TL;DR
This paper investigates the spatial and temporal drivers of shared e-scooter usage in Melbourne, Australia, using a mix of machine learning regression and GIS buffer analysis. The study reveals that nearly 30% of e-scooter trips end directly on footpaths and identifies population density and car ownership as the primary predictors of where these "micro-mobility" trips occur.
What Problem Does This Paper Address?
As cities embrace e-scooters to reduce carbon footprints and traffic congestion, they face a growing "clutter" problem. E-scooters are often left haphazardly, obstructing pedestrians and creating safety hazards.
The authors argue that existing research often takes a "one-size-fits-all" approach, looking at either spatial factors (where people go) or temporal factors (when they go) in isolation, or relying on aggregated data that masks local nuances. To create better urban policies—like where to build parking hubs or how to redistribute fleets—planners need a granular, city-specific understanding of how e-scooters interact with existing infrastructure and weather patterns.
Methods
The researchers employed a multi-pronged analytical approach:
- Spatial Regression: They compared four machine learning models—Negative Binomial Regression (NBR), Support Vector Regression (SVR), Neural Networks (NN), and Random Forest Regression (RFR)—to predict trip density across 350 Statistical Areas (SAs). RFR emerged as the most accurate.
- Temporal Analysis: They modeled hourly trip counts against weather data and time-of-day variables to understand what keeps riders off the streets.
- GIS Buffer Analysis: This is the "footpath" part of the study. They created 60m buffers around Points of Interest (POIs) like offices and residences, and 10m buffers around different path types (footpaths, cycle lanes, shared paths). This allowed them to calculate exactly what percentage of trips started or ended on specific types of infrastructure.
- Multi-Resolution Modeling: Instead of one global model, they built separate models for different time windows (e.g., weekday mornings vs. weekend nights) to capture shifting usage patterns.
Datasets
The study utilized a rich variety of real-world data from Melbourne (August to October 2022):
- E-scooter Trips: Anonymized data from the Lime service via the General Bikeshare Feed Specification (GBFS) API, totaling ~175,000 starts and ~180,000 stops.
- Demographics: 2021 Australian Bureau of Statistics (ABS) data covering population, age, gender, and car ownership.
- Land Use: The City of Melbourne’s CLUE (Census of Land Use and Employment) dataset for POI locations.
- Infrastructure: Shapefiles for footpaths, cycle lanes, and shared paths from the City of Melbourne Open Data portal.
- Weather: Hourly metrics (humidity, rain, wind, temperature) from the Australian Bureau of Meteorology.
Key Insights and Ideas
- Humidity Trumps Rain: Surprisingly, the model found that humidity is a more significant predictor of lower trip counts than actual precipitation. While riders might brave a light drizzle, high humidity (likely associated with discomfort or extreme heat/cold) is a major deterrent.
- The "Car-Free" Demographic: There is a strong negative correlation between car ownership and e-scooter usage. This suggests e-scooters aren't just a "last-mile" solution for commuters; they are a primary mode of transport for those who choose not to own vehicles.
- Age Matters (Specifically 30-39): While US-based studies often point to the 18-24 age group as the primary users, this Melbourne study found the 30-39 age group to be the most significant demographic predictor.
- Footpath Encroachment is Real: The buffer analysis confirmed the paper's title: footpaths are the primary "parking" infrastructure. Usage of footpaths for parking significantly exceeds that of cycle lanes or shared paths.
Key Results
- Model Performance: The Random Forest model outperformed others, achieving the lowest Mean Absolute Error (MAE) and Mean Squared Error (MSE) for spatial predictions.
- Parking Stats: Approximately 29% of trips ended on a footpath, and 27% started there.
- Commuter Flow: Between 6:00 AM and 10:00 AM, there was a statistically significant trend of more trips ending near offices than starting there, confirming the e-scooter's role in the morning commute.
- Top Predictors: Population density was the #1 feature associated with high usage, followed by (low) car ownership and the percentage of families without children.
Limitations and Future Work
The authors acknowledge several constraints:
- Data Granularity: Due to privacy-preserving ID rotation (every 15 minutes), the researchers could not link specific start points to specific end points to track individual trip trajectories.
- Single Provider/City: The study focused on one provider (Lime) in one city (Melbourne) over a three-month period.
- Future Directions: The authors suggest expanding the study to include multiple providers and longer durations to account for seasonal variations. They also propose using these models to evaluate the effectiveness of "designated parking hubs" in reducing footpath encroachment.
Summary updated