The American West is equal parts myth and reality. It’s a place known for its incredibly beautiful and equally savage landscapes, from scorching-hot Death Valley in the east to the ice-cold Pacific Ocean in the west to the snow-capped Sierra Nevada in the north, where names like Donner Pass serve as a testament to the lengths man will go to survive.

Over the past few years, another near-mythic danger has arisen in the West: wildfires.

Miles and miles of forests erupt into flames each year during the annual wildfire season. And the impact of wildfires grows exponentially each year. Megafires — wildfires greater than 100,000 acres — are growing more and more common as forested areas face drier and hotter conditions. We’re barely two months into wildfire season and this year’s California blazes have burned 726,667 acres according to the California Department of Forestry and Fire Protection (CALFIRE). This is 18 times the five-year average of acres burned through mid-July of 39,000.

Wildfires don’t just impact the American West. They’re a global existential and ecological threat. Among other environmental effects, wildfires contribute to air pollution and degrade groundwater and soil. By burning canopy vegetation and replacing ground cover with ash (a hydrophobic material), wildfires cause more water to flow over the land surface during storms. This leads to increased flooding, landslides, erosion, and runoff of sediment, ash, and other pollutants (including legacy mining waste, formerly covered by revegetation) into surface water. This can result in lower water quality, decreased reservoir storage capacity, loss of soil nutrients, and stream habitat degradation. 

Wildfires also release millions of tons of carbon dioxide into the atmosphere in an age where humanity is fighting climate change, leading to increased respiratory illness even in far-removed locations. And they lead to loss of life and property. A report examining California wildfire data from 2017 to 2021 estimated that the state incurs roughly $277 million in wildfire-caused property loss annually.

Photo from GRID Arendal – Flickr

Considering the damage that is done by intensifying wildfires, it is crucial for us to harness technology to adapt our fire mitigation response. Two interrelated tools that hold promise in enhancing wildfire management are artificial intelligence (AI) and machine learning (ML).

Predicting wildfire occurrence

The fire triangle consists of three factors — oxygen, fuel, and heat — which, when present together cause a fire. 

Source: Wikimedia Commons 

Three things, often called the “fire triangle,” must all be present at the same time in order for there to be a wildfire: an ignition (lightning, a matchstick, etc.), fuel (most often vegetation), oxygen (most often winds over 19-20 mph). There are various AI and ML models trained to detect and predict fires based on these three elements.

1. Weather signals and thermal data

ML models can be trained to detect weather signals such as changes in temperature, humidity, and the air pressure differential, which indicate the likelihood of a fire occurring. ML data is especially effective in predicting wind speeds, both now and in the near future, and the temperature of the wind. 

Various NOAA and NASA tools provide the data needed by ML systems to predict future wildfires. They collect visible and infrared images of weather phenomena; map out heat intensity; and analyze aerosol properties. 

Two examples of such tools are the Visible Infrared Imaging Radiometer Suite (VIIRS) instrument and the Moderate Resolution Imaging Spectroradiometer (MODIS). VIIRS collects visible and infrared images of various weather phenomena including aerosols, fire, smoke plumes, and chlorophyll. The data from the VIIRS tool is used in detection of fires and even measuring atmospheric pollution. MODIS measures the properties of aerosols in the atmosphere. An abundance of aerosols indicates some source of pollution, which can be narrowed down using MODIS’ analysis. Past applications of MODIS’ data have been to detect where biomass is being burned, and where dust storms, volcanic eruptions, and forest fires are occurring.  Sometimes, instead of satellites, thermal aircraft are used to get heat data on specific regions.

There’s a long history of AI/ML-enhanced predictive models in hurricane forecasting. Hurricane forecasting often uses deep learning-enhanced ensemble models. These models run simulations from a varied set of initial weather conditions, which are used to predict multiple outcomes of how intense a hurricane will be and what its path will be. This approach is superior to using deterministic models, which only provide one outcome based on one initial set of data.

2. Predicting ignitions

Recalling the fire triangle, it is crucial for an ignition source to be present alongside fuel and oxygen for a fire to occur. If there’s wind and vegetation but no ignition, then there won’t be a wildfire. The most common source of  ignition for naturally-occurring wildfires is a lightning strike.

The LightningCast AI model attempts to predict where lightning will strike within the next hour. The model gets a feed from two National Oceanic and Atmospheric Administration (NOAA) GOES-R satellites and goes through more than 6,600 images to determine where the wildfire will occur. 

In Türkiye, as a part of the World Economic Forum’s wildfire technology initiative called Fire AId, KoçDigital developed a wildfire risk map for the South Aegean and West Mediterranean regions of the country. Four categories were used — very high risk, high risk, medium risk, and low risk. In an attempt to account for human activity and probability of human-caused wildfires, the company also considered in its maps the presence of roads and farm areas and population density. 

3. Commentary

A predictive ML system should also have as part of its data set a nationwide map of distribution lines — both those that are new and those being developed. This addition will allow emergency services and infrastructure providers to make the most informed decisions

The major limitation of using AI with wildfire management is that a one-size-fits all model is especially difficult because of our planet’s multiple landscapes, teleconnections (faraway weather anomalies), and limited local data.

We need more insights on a local basis, and what allows these insights to be collected is the development and support of wider weather station networks. For example, in California, public utilities have sponsored the creation of thousands of weather stations — 1500 fall within Pacific Gas & Electric’s territory in Northern California. These stations provide data for public consumption and for use in utilities’ own risk models. The abundance of local data allows utilities to take national and international weather models, combine them with local models, and make more informed predictions and decisions.

Detecting wildfires

We can more effectively detect wildfires using AI. Two applications of AI in wildfire detection are in cameras and in remote sensing systems.

