Validation of an Augmented Parcel Approach for Hurricane Regional Loss Assessments

While simulation environments for the study of community resilience are rapidly advancing, they remain constrained by the completeness of inventory data. This paper presents an augmented parcel approach leveraging various sources of open data, machine learning modules, and time-evolving rulesets to support Hazus-compatible risk assessments on a wide class of buildings under hurricane wind and flood hazards. These techniques are implemented within the open-source regional hurricane loss assessment workflow of the Natural Hazards Engineering Research Infrastructure (NHERI) SimCenter. Illustrative examples demonstrate building inventory generation in both data-rich and data-scarce environments. The study's validations of computer vision-based modules underscore the importance of training on "in the wild" images labeled with explicit knowledge of the region and representative of architectural nuances such as carports. Validations further reveal the challenges of simplifying complex contemporary roof geometries to the simplified shapes adopted in Hazus and the criticality of accurate year built data, given the augmented parcel approach's reliance on time-evolving code-based rulesets. Published field observations collected in Lake Charles, Louisiana, following the landfall of Hurricane Laura, demonstrate that the use of an augmented parcel inventory within the SimCenter's workflow for Hazus-compatible loss assessments yields damage states consistent with ground-truth observations for minor to moderate damage states. Simulations of extreme damage states (characterized by fewer ground-truth observations) bias toward minor damage for undamaged structures and plateau at moderate damage even for severely damaged and collapsed buildings. This trend persists when considering uncertainty in hazard intensity, as well as the low rates of shutter compliance. Root causes of inconsistencies revealed in this validation exercise will require further processing of street-level panoramic images to generate more samples of severely damaged and collapsed buildings as well as post-2007 construction.