Using Crowdsourced Datasets to Assess and Mitigate Impacts of Recurrent Flooding on Roadway Networks

Recurrent flooding is becoming an increasingly common phenomenon in coastal cities all over the world. In the US, many cities on the east coast are facing the problem of recurrent flooding. This is a phenomenon caused by climate change and sea-level rise, yielding higher high tide events, rise in river water levels, and more severe rainfall events. In recent decades, recurrent flooding events are becoming more intense and frequent. These events typically last for a few hours, and while they do not necessitate evacuation plans, can cause significant impacts on traffic operations with roadway inundation. There have been various studies in the hydrology and transportation domains that cover the impacts of flooding on the transportation network. However, past research uses either static traffic volumes while altering the roadway network, or use simulated traffic flow under altered roadway network conditions. Previous research has mainly been conducted this way due to the lack of reliable, comprehensive empirical data that can provide information about the location of flood incidents, water depth on the streets, and traffic volumes at the locations of the flood incidents. There are very few cities where agencies collect any hydrology or transportation data related to recurrent flooding, and the few datasets that do exist do not provide sufficient spatial coverage or temporal resolution. Apart from agency datasets, there are also crowdsourced datasets providing information on roadway incidents such as flooding. However, the crowdsourced datasets are unregulated and can be prone to erroneous reporting. While there exists research exploring the potential of crowdsourced datasets, there are far fewer studies which establish the reliability of crowdsourced data.
This dissertation aims to address the knowledge gap between the incidence of recurrent flooding and guidance provided to local agencies to mitigate subsequent transportation impacts by establishing four research objectives aimed at closing the research gaps mentioned above (in three papers). First, the research addresses the lack of spatially comprehensive and disaggregate agency-collected traffic volume data. This dissertation proposes a machine learning model which uses unregulated crowdsourced traffic count data to build a traffic volume estimation model. Using these estimated traffic volumes, recurrent flooding impact are calculated on a citywide level using agency-provided flood incident data and on a local level using crowdsourced flood incident data. The second part of the research addresses the flaws of using an unregulated crowdsourced dataset by creating a model to estimate the trustworthiness of crowdsourced flood incident reports. With the ability to distinguish high-confidence crowdsourced reports from low-confidence reports, the filtered dataset can be applied towards the goal of real-time traffic management under recurrent flooding conditions. The last part of this dissertation assesses dynamic vulnerability of different locations throughout the roadway network under predicted recurrent flooding, through a traffic impact index. This research establishes a relationship between the hydrological, roadway, and environmental characteristics of the locations of the high-confidence crowdsourced flood reports and the observed traffic impact index. This relationship can then be used to predict near-real time traffic impact index at locations with insufficient data, such that emergency services can prioritize the locations predicted to experience high impact for faster mitigation of recurrent flooding disruption.
The results of the research show that unlike major disasters, recurrent flooding impacts are not well-observed when generalized on a citywide scale. While citywide impacts in the Norfolk case study showed a 3% overall decrease in vehicle-hours of travel (VHT) for agency-collected flood incidents, the localized impacts analysis showed a 7% decrease in VHT and a 12% reduction in VMT in the sub-areas of flood incidents for crowdsourced flood incidents. These impacts of recurrent flooding are also not uniform and vary across different roadway types, time of day, and location. Since these crowdsourced flood incident reports are unregulated, a trustworthiness model is created using contextual data to separate trustworthy and untrustworthy reports. It was found that the model’s prediction accuracy was about 91%. When applied to the crowdsourced flood incident data, about 72% of the data was deemed trustworthy. Using the trustworthy crowdsourced flood report data and traffic volumes data, a machine learning model is created to estimate the traffic impacts at locations where near-real time traffic volume information cannot be estimated. The preferred prediction model showed a normalized root-mean-square-error (n-RMSE) of 14% in the dataset. About 67% of the data was predicted within one standard deviation of the observed values. These models are limited in interpreting the results due to a small sample size. Nevertheless, this dissertation demonstrates a framework for using crowdsourced data that has a potential to quickly identify and predict likely high impact flood event locations so that appropriate emergency response measures can be taken to reduce the traffic impacts due to flooding.
This research shows the potential of crowdsourced data as more than just information and as a significant reinforcement to the spatially and temporally limited agency datasets. With the use of various crowdsourced datasets, spatially and temporally disaggregate analyses on the roadway network are now possible, with potential applications in the decision-making framework for the cities experiencing deteriorating impacts of recurrent flooding. The various trustworthiness, volume estimation, and traffic impact prediction models created through this dissertation have a broad spectrum of applications within and outside of transportation networks. The models developed in this dissertation will help to close the knowledge gap in existing practice and will also ensure the best use of available resources in applications that examine the use of crowdsourced data in disruption mitigation efforts.