Machine Learning Framework Assesses Roadway Vulnerability Using Aerial Imagery

When Hurricane Idalia slammed into Florida’s Big Bend in late summer 2023 as a Category 3 storm, it left a trail of toppled trees, downed power lines, and blocked roads across rural Taylor County. Barely a year later, in August 2024, Hurricane Debby arrived as a far weaker Category 1 system—yet its slow, rain-soaked passage over Apalachee Bay submerged many of the same roads for days, compounding damage that had never fully been repaired. Two very different storms, one shared stretch of asphalt. Now, a research team led by Samuel Takyi of the FAMU–FSU College of Engineering, together with Eren Erman Ozguven, Mark Horner of Florida State University, and Ren Moses, has turned that natural experiment into a rigorous quantitative framework, publishing in the International Journal of Disaster Risk Science a machine learning pipeline that can automatically detect, classify, and score roadway damage from high-resolution aerial imagery captured after hurricanes.

The heart of the study is a pair of new metrics designed to do what traditional damage assessments cannot: capture both the immediate severity of a storm’s toll on a road network and the longer-term pattern of vulnerability that emerges when successive hurricanes strike the same infrastructure. The first, the Road Closure Impact Index (RCII), quantifies how badly a storm disrupted access in a single event. The second, the Roadway Vulnerability Index (RVI), looks across multiple storms to flag road segments that repeatedly fail—those that consistently close, or that drift between partial and full closure, storm after storm. Together, the researchers argue, these indices transform scattered post-disaster imagery into an actionable map of where infrastructure investment and emergency response resources should flow first.

The technical machinery behind the framework combines two deep learning models working in sequence. The first is a multi-task roadway extraction model built on a ResNet-34 backbone, a convolutional neural network architecture whose 34 layers and skip connections preserve fine spatial detail across image scales. Inspired by an earlier multi-scale road extraction framework, the team customized it with task-specific loss functions, adaptive learning rate scheduling, and preprocessing tailored to post-disaster imagery. The model simultaneously performs road segmentation and centerline extraction, and it performed impressively: on validation data it achieved a Mean Intersection over Union—a standard measure of how well predicted road pixels overlap with actual ones—of 0.876, indicating high segmentation accuracy with minimal overfitting during training.

Once roads were delineated, the second model took over. The team trained YOLOv3—You Only Look Once, an object detection network prized for real-time performance—to classify road conditions into three categories: open, partially closed, and fully closed. YOLOv3 uses the Darknet-53 feature extractor, 53 convolutional layers originally trained on ImageNet, augmented to a fully convolutional 106-layer architecture, and the researchers chose it in part because it is the default detection model within ArcGIS Pro’s deep learning toolbox, allowing seamless integration with the geographic information system where the vulnerability maps were assembled. Training data consisted of 600 manually labeled bounding boxes drawn initially from aerial imagery of Lee County after Hurricane Ian, then expanded roughly fourfold to about 2,400 instances through rotation, scaling, and flipping. Hyperparameters—a learning rate of 0.001, 20 epochs, a batch size of 4, 256-by-256 pixel input tiles, and a non-maximum suppression threshold of 0.3—were tuned on a validation split of 10 percent of the data.

Detection performance was strong for the most operationally critical categories. F1 scores, which balance precision and recall, exceeded 84 percent for both the open and fully closed classes. In Hurricane Debby imagery, roughly 89 percent of fully closed predictions were accurate, and open-road predictions achieved perfect recall, meaning no passable road was wrongly flagged as damaged. The weakest link was the partially closed class, where recall dropped to 47 percent in Idalia imagery—a reflection of the genuine difficulty of identifying partial obstructions from above, where debris, water levels, and vegetation can obscure the visual signature of a road that is passable but compromised. The researchers note that YOLOv3 can struggle with small, occluded, or visually ambiguous road segments, and they suggest newer architectures such as YOLOv5 or transformer-based detectors as candidates for future refinement.

Applied to Taylor County, the framework produced strikingly different diagnoses for the two storms. Hurricane Idalia’s RCII came in at 0.54, corresponding to a normalized value of about 54 percent, while Debby’s reached 0.94—a normalized 94 percent, signaling a far more severe and widespread disruption of roadway accessibility. The interpretation tracks the physical character of each storm: Idalia’s destructive winds and storm surge caused immediate structural failures and debris-blocked corridors, whereas Debby’s prolonged rainfall and 3-to-5-foot surge drowned roadways in floodwater, keeping them closed for days, especially in inland and rural areas still weakened by the earlier hurricane. The index numbers thus encode something disaster managers intuitively understand but rarely measure: a slow, wet storm can cripple a road network far more thoroughly than a faster, stronger one.

The RVI added the temporal dimension. County road CR-38000037 topped the vulnerability table, having been fully closed in both hurricanes—an unambiguous candidate for priority reinforcement or redesign. Several state roads, including SR-38590000, SR-38540001, and SR-38540000, showed moderate vulnerability scores, oscillating between partial and full closure across the two events and suggesting intermittent but real susceptibility. Meanwhile, roads such as SR-38514001 registered an RVI of zero, remaining open throughout. When mapped in ArcGIS Pro with red, yellow, and green symbols denoting high, moderate, and low vulnerability, the result is a spatially explicit risk portrait that county planners and emergency managers can consult before the next storm forms in the Atlantic.

The choice of study area was deliberate. Taylor County is home to roughly 21,800 people, about a fifth of whom are 65 or older—a demographic particularly exposed during evacuations. Its 1,232 square miles blend coastal lowlands and dense forest, and its transportation spine runs along U.S. Route 98 parallel to the Gulf and U.S. Route 221 heading inland, both critical for evacuation and recovery. The imagery underlying the analysis came from the National Hurricane Center and NOAA, spanning resolutions from 1.5 feet per pixel down to 0.25 feet per pixel, with most images at roughly 0.15 meters per pixel—fine enough to reveal subtle damage, debris fields, and flood extents. Roadway shapefiles and evacuation route data came from the Florida Department of Transportation, allowing detected damage to be overlaid precisely on the real network. Images were mosaicked, georeferenced, and resampled with nearest-neighbor interpolation so that Debby imagery matched Idalia’s resolution, ensuring consistent feature detection across storms.

What elevates the study beyond a methodological demonstration is its finding about compounding, sequential disasters. The team documented that residual damage from Idalia measurably exacerbated Debby’s impacts: roads that had been partially restored after the first hurricane were the first to fail under the second. This cumulative vulnerability, they argue, is invisible to static assessment models built on historical data and manual inspections, which remain slow, resource-intensive, and poorly suited to the compressed timeframes of real disaster response. The RCII and RVI, by contrast, can be recalculated as new imagery arrives, and the underlying models can be retrained as fresh data emerge—properties the authors say make the framework adaptive rather than archival, suited to prioritizing debris removal, drainage upgrades, and evacuation route hardening in near real time.

The researchers are candid about limitations. The analysis covered only two storms in a single, predominantly rural county, and results may not generalize to regions with different geography or infrastructure standards. Aerial imagery remains hostage to resolution, weather, and availability, and gaps in coverage can translate into gaps in assessment. Nonetheless, the authors point to clear paths forward: integrating LiDAR and satellite data to enrich the input stream, improving predictive modeling so that vulnerability can be forecast before a storm rather than measured after it, and engaging affected communities to ensure that resilience investments reach the low-income areas that historical hurricanes, from Katrina to Harvey, have disproportionately devastated. For coastal communities on the front line of a warming Atlantic, the message of the work is stark but useful: the roads that fail once will likely fail again, and now, for the first time, there is an automated, quantifiable way to know exactly which ones.