Constructing and Validating Geographically Refined HAZUS-MH4 Hurricane Wind Risk Models: A Machine Learning Approach

Abstract

Accurate estimation of risk to structures due to hurricane wind hazards is critical for emergency planning and post-event recovery. A widely used model for assessing risks due to hurricane winds at the county or zip code level is the HAZUS-MH4 suite of programs developed by the Federal Emergency Management Agency (FEMA). Very few studies have attempted to assess the predictive accuracy of HAZUS wind risk models at levels of geographical specificity below counties and zip codes. In this paper, we present a methodology for computing HAZUS predictions of wind damage aggregated to the level of one kilometer square blocks. We use detailed damage data on over 700,000 single-family residences in Harris County collected in September 2008 after Hurricane Ike for evaluating the predictive performance of HAZUS. We show that at this level of aggregation, HAZUS has a predictive accuracy of 29.5%. This finding raises the question of how to improve the predictive accuracy of HAZUS-MH4 models at this level of spatial specificity. It is difficult to manually analyze errors made by HAZUS over a data set with 700,000+ structures, so we devise a statistical machine learning approach to distinguish structures whose damage is correctly predicted by HAZUS from those that are incorrectly predicted. We identify ten potential explanatory variables at the level of each building. These variables include number of floors, terrain, age, whether a building was remodeled, building and land value, proximity to highway, and more refined wind measurements including wind speed, wind swath and wind direction. Our machine learning classifier reveals the relative importance of these variables for improving HAZUS’s predictive performance. In particular, we find that HAZUS’s fragility curves for one and two floor residences in urban terrains need to be refined by incorporating building and land value, age, proximity to a major highway, as well as wind direction and wind swath. These results can be used to build the next generation of HAZUS fragility curves for accurately predicting hurricane wind damage to structures at a spatial specificity of one kilometer square blocks.