In a world overshadowed by the rapid deterioration of its climate, AI engineers are beginning to apply machine learning technology to issues of environmental justice and crisis. In California, where wildfires have ravaged more than 570,000 acres of land and forced 41,000 evacuations already this summer, AI is being used to prevent and monitor forest fires throughout the state (1) (2).

Last summer’s wildfire season left 4.3 million acres of land destroyed, marking a record high for California. With peaking hot and dry conditions, this year’s wildfire season is starting earlier and has the potential to match or exceed last year’s levels (3). Scientists estimate that more than half of the state’s forests are at high risk of forest fires. 85% of California is in an extreme drought, 41 of the state’s 58 counties have declared a drought emergency.

With government budgets stretching to accommodate new fire agency demands and an immediate need for human resources to fight forest fires, the future of these Californian communities and habitats might lie in the hands of AI.

Screen Shot 2021-08-28 at 1.37.42 PM.png

Geospatial AI gleans insights from satellite and camera imagery, as well as other forms of remote sensing, and is being used worldwide to address large-scale environmental events (4). With the recent increase in forest fires across the Western United States, AI technicians and environmentalists are working together to apply the technology to preventing, detecting, and monitoring wildfires on the West Coast.

AI technology is instrumental in interpreting large volumes of visual data, and there are numerous ways in which visual data can be leveraged to prevent wildfires. Some initiatives, such as the California Forest Observatory, are leveraging AI to identify regions at high risk of ignition, and working to improve resilience in those regions. Others, such as the ALERTwildfire project, are using AI to reduce the time it takes to alert forestry departments of new fires, with hopes of stifling them before they have spread. Finally, projects like Maxar Technologies are using AI as a tool for tracking and monitoring fires as they grow, which helps local governments plan evacuations and extracts important information for civilians curious about the fate of their homes and communities. With mixed results and successes, these different approaches all utilize geospatial AI to mitigate the threat of forest fires.

One approach to hindering wildfires is to highlight regions that might be vulnerable to new fires and prevent them before they ignite. Conservation technology companies Salo Sciences and Planet have taken this approach in their joint project, the California Forest Observatory (CFO). The CFO uses real-time imagery from the sentinel-1, sentinel-2 and Planet Labs satellites to map the various drivers for wildfires including topography, weather patterns, vegetation fuel, and infrastructure (5). Salo Science’s website cites the objectives of the project to be improving forest and community resilience in high-risk areas, and to provide scientists and government agencies such as CALfire with accurate and up-to-date information about fire threats throughout the state, a task that would be impossible using only human eyes (6).

A more targeted approach is to survey landscapes for small fires in hopes of reducing the time it takes for agencies to become aware of new blazes.

As they grow, wildfires create rapid air movement as heat rises from the flames and cold air is sucked towards the fire like a vacuum. Under the correct meteorological conditions, the vacuum can create violent winds that rival events like hurricanes and tornadoes (7). In an interview with The Atlantic in 2018, wildfire expert Park Williams remarked that “When it’s that big, and there are helicopters dropping water and retardant on it, they’re doing nothing. When you see firefighters spraying hoses at it, [the fire] is so hot that they can’t even be close enough to be within a hose-shot.” Williams went on to say that “The fire, to me—it’s like an ocean. It’s so strong that we don’t really stand a chance of doing much to it” (8). Thus, locating a fire as quickly as possible is crucial to firefighting; it provides firefighters with access to more techniques and opportunities to effectively put out the fire (9).

Traditionally, local governments have relied on civilian 911 calls, airplanes, and lookout towers to report fire incidents. These reports are often unreliable and come too late, allowing the flames to spread before firefighters arrive on scene. Projects in California, New Mexico, and elsewhere are leveraging geospatial AI technology to reduce the time that it takes for fires to be reported and extinguished. AI algorithms assess in real time massive amounts of camera and satellite footage and identify smoke and flames from new wildfires. The systems then alert local authorities and dispatchers of the new ignition.

