A better way to measure fire risk? Artificial intelligence may be answer
by Sarah Holder, CityLab | posted: February 24, 2020
Driving through California’s Sonoma County, the remnants of last year’s fire season – and the season before that – are impossible to ignore. Blackened trees still line the highway. A sign, one of many like it in the county, reads, “From the ashes, we will rise.”
Indeed, new homes in various states of construction are rising in neighborhoods that were recently reduced to cinders by the 2019 Kincade Fire, which destroyed more than 350 structures on 77,000 acres in Sonoma County, and the 2017 Tubbs Fire before it, which was smaller but more devastating. Many California homeowners chose to rebuild, but insurance companies have preferred to retreat. Over 340,000 Californians were stripped of their wildfire insurance plans between 2015 and 2018, according to the insurance news site Gavop; other residents were served with premiums that jumped by up to 500 percent.
These rate hikes and non-renewals are informed in part by disaster risk models, which insurance companies use to assess potential damage and decide how to underwrite a property – models that, in California, are getting increasingly conservative. And no wonder: Insurance companies spent $24 billion on wildfire damage in the state between 2017 and 2018. In a warming world, future fires stand to be even bigger and more destructive.
For the more than 1 million California homeowners who live on land deemed by the Department of Forestry and Fire Protection in 2010 as a high-risk wildfire zone, and the 4.5 million other homes that lie along the path of potential destruction, in what’s called the broader wildland-urban interface (WUI) zone, losing insurance protection could be catastrophic. In an effort to slow the insurance exodus, California’s insurance commissioner, Ricardo Lara, has ordered insurance companies to keep protections in place for 1 million homes in wildfire-prone counties for the next year. And state assemblywoman Lorena Gonzalez co-introduced a bill on February 18 that would force insurance companies to preserve plans for residents who have proved they’ve made progress in fireproofing their property, according to the Los Angeles Times.
But what if the current methods of modeling wildfire risk aren’t strong enough?
“We can say with confidence that close to three quarters of California has no to fairly limited [wildfire] risk,” said Attila Toth, the founder of Zesty.ai, a start-up that uses artificial intelligence to develop disaster risk assessment tools. “This is a low single-digit percentage problem in California.”
Zesty.ai is one of a growing number of “insurtech” startups that are hoping to disrupt the decades-old methods used by the insurance industry by introducing a more detailed modeling system. Armed with artificial intelligence, high-resolution satellite imagery and 100 years of wildfire data, Toth says his company can provide a more granular idea of what homes are truly threatened by fire, and how severely.
Gaining a sharper image of risk isn’t just about keeping homes safe, Toth says. It’s about shaping the state’s development patterns, making decisions about where to rebuild or retreat – and, more immediately, putting a more precise price on each home’s insurance.
Legacy players largely welcome the disruption, says Janet Ruiz, the Insurance Information Institute’s California spokeswoman: “We understand the need.”
Zesty.ai has a partnership with MetLife, which is expanding its homeowner insurance base in the West Coast. The venture capital arm of State Farm, the largest home insurance company in California, has invested in Cape Analytics, a San Francisco-based company that uses its own AI models to analyze geospatial property data. Insurance companies like The Hartford, Security First Insurance, and CSAA use the platform for their disaster modeling.
The problem these AI-driven companies are set up to solve is that the old model of disaster risk assessment can be imprecise: Neighborhood-level footage or risk maps drawn by the Department of Forestry and Fire Protection inform assessor decisions for whole zip codes. The way risk scores for each property are calculated is generally opaque, as the Voice of San Diego determined in a 2019 report, and the method varies by model. Some are developed by third parties, and some are unique to insurance companies themselves.
FireLine, one of the most popular risk assessment tools used by insurers like AAA and Mercury Insurance, rates properties on a scale of 0 to 30, with 30 being the highest risk. It gives a score based on a quarter mile radius around each home, a vice president of Verisk, which developed FireLine, told the San Francisco Chronicle; the model also takes into account vegetation, topography, and how easily firefighters might be able to intervene. State Farm has its own in-house scoring system, says representative Sevag Sarkissian, which “considers factors such as the location of the property relative to natural hazards, condition of the property, and the customer’s past claim activity and history.”
It’s these kinds of risk models that produced the wildfire insurance conundrum California is in today, critics say. In a 2018 report, the California Department of Insurance recommended that the state legislature introduce more oversight over how risk is calculated, and called into question the current rating system’s accuracy. Left with fewer private options, the number of homeowners that turned to the state’s public coverage option grew 177 percent between 2015 and 2018, the department found.
At the Zesty.ai office in Oakland, Toth showed me a map to illustrate the gap between risk predictions and reality. The Tubbs Fire of 2017 destroyed about 5,500 structures (marked in blue), but only about 600 of them were within the land officially deemed high-risk by the state (marked in pink). Of course, that fire was extraordinary – the worst in the state’s history – meaning there was little precedent to draw from. Still, Toth says, with more data points, the map could have looked different.
“Many of them are amputating limbs when they have a nail infection,” said Toth of the legacy models. “Our model is looking at it surgically.”
