TL;DR
This paper introduces a novel data-driven method to classify bus accessibility over time, using floating car data to create dynamic "actual isochrones." By applying dynamic time warping and k-medoids clustering to these time series, the research reveals distinct spatial patterns of bus service accessibility, offering valuable insights for urban planning and public transit management.
What Problem Does This Paper Address?
Traditional accessibility studies often focus on static measures or individual-centric analyses. While temporal dynamics are increasingly recognized as crucial, place-centered accessibility research has lagged in adopting data-driven, time-series classification methods. Existing approaches often rely on subjective metrics or pre-selected snapshots, potentially missing nuanced spatiotemporal patterns. This paper aims to fill this gap by developing a method to classify geographical areas based on the temporal behavior of their bus accessibility, using readily available floating car data.
Methods
The core of the methodology involves several key steps:
- Data Preparation and Bus Network Construction: Floating car data (FCD) from buses is cleaned, snapped to a road network, and used to construct a dynamic bus network. This network incorporates time-varying bus speeds and travel times between stations, accounting for walking times to and from stops.
- Actual Isochrone Time Series Calculation: For each hexagonal geographical unit (chosen for their spatial contiguity), actual isochrones are calculated at various time intervals (e.g., 30, 60, 90 minutes) based on the dynamic bus network. The area of these isochrones serves as a measure of accessibility for that unit at a specific time. This process is repeated for numerous time windows throughout the day, creating an accessibility time series for each unit.
- Time Series Classification:
- Dynamic Time Warping (DTW): DTW is used to calculate the "distance" or dissimilarity between the accessibility time series of different hexagonal units. DTW is particularly effective for time series that may vary in speed or have slight temporal shifts.
- k-Medoids Clustering: Based on the DTW distances, the k-medoids algorithm is employed to group hexagonal units into a predefined number of classes (K). K-medoids is chosen for its robustness to outliers, making it suitable for non-Euclidean distances like DTW.
- Evaluation: The optimal number of clusters (K) is determined using a combination of the elbow method (minimizing within-cluster sum of distances) and the silhouette score (measuring how similar an object is to its own cluster compared to other clusters). Manual examination of time series plots and spatial distribution maps further validates the classification.
Datasets
The study utilizes:
- Road Network Data: Obtained from OpenStreetMap.
- Bus Floating Car Data (FCD): Collected from Anhui Mobile Communication Co., Ltd. for a week-long period in Hefei, China. This data includes GPS location, timestamps, bus line information, and service conditions.
- Bus Station Data: Provided by the bus company, including precise locations and connected bus lines.
No new benchmark datasets were introduced; the focus was on applying existing data types to a novel analytical framework.
Key Insights and Ideas
- Dynamic Isochrones as Accessibility Metrics: Using the area of actual isochrones derived from real-time bus FCD provides a more nuanced and data-driven measure of place-centered accessibility than static methods.
- DTW for Temporal Pattern Matching: Dynamic Time Warping proves to be a powerful tool for comparing the temporal shapes of accessibility time series, even when they don't align perfectly in time. This allows for the identification of similar accessibility dynamics across different locations.
- Spatial Clustering of Temporal Accessibility: The classification reveals that distinct temporal accessibility patterns are not randomly distributed but form clear spatial clusters. This suggests that factors influencing bus service accessibility (e.g., network density, demand, traffic conditions) are spatially correlated.
- Data-Driven Insights for Transit Planning: The method offers a way to automatically identify areas with similar or contrasting accessibility behaviors, which can directly inform targeted interventions for public transit planning and service improvement.
Key Results
- The proposed method successfully classified hexagonal units into distinct accessibility time series classes.
- The case study in Hefei, China, demonstrated the feasibility of the method, yielding a "fair" silhouette score of approximately 0.3.
- The classification revealed clear spatial patterns:
- 3-Class Classification: Showed a concentric pattern with Class 1 in the city center, Class 2 encircling it, and Class 3 in the outskirts.
- 4-Class Classification: Also showed spatial patterns, with notable differences in the southwest part of the city between weekdays and weekends, where classes 2 and 3 became interwoven.
- Class 1 (city center) generally exhibited higher average accessibility and distinct dynamic patterns (e.g., a noon peak) compared to other classes.
- Weekday and weekend accessibility patterns differed significantly, with weekends showing less circular patterns for classes 2 and 3.
Limitations and Future Work
- Spatial and Temporal Granularity: The choice of 250m hexagonal units and 20-minute time windows was based on data characteristics and computational constraints. Further investigation is needed to understand the sensitivity of the results to these choices.
- K Selection Subjectivity: While multiple methods were used, selecting the optimal number of clusters (K) still involves some degree of subjectivity and manual inspection. Future work could explore fully automatic K optimization methods.
- Bus Service Only: The study focused solely on bus accessibility. Integrating other transit modes like subways, which have different operational dynamics and are less affected by road traffic, would provide a more comprehensive picture.
- Statistical Correlation: While spatial patterns were observed, the study did not statistically investigate their correlation with geographical features (demographics, road networks). This is suggested as a fruitful avenue for future research.
- Individual-Centered Accessibility: Incorporating individual-centered accessibility into place-centered classification could offer richer insights.
- Multimodal Integration: Exploring how to integrate subway schedule data with bus FCD for place-centered accessibility analysis is a promising direction.
Summary updated