MULTI-SCALE MODELING OF BUILDING VULNERABILITY IN WILDFIRE

Abstract

In the United States, fire regimes are changing because of climate change, fire suppression, and increased ignitions. Meanwhile, the Wildland Urban Interface (WUI) is growing, exposing more people and buildings to destructive wildfires. In this thesis, I advanced understanding of how to best predict building destruction in multiple ecosystems, focusing on how landscape context and scale influence outcomes. In my first chapter, I used Convolutional Neural Networks (CNNs) to predict building destruction in eight high-intensity Sierra Nevada fires using three different imagery extents (15, 30, and 60 m). In my second chapter, I measured drivers of home destruction in the wetlands of the Atlantic Coastal Plain at 35 extents from 30 to 4,110 m and determined the optimal extent or extents for each factor. I compared models using these optimal extents to models using fixed extents to understand the influence of scale on the ability to predict destruction. In the first chapter, I found that CNNs performed very well for training fires but could not be extrapolated to other fires at any extent. The 15 and 60-m models both outperformed the 30-m model, showing that the extent used for analysis changes outcomes. In the second chapter, I found that the drivers of building destruction in these wetlands differed from drivers in other ecosystems, and that the influence of variables fundamentally changed when measured at different extents. My work in this thesis shows that the extent to which variables are measured greatly influences predictions of building destruction, and that if extent is not chosen carefully, research can lead to misleading conclusions.