Keywords

1 Introduction

Bikeshare promotes sustainable travel, health benefits, and economic growth (Qiu and Chang 2021). Dockless bikeshare (DBS), compared to the docked bikeshare system, is getting more popular in the last decade due to benefits like accessibility and convenience (Gu et al. 2019).

There are observed research gaps in DBS seasonal study: (1) the development of mobile applications and cashless mobile payment have make bike sharing usage even more prevalent (Guo et al. 2022). However, as a new mode of transportation, DBS has received less attention than docked bike sharing (Guo et al. 2022). (2) A majority of precedent studies of DBS focus mostly on where a trip starts and ends, rather than the cycling experience itself. (3) limited examination of how seasonal streetscape affects cycling. Although a small number of studies consider temporal scale, yearly comparisons (Li 2021) offer limited help in comparing seasons, and studies addressing the association between seasonal climate and bike sharing didn't examine other seasonal environmental characteristics(Li and Kamargianni 2017).

The study (1) provides a quantitative study of DBS focusing on perceived environmental elements along the journey, (2) integrates SVI and CV to estimate how seasonality of street built environment impacts DBS usage at a fine spatial scale. (3) considers previously ignored seasonal environmental features like vegetation color.

2 Data and Methodology

2.1 Study Area and Methodology

Our study area includes the Town of Ithaca and a few nearby neighborhoods (Fig. 1).

Fig. 1.
figure 1

Study area

Full size image

Figure 2 illustrates the framework of this study: (1) Using GPS bike trajectory data from LIME, a DBS system in Ithaca USA, we computed Seasonal Weighted Rides (SWR) to capture the cycling volumes of road segments in summer and autumn. (2) We collected SVIs in summer and autumn with Google SVI API. (3) We used PSPNet to compute the view ratio indices of streetscape elements, and used Mask R-CNN to count the number of streetscape objects. We also computed CV indicators (color deviation, L, A, B values in CIELAB color space) to present the seasonal color change of these three variables: tree, plant, and grass. (4) We quantified four subjective perception scores (accessibility, ecology, enclosure, scale) of street environment in summer and autumn with ML. (5) We collected and computed non-seasonal variables (typical POI, landmark POI, infrastructure, road type, land use mixed score, street network connectivity, terrain). (6) With OLS regression models, the seasonal environment attributes are comprehensively analyzed with their impacts on DBS volumes in summer and autumn.

Fig. 2.
figure 2

Research framework

Full size image

2.2 Data and Processing

2.2.1 Dependent Variable: DBS Data and Seasonal Weighted Rides (SWR)

LIME provided us with the DBS trips data for this study, accessible through an API. LIME’s app was used to collect the data. After the users agreed to LIME’s terms and conditions for using the services, LIME recorded and analyzed their journeys. The dataset does not include identifiable individual information. To clean the data, trips that started or ended outside the Greater Ithaca area (Qiu and Chang 2021), trips with distances shorter than 0.05 miles or 264 feet (Qiu and Chang 2021), and trips with durations less than 3 min or more than 120 min were removed. The validated dataset has 102,178 trip records.

To make cyclists’ choice and preference of routes comparable across different seasons, we have to capture the popularity of road segments in each season with Seasonal Weighted Rides (SWR). This conversion can be found in Eq. (1).

$$SW{R}_{j} =S{R}_{j} / ( \sum_{j=1}^{n}{r}_{j} )$$
(1)

For segment \(j\), \(S{R}_{j}\) represents the volume of rides on it in this particular season, \(n\) represents the number of segments with rides on them in this particular season, and \(\sum_{j=1}^{n}{r}_{j}\) represents the total number of rides on all segments in Ithaca during this particular season. The \(SW{R}_{j}\) that is aggregated to the sampling points \(i\) is the dependent variable of this study, and has the precision of street segments, so sampling points on the same segment would have the same value for one season.

To reduce the bias caused by data sparsity, we removed segments with a total annual volume of less than 500 riders (approximately 1.5 riders a day) from the aggregated data. Then we sampled points every 25 m in segments with lengths more than 50 m (which is too short). Therefore, 671 out of 10508 road segments were selected, and 2,082 sampling points were obtained from them.

