Researcher(s)
- Kevin He, Applied Mathematics, Johns Hopkins University
Faculty Mentor(s)
- Pinki Mondal, Geography and Spatial Sciences, University of Delaware
- Mehrshad Amini, Department of Ocean Engineering, University of Rhode Island
Abstract
First Floor Elevation (FFE) shapes how exposed a building is to surge and wave forces during flooding. One foot of freeboard above Base Flood Elevation (BFE) cuts Average Annualized Loss by 90–93%; a second foot brings that down to 96–99%. However, 90% of structures have no official elevation certificate, largely because manual surveys are expensive. Existing AI methods for estimating FFE mostly detect front doors and apply fixed-height assumptions, which is ineffective when the door sits above the first floor or doors come in irregular sizes.
We built a two-stage pipeline that estimates FFE automatically from Google Street View imagery. First, we fine-tuned a YOLO11l model to recognize architectural features (doors, windows, garages, elevated space, floors, roofs, stairs, and the building outline itself) across 600 annotated images from Fort Myers Beach, Florida, each tied to a ground-truth FFE measurement. We then fed those detected features, along with the imagery, into a DINOv3 Vision Transformer to regress a continuous FFE value. To check how much the detected features actually helped, we compared this against a version of the model that saw only the raw street-view photo.
Our trained YOLO11l model outperforms zero-shot Grounding DINO, achieving an mAP@0.5 over double that of Grounding DINO, indicating that domain-specific training is necessary. Incorporating annotation-derived architectural features into the DINOv3 regression halved the Mean Absolute Error (MAE) relative to an RGB-only baseline and raised R² substantially. Backbone comparisons also show that among DINOv3 models ViT-B/16 has the best overall accuracy, while ViT-S+/16 delivers comparable accuracy at lower computational cost. As a result, our findings suggest that these automated, feature-aware pipelines can support scalable, accurate FFE measurement, providing a strong path toward improving the accessibility of elevation certification for flood risk assessment.