. According to the results, parking patterns around footpaths is different to cycle lanes and shared paths. More trips have started and ended on the footpath (28.2%) compared to the area around footpaths (5m = 25.3%, 5m-10m = 19.1%). The opposite is observed for cycle lanes and shared paths. It indicates the higher utilization of footpaths for parking e-scooters.