Using low-flying planes, satellite imagery, and building codes, Toth says Zesty.ai looks at 65 risk factors – not only geography but each particular home’s physical attributes, like the radius of vegetation it’s nuzzled by, its roof material, the slope of the land it sits on, and the risk level of the house next door. This helps them pinpoint the likelihood that a particular property will be destroyed in a fire. Instead of condemning large swaths of the California population to the same risk level, the company’s tool might hand vastly different scores to two homes on the same block – reflecting the reality that when a fire hits, neighbors can have very different fates.
Cape Analytics uses satellite and aerial footage to get a similarly layered view of risk, but gives the information to an insurance company without translating the data into a risk score, as Zesty.ai does. This makes its findings more transparent, said Kayvan Farzaneh, a representative for Cape Analytics.
AI interventions should be seen as tech-powered tweaks on old rating models, not a fundamental revolution, says Kevin Van Leer, a client solutions manager at Cape Analytics. “The idea of understanding all the risk factors that go into when a wildfire will occur … that’s not really new science, it’s been around for a while.” These tools can do that analysis faster, with fewer resources, and, depending on the image sources, with a higher-resolution view.
If harnessed correctly, they can also help address another problem the CDI identified, and that Gonzalez hopes to address with her legislation: the fact that many of the existing models fail to take into account the efforts made by homeowners to mitigate their risk.
Both Zesty.ai and Cape Analytics can quickly reassess risk scores based on consumer behavior, representatives from the companies say. To prove that a property owners’ hedges have been trimmed, for example, or that they’ve installed ember-resistant vents, an assessor doesn’t have to swing by the house to check it out visually. Instead, the program can scrape another picture of the same plot and get a new score immediately.
The tools help customers understand their policies, as well, Toth says: If an insurance package is denied, or a premium hiked, companies can tell customers exactly why – your neighbor’s vines are hanging over your patio, for example – and give them the chance to fix it.
Like many technological fixes meant to streamline processes usually mediated by a human, AI-powered risk assessments could bring unintended consequences by introducing new biases or shifting the burden of paying for risk. “We are seeing insurance companies over-rely on technology, and the consumers are paying the price,” Emily Rogan, the COO of United Policyholders, which advocates for insurance customers, told Axios.
Consumer Watchdog, a group that advocates for consumers, wrote a letter to the CDI’s Lara in August, encouraging him to pass emergency regulations protecting policy-holders from rate hikes by limiting the use of so-called “black box” risk models such as FireLine without clear disclosures about how they work. “These models use algorithm-based projections of wildfire losses generated by third-party vendors to support massive rate increases,” they wrote. “Emergency regulations should bar the use of any models that lead to excessive or discriminatory rates.”
While AI-driven models profess to be more clear about the breakdown of risk factors, the consumers’ access to information is still granted at the insurance company’s discretion. (Consumer Watchdog did not respond to a request for comment.)
Wildfire doesn’t discriminate, says the Insurance Information Institute’s Ruiz. But it does reflect the geographic disparities of California real estate. Coastal properties with high property values are most at risk, and are more expensive to insure and rebuild – Sonoma County, for example, is wine country, full of expensive second homes and vineyards.
Meanwhile, research in the science journal PLoS ONE shows that, though “fire-prone places in the U.S. are more likely to be populated by higher-income groups,” low-income households are still more vulnerable overall, with fewer resources to pay for fire prevention tools before an incident, and, later, rebuild. In California, residents of higher-poverty exurbs in fire-prone areas would find it harder to afford the mitigation strategies that bills like Gonzalez’s would reward.
Nick Allain, a spokesman for Zesty.ai, shared results of an internal analysis that showed no correlation between property value and “Z-FIRE” score. Still, less-precise maps that spread similar insurance premiums across a wide area help spread the costs of disaster preparedness more evenly. Critics of leaning on technological advances say collecting more information—however accurate—may inevitably result in wider cost disparities between homeowners.
Ruiz says that’s just how the insurance industry works. Buildings that are in danger of being destroyed year after year deserve to be assigned higher risk scores, she says, regardless of who lives within.
But she says that having more granular information may compel painful—yet practical—decisions. “If you’re living on the edge of the cliff and it’s falling apart, you have to decide whether you can afford that.”
If insurers and state insurance departments want to protect vulnerable neighborhoods, they should take a more proactive role in funding fire prevention, Cape Analytics’ Farzaneh added. “It’s kind of like vaccines,” he said. “If you have a community that’s done mitigation community-wide, it makes the entire community safer. … Hopefully having more granular data will help that happen.”
As the planet warms, the stakes of these decisions will grow graver. Though Toth paints a more optimistic picture of current wildfire risk than existing models do, he echoes scientists in saying that higher temperatures and intensifying dry seasons are already making the state’s wildfires larger. Looking further into the future, the company recently prepared a 10,000-year simulation showing where wildfires might one day burn across California. The southeast part of the state is predicted to keep its low frequency of wildfire events, while parts of the middle of the state and the coastline (shaded yellow, orange, and red) could see more ignitions.
“[It’s] very much directionally aligned with ignition history over the past 100 years,” said Toth. “But given more vegetation growth, and more urban development within the wildland urban interface, we expect that these events are going to create higher losses in the future.”