Clustering neighborhood morphology to explain hurricane damage susceptibility in Bay County, Florida

Abstract

Hurricane events do not impact buildings uniformly across a city. Neighboring structures share similar traits, such as construction era and elevation, so neighborhoods may carry information that models for single buildings miss. Urban morphology has classified neighborhoods into reproducible archetypes considering density and spacing variables, but whether these features explain observed disaster damage has not been evaluated. To test this, we developed a pipeline that clusters neighborhoods by morphology and evaluates them against observed post-disaster damage. Using xBD satellite damage footprints and OpenStreetMap roads for Bay County, Florida (Hurricane Michael, 2018), we extracted four features per 250 m grid cell: building count, coverage ratio, median building spacing, and road density. A Gaussian mixture model clustered 1,754 cells into three neighborhood archetypes, with cluster count chosen by seed stability and bootstrap reproducibility. While the clustering is reproducible, the three resulting archetypes reveal the limits of what these features can distinguish. Two residential archetypes have nearly identical median footprints (238 m² vs. 242 m²), with building count as the distinguishing factor. Isolated and often larger buildings characterize the third, with a median building spacing of 45.1 m compared to 4.7 m and 8.7 m for the residential archetypes, and a 634 m² median footprint. That isolated archetype sustains severe damage at 1.9 to 2.4 times the residential rates (26.7% vs. 11.2% and 13.6%), while the residential archetypes barely differ. However, neighborhood archetypes explain under 1% of variance in damage between cells, so what distinguishes vulnerable neighborhoods appears to lie beyond density and spacing. Adding features describing building shape, height, and land cover, alongside hazard covariates, would test whether richer morphological descriptions can close this gap.