HomePress ReleaseZ-FIRE™

Wildfires in the U.S. are more common and catastrophic than ever before. California, in particular, just experienced the deadliest and most destructive wildfire season in its history with an intense series of fires destroying almost 2,000,000 acres and killing 86 people. In addition to grabbing the attention of citizens, the media and government at all levels, these wildfires have become the top priority for the $2.2 trillion property and casualty insurance industry.

The numerous CA wildfire events during the past 15 months are estimated to reach cumulative insured losses of $27-31 billion according to CoreLogic.  In the wake of the first catastrophic fire events data vendors to the insurance industry reassessed fire risk and out of abundance of caution categorized the vast majority of densely populated areas of California as “High and Extreme High Wildfire Hazard” making it impossible for a carrier to accurately assess property-level fire risk in the most populous state of the United States

Existing methods of assessing wildfire risk rely on 1980’s technology and are broad-based, quantifying risk at a regional level as opposed to at the property level, and are not updated frequently, despite significant changes in vegetation, building stock and weather patterns. In addition, they rely heavily on fuel source as a modifier without taking into account many other factors that significantly impact wildfire risk. Consequently, current methods previously designated only 10% of the properties damaged by the Tubbs Fire in California in 2017 as high-wildfire risk properties, leading to $10+ billion dollars of losses from that event.

 

Map of High Fire Hazard Zones before the 2017 Tubbs Fire and the destroyed buildings from the fire.

Zesty.ai’s WildFire Risk Score (Z-FIRE)

Zesty.ai leverages Artificial Intelligence to analyze 20+ property level attributes related to vegetation, building and roof material, topography of the land parcel among many others to quantify the risk of wildfire at a given property with the Z-FIRE Score.

Our approach to short-listing the variables that impact Wildfire Risk was based on a comprehensive understanding of wildfire science including how wildfires originate, spread, damage properties and eventually get extinguished. For example, spread of wildfires is not always contiguous i.e., they don’t spread from one property to the neighboring property. Depending on wind conditions, fire embers travel farther than the front line of the fire, and they tend to accumulate in nooks and crevices of roofs and decks. In addition, openings in roofs such as unclosed vents create an opportunity for the embers to enter the attic and start a fire. Building and roofing materials are very important in understanding the extent of property damage from wildfire. For example, properties with wooden roofs are more likely to be impacted from wildfires than properties with metal roofs. In addition, the presence of fuel, such as flammable vegetation, other buildings, and other factors. in the primary (0 to 30 feet around the building perimeter) and secondary (30 to 100 feet around the building perimeter) defensible spaces around a property, are considered in determining if a wildfire will spread to a property or not.

Zesty.ai’s approach addresses both of the following key questions. a.) What is the risk of a wildfire occurring at or in the vicinity of a property, and b) If a wildfire occurs in the vicinity, what is the likelihood and extent of damage to the property? Existing wildfire models focus primarily on the former question and can only analyze risk at a regional level. Zesty.ai addresses both questions and quantifies wildfire risk at the individual property level.

To quantify how wildfires damage properties, Zesty.ai accessed aerial imagery immediately before and after several wildfires and ran its computer vision models to automatically categorize all properties into one of three options: not damaged, partially damaged, and fully damaged. Refer to example imagery shown below from the Carr Fire. In the Damage Assessment image on the right, homes that weren’t damaged are in green and homes that were damaged are in red.

Peak Under the Hood

PRE FIRE IMAGERY

POST FIRE IMAGERY

DAMAGE ASSESSMENT POST FIRE

We collected data from several wildfires including the most recent ones from 2017 and 2018. Then, on the pre-wildfire imagery, Zesty.ai ran its property level AI-models to assess several property features related to vegetation, building and roof material amongst many other analyzed attributes prior to the fire. Zesty.ai developed AI-models for each of its property features by ingesting training data – typically hundreds of thousands of sample imagery that were labeled with various values of the property feature. The AI model is trained to recognize patterns (just like the human visual system) in determining property features from imagery at the property level.

Z-FIRE was created using cutting-edge AI-based modeling techniques, which enable the modeling of interactions between variables (e.g. modeling how the presence of vegetation interacts with the health of vegetation). This results in a radical improvement over previous methods in which interaction features need to be explicitly included in the model. Properties are assigned a score 1-10, 1 being the least likely to be destroyed in a wildfire, and 10 being the most likely. Properties with a Z-FIRE Score of 10 are 30 times more likely to suffer damage as a result of a wildfire than properties with a Z-FIRE Score of 1.

Zesty.ai’s Fire Risk Score (Z-FIRE) is highly predictive of which properties would be damaged in the event of a wildfire. Z-FIRE adds tremendous value for carriers and re-insurers with significant exposure to wildfire prone geographies. Z-FIRE can be leveraged as an additional layer to existing underwriting frameworks to decline policies with severe exposure to wildfire risk and to also align pricing commensurate with true wildfire risk.