Source-linked AI summary
Using coarse GPS data to quantify city-scale transportation system resilience to extreme events
Brian Donovan, Daniel B. Work
TL;DR
Transportation resilience requires quantitative monitoring, but conventional city-scale traffic sensing is costly and taxi GPS data are comparatively coarse and noisy. The paper computes historical origin-destination pace distributions and uses pace deviations to detect and measure unusual events. Applied to New York City taxi trips, the method finds prolonged and severe effects from Hurricane Sandy, with minor evacuation disruption but substantial post-disaster reentry delays.
Problem
City-scale transportation resilience needs quantitative monitoring, while traditional traffic sensors are costly and taxi GPS data contain noise and incomplete trip coverage.
Method
The method computes historical pace distributions between city regions and uses time-dependent Mahalanobis distance to detect statistically unusual pace deviations.
Results
Hurricane Sandy was the longest event in the four-year dataset, caused over two minutes of delay per mile at its worst, and produced its largest delay spike during post-storm reentry.
Takeaways & Limitations
Taxi GPS data can support low-cost city-scale resilience assessment, distinguishing the relatively efficient pre-Sandy evacuation from post-Sandy reentry gridlock.
Takeaways & Limitations
About 7.5% of trips are discarded as obvious errors, and event severity is difficult to evaluate without a known ground-truth training set.
Abstract
from arXiv · showhide
This article proposes a method to quantitatively measure the resilience of transportation systems using GPS data from taxis. The granularity of the GPS data necessary for this analysis is relatively coarse; it only requires coordinates for the beginning and end of trips, the metered distance, and the total travel time. The method works by computing the historical distribution of pace (normalized travel times) between various regions of a city and measuring the pace deviations during an unusual event. This method is applied to a dataset of nearly 700 million taxi trips in New York City, which is used to analyze the transportation infrastructure resilience to Hurricane Sandy. The analysis indicates that Hurricane Sandy impacted traffic conditions for more than five days, and caused a peak delay of two minutes per mile. Practically, it identifies that the evacuation caused only minor disruptions, but significant delays were encountered during the post-disaster reentry process. Since the implementation of this method is very efficient, it could potentially be used as an online monitoring tool, representing a first step toward quantifying city scale resilience with coarse GPS data.
1 Introduction
The paper develops a city-scale transportation-resilience method using taxi GPS data, addressing the need for quantitative monitoring without costly traditional sensors. It measures statistically significant disruptions through pace deviations, recovery time, and peak degradation, and applies the approach to compare extreme events.
- 1.1 Motivation: Taxi GPS fleets provide a lower-cost basis for monitoring transportation infrastructure resilience than installing traditional traffic sensors.The method uses existing taxi sensors to study city-scale impacts relevant to evacuations and post-disaster emergency services.
- 1.1 Motivation: The method measures resilience at city scale using recovery time and peak pace deviations rather than individual streets or vehicles.Pace is travel time per mile, normalizing for varied taxi-trip lengths.
- 1.1 Motivation: Statistically significant disruptions help separate event-related changes from noisy, variable traffic conditions and unmodeled human factors.The design explicitly targets signal-versus-noise separation in taxi data.
- 1.1 Motivation: The computation is intended to be O(N), ideally requiring one pass through taxi-trip data and supporting realtime streams.This requirement follows from the high frequency of taxi trips in large cities.
- 1.3 Outline and Contributions: The method uses historical pace distributions to detect unusual events and quantify peak disruption and recovery time.Its city-scale application compares events including Hurricane Sandy, highlighting efficient evacuation before Sandy versus post-Sandy reentry gridlock.
2 Methodology
The method aggregates taxi trips by origin, destination, and hour, then uses distance-normalized pace distributions to identify unusual city-scale traffic conditions. It detects deviations from time-dependent typical behavior and summarizes events through disruption and recovery statistics.
- Trip aggregation: Taxi trips are aggregated by origin-destination pairs and hourly departure time to extract city-scale traffic features.The city is divided into k regions, producing k^2 origin-destination pairs.