2.2.2 Independent Variable: Street View Measuring

Google Street View (GSV) is only available in summer (Jun, July, Aug) and fall (Sept, Oct, Nov) in our study area, and out of 2,082 sampling points, there are only 1,170 points having GSV in both seasons. With GSV panorama ID and Street View Download 360 software (Street View Download 360, n.d.) we download the panorama in both summer and fall of each point. To extract the count of elements in a panorama SVI, we use a Mask-RCNN pre-trained on COCO dataset with ResNet-50 backbone. Then we use Pyramid Scene Parsing Network (PspNet) with pre-trained model psp_resnet101_ade to conduct image segmentation.

Undesired segmentation distortions might occur near the top and button of a panorama image, so we unwarp the panorama into images in 6 directions (Forward, Back, Left, Right, Up, and Down) with py360convert package and extracted the four directions at eye-level: forward(F), back(B), left(L), and right(R) (Fig. 3).

Fig. 3.
figure 3

Panorama object detection with Mask_RCNN, unwarping(in 4 directions: F, R, B, L), and semantic segmentation with PSPNet. Left: Summer SVI. Right: Autumn SVI

Full size image

From percentage that the pixels of the specific visual element take-up of the total pixels of an SVI we calculated the visual ratio of an element in the image. Not all objective view indices will be input to the regression model after the Variance Inflation Factor (VIF) test, only 20 out of 28 visual ratio are kept, including: tree, road, grass, car, streetlight, wall, building, sidewalk, earth, water, plant, awning, van, person, bridge, railing, bicycle, minibike, ceiling, chair.

To study color and change of street greenness from urban cyclists’ perspective, we used the CIELAB colorspace and extracted PSPNet pixels for three types of greenness. Converted from RGB to CIELAB using the Python-colormath library, we calculated average A and B values for each pixel and the standard deviation from actual values. To eliminate potential interference, we eliminated the L value (brightness) as SVIs are taken in different conditions.

We evaluate street perceptions using a 300-point pre-labeled dataset and an ML framework developed by related research (Qiu et al. 2023; Su et al. 2023) to evaluate the subjective perceptions score of the streetscape: Accessibility (accessibility to activities, attractions, and amenities), Ecology (detecting living organisms, animals, plants, humans, and their physical environment), Enclosure (the degree to which buildings, walls, trees, and other vertical elements define streets and public spaces visually), and Scale (human-sized and proportional elements). The dataset was collected from a crowdsourced visual survey of an expert panel and includes SVI input variables and 4 perceptual scores as output labels. 75% (225) of the dataset is used for training and 25% (75) for testing. Multiple ML algorithms are used, including K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Random Forest (RF), Gaussian Process (GP), Gradient Boosting Regression (GB), ADA boost, and Bagging Regression. GP is chosen as the optimal model for predicting the target perceptions (Table 1).

Table 1. Performance of GaussianProcessRegressor (GP) predictions
Full size table

2.2.3 Independent Variable: Non-seasonal Variables

We use Open Street Map (OSM) to get typical POI, infrastructure (transit facility), road types, and landmarks POI. The buffer zone radius is 500m for typical POIs, 100 m for infrastructure, and 1000 m (typical 5 min bike rides) and 3000 m (typical 15 min bike rides) for landmark POIs.

We use land use data collected from Tompkins County Open Data Portal (Land Use and Land Cover 2015; Tompkins County Open Data Portal, n.d.), and use a method originally developed to calculate the evenness of distribution of the area of different use types (Frank et al. 2005). The calculation is shown in Eq. (2).

$$LanduseMix =\left(-1\right)*\left(\sum_{i=1}^{n}{p}_{i}\mathrm{ln}\left({p}_{i}\right)\right) / \mathrm{ln}\left(n\right)$$
(2)

With \({p}_{i}\), the proportion of land use \(i\) within the 500m buffer is divided by the total area within the 500 m buffer, while \(n\) represents the number of exclusive land use types within the 500 m buffer. This formula would calculate the mixed land use level within range 0–1, a higher value means a higher level of mixed land use.

We use Depthmap X to run the space syntax calculation. However, we finally choose Connectivity and Angular Integration (with segment length weighted, or SLW) after removing space syntax score with high multicollinearity (VIF > 10). Angular Integration (SLW) has two metric radius parameters: 250 and 1000 m. For a particular segment, Connectivity describes quantity of other segments it connects with, and Angular Integration (SLW) captures accessibility and how close(integration level) of this segment is to all others in terms of the sum of angular change(Angular Integration Space Syntax—Online Training Platform, n.d.).

