Date of Award

Spring 2026

Abstract

Coastal roadways are increasingly vulnerable to storm impacts and sea level rise. One common failure mode during these events is roadway washout, defined in this study as the physical loss or structural failure of roadway segments due to erosion, overtopping, or hydraulic failure. Washout events can disrupt transportation networks, restrict emergency access, and isolate coastal communities, highlighting the need for improved predictive methods that move beyond index-based assessments toward data-driven identification of vulnerable roadways.This study develops a data-driven framework for assessing roadway washout vulnerability using observed damage data and environmental predictors. Machine learning models, including random forest and decision tree approaches, were trained using documented roadway damage locations in Maine and evaluated across multiple coastal regions. Predictor variables included physiographic and coastal exposure metrics derived from digital elevation models, shoreline data, and meteorological observations. Model transferability was assessed by applying Maine-trained models to additional datasets from New Hampshire, Rhode Island, Texas, and Mississippi. In addition, sea level rise scenarios were incorporated to evaluate how increasing water levels may shift patterns of roadway vulnerability. Results indicate that three variables, particularly roadway elevation and distance from the coast, capture much of the signal associated with washout risk. For example, the Full RF Model trained in Maine achieved a balanced accuracy of 0.74, an AUC of 0.83, and a damage recall of 0.79, while the Linear Regression Model trained in Texas achieved an accuracy of 0.86, an AUC of 0.95, and a damage recall of 0.99 when evaluated in their respective training regions. Models performed well in regions with similar coastal characteristics, such as the Maine-trained Full RF model applied to Mississippi (balanced accuracy = 0.74, AUC = 0.82, damage recall = 0.97), but showed reduced accuracy where geomorphology and feature distributions differed substantially, such as in Rhode Island (e.g., Maine Full RF accuracy = 0.65, damage recall = 0.25; Texas linear regression accuracy = 0.56, damage recall = 0.38). Sea level rise scenarios further suggest that vulnerable roadway segments are likely to expand inland under future conditions. These findings demonstrate a method to use roadway damage locations datasets to identify key drivers of coastal roadway vulnerability and support screening-level assessments for transportation planners and public works agencies.

Document Type

Master's Thesis

First Advisor

Jennifer M Jacobs

Second Advisor

Jo E Sias

Third Advisor

Eshan V Dave

Degree Name

Master of Science

Download

Share

COinS