Founded by Graham Kent, The ALERTwildfire system is one of the leading projects in California using AI to identify wildfires. In Sonoma County, just north of the San Francisco Bay Area, ALERTwildfire is using twenty one pan-tilt-zoom (PTZ) fire cameras to survey the county’s terrain for flames and smoke. Every 10 seconds, each camera sends an image for analysis by an AI algorithm trained to compare the photographs with historical images of the same landscape. Although the company has more than 800 cameras located across 8 states, it is only in Sonoma county that ALERTwildfire has incorporated artificial intelligence (10). The South Korean based AI company Alchera has partnered with ALERTwildfire to provide “image recognition AI”, which is being used to sort through the imagery at “a rate impossible for human eyes” (11).

The algorithm struggles with false-positive alerts, and is still in training to distinguish wildfire smoke from other aerosols like geysers and clouds. It is estimated that the system will need to witness 70 more fires before it is fully accurate. Although it is still a work in progress, incorporating AI has significantly decreased the time it takes for fires to be identified and reported. The company claims a 10 minute difference between ALERTwildfire alerts and 911 calls reporting the same fire.

One company based out of New Mexico has developed a similar system for reporting wildfires using satellite imagery instead of cameras. Descartes Labs uses visual and geothermal imagery from two of the NOAA’s (the National Oceanic and Atmospheric Administration) satellites, the GOES-16 and GOES-17 satellites. Both satellites are in geostationary orbit above the continental United States, meaning that they rotate in sync with the earth in order to remain “stationary” above the earth. Descartes Labs receives updated images of the New Mexico terrain every five minutes, and multiple geospatial AI algorithms are run to identify new wildfires. Among these is an algorithm that checks for abnormal spikes in the earth’s temperature in target areas and an algorithm that compares visual imagery of the land to previous images, similar to how the ALERTwildfire algorithm detects fires. Additionally, the algorithm takes into account the locations of large copper mills and oil refineries to reduce the number of false positive alerts (12).

NASA’s GOES-16 Satellite. Source: Ryan Howerter. License: https://creativecommons.org/licenses/by-nc-sa/2.0/legalcode

Descartes Labs is currently doing a trial with the New Mexico State Forestry Department. If the algorithms all conclude that a wildfire is present, a report is sent to the State’s foresters. According to the lab’s website, only nine minutes are needed for the entire process. This assertion highlights the critical role of time-frame in successfully smothering young fires.

Despite efforts to prevent new fires from igniting and spreading, this summer has already seen a number of catastrophic forest fires. As of August 23rd, Northern California’s Dixie fire is the second largest fire the state has ever seen. The fire has destroyed 1,259 structures and burned 725,821 acres of land (1). In early August, 107 wildfires were burning in 15 states across the country. And so alongside its applications in detecting and stifling new fires, AI technology is also being used to monitor large wildfires in order to extract valuable information about the movement and boundaries of the fires.

Historically, firefighters have sourced their information from airplanes that photograph the active fires after dark. This monitoring is inefficient and slow, and is made very difficult by billowing smoke that obscures the view for traditional satellite and camera imagery (13). Thankfully, companies like Maxar Technologies use imagery from satellites equipped with SWIR bands (short wavelength infrared radiation). SWIR bands photograph radiation at 0.9 – 1.7 μm wavelength, a range that is not visible to human eyes (14).

“SWIR is able to penetrate through wildfire smoke to still capture the Earth’s surface, providing a detailed view of the ground, including burned areas and active fire zones” (15).

Maxar Technologies sources its imagery from the WorldView-3 satellite. The satellite is equipped with eight SWIR bands, the last of which is best suited to penetrate wildfire smoke (16). The SWIR bands also provide information about where the fire is burning the hottest. Information about the movement and boundaries of a wildfire can help government officials identify regions and infrastructure that might be in danger. This allows them to distribute and conserve resources and to plan evacuations with more precision. Additionally, Maxar’s labelled satellite map is publicly available online so that anxious civilians have access to information about their homes and communities.

Despite urgent calls from climate scientists, AI engineers have been slow at applying their technology to environmental issues. Now, geospatial AI is slowly being leveraged in California and elsewhere in addressing climate change and its consequences. Although the above applications of AI in fighting wildfires are new and it is difficult to measure their impact, they indicate a growing intersection of AI and environmental monitoring, a space that might one day be instrumental in slowing climate change.

Resources