We use high resolution(1 m) Digital Elevation Model (DEM) from United States Geological Survey (USGS) collected in May 2020. The medium value of slope within 5 m of sampling point along the road is chosen as the slope value.

2.3 Model Architecture

We start with a simple OLS model Eq. (3).

$${Y}_{i}= \alpha +\sum_{m}{X}_{(i,m)}{\beta }_{m}+ {\epsilon }_{i}$$
(3)

\({Y}_{i}\) is the dependent variables of the sampling point. \(X\) is the independent variables that explains \(SWR\). \(\beta \) is the coefficient of variable \(m\) that reveals how and to what extent variable \(m\) is related to \(SWR\). Constant term \(\alpha \) refers to the average \(SWR\) when all other variables are zero. Error term \({\epsilon }_{i}\) captures elements that influence the \(Y\) but are not included in \(X\).

A baseline model (M1_Summer, M1_Autumn) with significant variables was constructed using all variables except subjective perception. VIF was calculated in the whole process and only variables with VIF less than 10 are kept. Then all 4 perception scores are added (M2_Summer and M2_Autumn) (Table 2).

Table 2. Dependent variable, independent variable, and method of different models
Full size table

3 Result and Discussion

3.1 Models Performance and Diagnosis Results

Because DBS volume is not statistically significant with all variables, we remove unrelated variables from models. The remaining significant variables are used to build models after removing variables with high correlations and multicollinearity with a VIF test. The Regression diagnosis results show that: (1) All Autumn models have higher R2 than Summer models. (2) M1 model and M2 model have very close R2 with or without subjective perception. (3) There is high correlation and multicollinearity between the new perception scores and other segmentation results.

M1_summer and M1_autumn results show consistent significance for some variables in both seasons, such as visual ratios for roads, cars, sidewalks, grass color deviation, land use mix scores, and road network angular integration. Some variables are only significant in summer, like building and wall ratios and number of education POIs, while others like color deviation and lab_A values of tree, lab_A and lab_B values of grass only show significance in autumn. For variables significant in both seasons, the positive or negative effects are consistent across seasons (Table 3).

Table 3. Coefficients between bikeshare volume and selected variables
Full size table

3.2 The Seasonal Variations of the Streetscape

Cycling volume is positively impacted by the presence of roads and sidewalks, with dedicated bike lanes potentially available on a larger network. Studies of the built environment often examine both walking and cycling behaviors together (Mertens et al. 2016), so sidewalks play an important role as well. A pedestrian-friendly neighborhood promotes sustainable mobility and slows down traffic. Trees positively impact cycling in both seasons. Street greenery offers ecological benefits to neighborhoods, including providing shade for microclimate control, and creating an enjoyable environment for cyclists (Li et al. 2018). Waterfront areas provide a desirable setting for cycling, which aligns with previous research findings (Ding 2016; Lee et al. 2021; Song et al. 2021).

3.3 Other Non-seasonal Variables

Land use mix score is positively related to DBS volume in both season, aligning with prior evidence that the mix ratio of land use affects travel behavior (Van Dyck et al. 2012; Kerr et al. 2016). Angular integration at 1000 m radius positively contributes to DBS volume in both seasons, in line with previous research suggesting that road network accessibility influences cycling behavior (Saghapour et al. 2017; Tucker and Manaugh 2018). However, angular integration at 250 m radius and connectivity are only significant in summer, which may be because autumn rides are more commuter-related. More research is needed to understand how space syntax impacts DBS at various scales and times. The number of points of interest (POI) is found to have a positive impact on DBS volume in both seasons. However, only educational POI has a significant positive impact in summer. This difference would need further research with more POI data.

4 Conclusion

Using GPS trajectory data, this study examines the correlation between DBS, streetscape, and spatial elements in different seasons. The study finds that: (1) seasonal streetscape factors such as roads, cars, sidewalks, tree, and vegetation color significantly influence DBS volume; (2) the significance varies in summer and autumn; (3) non-seasonal factors like mixed land use score, street network connectivity, etc., are significant in both seasons, some only show significance in one season; (4) adding subjective perception to both seasons improves explanatory slightly.

There are several limitations in this study that can be improved in future studies. Firstly, due to the data source limitation, only summer and autumn SVIs are collected in Ithaca. Finer temporal resolution can also be taken into consideration when SVI from more seasons or even months is available. Secondly, more advanced spatial model can be introduced to examine the spatial effect. Thirdly, microclimate-related data like temperature would better explain the seasonal variation.