Abstract
Hydrological parameters require careful calibration before they can effectively support urban decision-making. However, traditional calibration methods often fail in regions where runoff data are inadequate—the "ungauged catchment problem." To address this, we propose a Bayesian updating framework that leverages taxi GPS data as a proxy for hydrological state. By analyzing periods of road disruption where taxi presence drops to zero, we can calibrate parameters like catchment area and time of concentration. Validated across 10 flood-prone roads in Shenzhen, China, the framework correctly predicted runoff trends for 80% of the sites. This study demonstrates that integrating transportation big data with hydrological modeling offers a viable alternative for flood hazard mitigation in data-poor urban environments.
Introduction
Predicting urban flash floods is critical for city resilience, yet the accuracy of these predictions depends heavily on model calibration. Standard calibration requires physical measurements of water flow (runoff data), which are often unavailable for dense urban road networks due to the high cost of installing and maintaining physical gauges. This lack of data creates a significant barrier to effective flood management.
This research explores the use of "human sensors"—specifically, the collective behavior of taxi drivers—to fill this data gap. By treating the absence of taxis on a road during a storm as evidence of flooding, we can infer the underlying hydrological conditions and tune our models accordingly.
Methodology
The core of the proposed approach is a Bayesian updating framework that translates meteorological data into the probability of observing no taxis on a specific road segment. This involves a three-stage transformation:
- Rainfall to Runoff: The Soil Conservation Service (SCS) unit hydrograph method is used to convert rainfall excess into a flow hydrograph. The primary parameters for calibration are the catchment area and the time of concentration.
- Runoff to Road Disruption: An empirical runoff-disruption function was developed based on vehicle stability literature. This function maps water flow (depth and velocity) to the probability that a road becomes impassable or is avoided by drivers.
- Disruption to No-Taxi Probability: Using a Poisson distribution based on historical taxi arrival rates, the model calculates the likelihood of zero taxi observations within a 5-minute window given the calculated disruption probability.
The Bayesian framework uses these likelihoods to update the probability distribution of the hydrological parameters. As more storm events are observed, the model's "belief" converges on the most physically accurate parameter sets.
Data and Validation
The study focused on Shenzhen, China, utilizing several high-resolution datasets:
- Taxi GPS Data: Anonymized trajectories from May 2015, aggregated into 5-minute intervals.
- Rainfall Data: 5-minute accumulative rainfall from 115 stations.
- Water Level Data: Ground truth measurements from 171 flooding gaging sites used to validate the model's predictions.
Key Results and Insights
The framework was tested on 10 flood-prone roads. The results showed that the model could correctly predict runoff trends for 8 out of the 10 roads. For these successful cases, the correlation between predicted runoff and observed water levels was high, with Pearson coefficients ranging from 0.68 to 0.94.
Furthermore, the Bayesian framework demonstrated significant refinement of parameter estimates after just two storm events, moving from a broad initial guess to a focused, high-confidence estimate of the road's hydrological characteristics. A key insight from this work is the value of absence as evidence. While a few taxis passing through a street might provide ambiguous information about water depth, a total cessation of traffic during a heavy storm is a strong indicator of physical blockage.
Discussion and Limitations
While promising, the method has limitations. It is highly dependent on taxi density; in areas with low traffic volume, the "no-taxi" signal is too noisy to be reliable. Additionally, the model currently does not account for cascading failures, where a road might be empty not because it is flooded, but because an adjacent connecting road is blocked.
Future work will focus on developing algorithms to distinguish between direct flooding and network-induced disruptions and exploring more complex hydrological models that incorporate urban drainage infrastructure.
Conclusion
This Bayesian framework provides a scalable, cost-effective solution for calibrating hydrological models in ungauged urban areas. By leveraging the "human sensor" data already being collected by transportation networks, cities can significantly improve their flood prediction capabilities and enhance overall urban resilience.
Summary updated