Developing models for building damage estimation for wildfires using Earth Observation (EO) data, Wildland Urban Interface (WUI) maps, and structural vulnerability functions

Abstract:

In recent years catastrophic wildfire events in the Western US caused significant economic impacts and posed a major risk to public health, public safety, and the environment. Wildfires are becoming more severe, common, and expensive due to increased temperatures, decreased relative humidity, and in general a warmer, drier climate from the effects of climate change. Additionally, increased building development in the Wildland Urban Interface (WUI) is putting more human lives and structures at risk. This worsening wildfire situation is of major concern for first responders, government agencies, policymakers, and property insurers. The catastrophic economic losses in recent years underscore the need to more accurately estimate future risk and loss potential to support effective risk management decisions. This research focuses on several aspects of wildfire risk by combining Earth Observation (EO) data on wildfire detection and analysis, new models of Wildland Urban Interface (WUI), and development of structure vulnerability or response functions to estimate damage to residential buildings. For events in California between 2018 and 2020, wildfire impacted regions were generated using EO data from NASA’s real-time fire and thermal anomaly sensor data from MODIS (Aqua and Terra) and Suomi NPP VIIRS I-Band (375-meter resolution) sensor data. New and updated WUI areas were mapped and provided by the University of California Irvine (UCI) research team (Banerjee and Nguyen and their graduate students). Damage data and classification of severity for ten fires were obtained from the California Department of Forestry and Fire Protection (CAL FIRE). Building upon the EO data from wildfire events and the new WUI model, a data-driven framework can be developed to provide estimations of damage to buildings subjected to wildfire hazard in both active fire situations (i.e. using near-real-time EO data) and forecasted future events. Machine learning techniques can be used on existing data from recent wildfires (2018-2020) to classify buildings into various damage states by models trained at different levels of detail. Preliminary results will make use of parameters such as Fire Radiative Power (FRP) and brightness temperature (available from EO data), building density, and the existence of various types of WUI for a given building location to estimate its potential damage state. On a second and more detailed level, building-specific data (e.g. roof type, eaves condition, exterior siding material, etc.) gathered from damaged and undamaged buildings can be added as additional features to train a model capable of capturing some of the contrasting damage states observed in wildfire events for adjacent buildings.