- Pace measurement: Distance-weighted average pace normalizes travel time by metered trip length and gives longer trips more influence.Pace is defined as travel time per mile.
- Typical behavior: Historical pace vectors are modeled with a weekly periodic reference distribution whose mean and covariance capture expected conditions, variability, and correlations.The reference set uses matching hours from other weeks, while the full covariance matrix accounts for correlated origin-destination paces.
- Event detection: A time-dependent Mahalanobis distance scores deviations from typical pace behavior, with events detected when the score exceeds a threshold.The threshold is set to the 95% quantile of M(t), and noisy threshold crossings are merged when separated by fewer than six hours.
- Event characterization: Detected events are characterized by recovery time and maximum pace deviation, enabling comparisons of how different events affect the city.Long-lasting events can be examined across different stages.
3 Application to Hurricane Sandy with New York City Taxi Data
The New York City application uses nearly 700 million taxi trips to construct city-scale pace features, process them efficiently, and detect Hurricane Sandy’s transportation impacts. Sandy produced unusually severe and prolonged travel-time disruptions, although some Lower Manhattan routes became faster as demand declined.
- Dataset and processing: 7.5% of trips were discarded for invalid coordinates, distances, times, velocities, or other obvious errors.The data records only occupied taxi trips, excluding non-occupied travel.
- Dataset and processing: 697,622,444 taxi trips were aggregated into 35,064 hourly mean pace vectors covering the four-year dataset.Because trips were chronologically sorted, the vectors were computed in a single pass.
- Computational implementation: O(N) processing extracted all 35,064 mean pace vectors in about 75 minutes on an 8-core machine, using roughly 40MB of RAM per process.Subsequent Mahalanobis-distance computation also ran in linear time and took less than 10 seconds.
- Pace features: Four large regions—Upper Manhattan, Midtown, Lower Manhattan, and East of the Hudson River—define 16 hourly origin-destination pace categories for city-scale analysis.The regional design captures geographic and infrastructural properties while avoiding street-level analysis.
- Hurricane Sandy findings: The slowest traffic occurred on Wednesday October 31st, almost two days after Sandy struck land, while some Lower Manhattan routes remained nearly five standard deviations faster than average.The authors suggest reduced demand in Lower Manhattan until infrastructure recovery as a likely explanation for the faster routes.
- Hurricane Sandy findings: Over five and a half days were required for travel times to return to normal after Hurricane Sandy, more than three times Hurricane Irene’s recovery time.Sandy added over two minutes per mile at its worst, whereas Irene added less than forty seconds.
- Caveat: Accuracy of the event comparisons is difficult to evaluate because the true severity of each event is unknown.With a training set, the detection threshold could be adjusted to balance type I and type II errors.
4 Conclusion
The method detects and measures unusual transportation events from taxi GPS data by comparing origin-destination pace deviations with typical traffic statistics. Applied to New York City, it finds Hurricane Sandy produced severe, delayed post-disaster reentry disruptions.
- Taxi GPS data can detect and measure unusual-event effects on transportation infrastructure without additional sensors.The method quantifies deviations in origin-destination pace from typical traffic conditions.
- Typical traffic statistics distinguish significant event-related disruptions from random day-to-day fluctuations.
- Over two minutes per mile of delay was the peak disruption attributed to Hurricane Sandy.The hurricane was also the longest event in the four-year dataset and one of the most severe by peak pace deviation.
- The largest delay spike occurred two days after landfall during residents’ return to New York City.The post-disaster reentry process was extremely slow compared with evacuation before the hurricane.
5 Future Work
Future work focuses on testing alternative city-region partitions and potentially defining regions automatically. Increasing the number of regions would require more GPS data for adequate origin-destination samples, while link-level analysis remains computationally expensive.
- Regions are currently chosen manually, so alternative partitioning schemes or clustering-based definitions could be evaluated.
- More regions require more GPS data to obtain adequate samples for each origin-destination pair.
- Link-level pace analysis is a possible extension, but existing link travel-time estimation algorithms are computationally expensive.