1. Cameras

If used as part of a camera system, AI can examine images to identify visual indicators of fire like smoke and heat shimmers, which suggest a fire has broken out. Thanks to machine learning, the AI system would never stop learning, so it would only increase in accuracy with time. 

ALERTCalifornia, a result of a public-private partnership between Digital Path and the University of California, San Diego, is a leader in wildfire detection camera systems. The organization has more than 1,080 high-definition cameras with 360-degree sweep capability deployed across California (as of June 2024). You can check out live footage from ALERTCalifornia’s cameras here. The cameras also have near-infrared capabilities for night monitoring. The cameras can view up to 60 miles on a clear day and 120 miles on a clear night. ALERTCalifornia worked with its parent company Digital Path as well as CALFIRE to develop an AI tool that TIME named among the best inventions of 2023. 

Here’s how it works: When even one camera spots a potential fire, all the cameras around the fire are pivoted to look at the same spot and to assess the threat level. Once it is determined that the fire is dangerous, the fire alert (which includes a percentage of certainty, and estimated location for the incident) is sent to CALFIRE’s dispatch. After human vetting, firefighters are sent into the field. Camera data is also used to inform other fire suppression tactics like sending alerts to evacuate specific areas or to park on asphalt when parking on the side of the road, or declaring a certain area a “no barbecue zone” or “no camping zone.”  

As seen in the case of ALERTCalifornia’s tool, AI’s role in shortening response time is crucial to timely and effective wildfire management.

2. AI on remote sensors

Dryad’s ultra-early detection remote sensor is placed on trees in California’s densely forested areas.

Source: Washington Post (licensable)

AI is being used on small remote sensors that use chemical indicators to detect wildfires. These sensors are placed on poles, on drones, on trees, or even dropped like confetti into the forest. Depending on the type of sensor used, they can provide up to 10 km² of coverage. They don’t just detect heat, smoke, and changes in humidity, but also detect presence of gasses. AI on these sensors can be trained to recognize a wildfire from a campfire or construction dust, allowing for fewer false positives. 

Here’s a breakdown of how remote sensing systems work: When the system gets an alarming chemical reading, it sends data to a cloud-based platform enhanced by AI to identify patterns, anomalies, and trends and accurately detect fires. After the system validates that there is a wildfire happening (and humans have the final say), alerts are sent out to emergency response teams, first responders, city officials, or system administrators.

CALFIRE has partnered with Dryad to place 400 of their ‘ultra-early’ wildfire detection sensors in Jackson Demonstration State Forest, an area prone to wildfires.  These sensors measure presence of gasses like hydrogen and carbon monoxide, as well as temperature, humidity and air pressure. They can detect fires within minutes of them beginning, when they’re “smoldering.” By contrast, cameras take a few hours more when wildfires have already become “open fires,” and satellites often only pick up on fires when they are rapidly spreading and out of control. 

Increased adoption of remote sensors should be a priority, especially since many existing systems rely heavily on visual data sources (like cameras and satellites) which falter in poor weather, darkness, or when terrain makes it hard to spot smoldering fires. 

Mapping the path of ongoing wildfires

Once a fire has reached a certain size, the only recourse available to Emergency Management Services (EMS) agencies is to consider how they can most effectively contain the fire, for which they need to understand its projected path. Both generative AI and ML simulations are being used to generate accurate models of the projected path of ongoing wildfires.

1. Using generative AI

An early study proof from researchers at the University of Southern California (USC) has shown how satellite data of a forest fire’s real-time progress (along with some historic data) can be used by AI to forecast trajectory, speed, and what is in the fire’s path. By carefully studying the behavior of past wildfires, the team of researchers understood the patterns in how they started, spread and were eventually contained. They saw that weather, topography (wind-land alignment, vegetation, vegetation density, etc.), and other factors also influenced wildfires. Then they created a generative AI model trained to understand the fire triangle and the added factors they noted, and to recognize patterns in satellite images that match up with how wildfires spread in their model. Using AI in this way helps in building a fire potential index model which provides insight about the relative risk of wildfire.

2. Using simulations

Another company making strides in situational awareness technology for wildfires is TechnoSylva, which serves both electric utilities and fire authorities. Using weather and fuel moisture prediction systems and simulations, TechnoSylva’s platform generates daily wildfire risk forecasts, real-time wildfire spread predictions, and calculations of potential impact. 

CALFIRE is currently using TechnoSylva’s platform to run simulations to track ongoing fires, including the Park Fire, and inform decisions on retardant deployment and other fire suppression measures. 

PG&E and other electric utilities use TechnoSylva’s risk forecasts to determine risk of ignition by utility assets. The tool has a four-day horizon and informs utilities of which lines to de-energize. Data from TechnoSylva can also be used by utilities to make asset-hardening decisions; that is, to decide which lines/poles to make more resilient against fire threats, and how to do so. 

3. Using machine learning

A government effort that uses machine learning is NASA’s AUDREY (Assistant for Understanding Data through Reasoning, Extraction, and sYnthesis). The tool collects critical information like temperatures, the presence of gasses, how quickly a fire is developing, how the items being burned will affect fire growth, and fire flow paths. It then uses this bank of information to help first responders tackle wildfires in the most efficient and safe way possible. Currently, the software is being refined so it could be used to manage a whole team of firefighters.

Looking ahead

Availability, accessibility, volume, and content richness of data, along with inability to predict human behavior and lack of an ultra-high performance AI system are hurdles in AI-assisted wildfire risk prediction. But these are small obstacles in a field that is ever-evolving. As new AI, machine learning, and deep learning approaches emerge, and research into quantum computing and other emerging technology begins, the possibilities for innovation in wildfire risk management technology are endless.