As fires get larger and more dangerous, various government and private agencies have turned to AI to detect, and potentially predict wildfires. The National Guard has been carrying out reconnaissance flights in California during the late summer and fall for the pew years, but now the drones used to carry out these flights have received upgrades with AI algorithms intended to automatically generate maps of fires within a particular region.
Creating fire maps is an incredibly difficult process that requires data analysis to map constantly changing fires as they move over rugged terrain. Both air and ground observations are used to make fire maps, and fire maps are typically only updated once every day or so. Large fires can move as far as 15 miles during a single day, as witnessed by some of the fires this fire season. Fire monitoring agencies need faster ways of gathering fire data and updating fire maps, and aerial drones combined with AI can address this need.
Fire-mapping systems, many of which rely on satellite data, typically utilize one of two different methods of detecting possible fires. The fires are detected either by sensing heat coming off the Earth’s surface (detecting unusually hot areas) or by analyzing aerosol emissions (detecting the smoke particles released into the air as biomass burns). After potential fires are detected, they can be confirmed through the use of high-resolution imaging systems like drones. The cameras that the National Guard drones are equipped with are capable of showing fires at a resolution of just 90 feet above the ground.
The National Guard has equipped its MQ-9 “Reaper” drones with AI algorithms intended to detect fires and generate fire maps. The AI algorithms are used to collect data on actively burning fires and detect “spot fires” that have been started by larger fires. The project was driven by the Joint Artificial Intelligence Center (JAIC), a division created by the Pentagon in 2018. The JAIC fire-mapping system utilizes machine learning algorithms trained on aerial footage of past fires with annotated boundaries. The algorithm can then take in unseen images with only location data and detect fires in them, outputting a map that shows which regions are burning. The location of spot fires is also marked.
Compared to the day-long fire map generation process used by other agencies, the JAIC fire-mapping system is much faster. The AI-powered fire mapping process can generate a new fire map roughly every half hour. According to the California Air National Guard, the maps produced by the new system are accurate and feedback from CalFire has been positive. If the maps continue to prove reliable and can be successfully integrated with CalFire’s operations, it could be deployed to help spot fires during next year’s fire season.
Beyond mapping the boundaries of current fires, AI can help fire-fighting teams predict the movement of fires. CalFire itself has recently begun working with a tool called WildFire Analyst Enterprise. The wildfire analysis tool was created by Technosylva, and it operates by combining various fire-spreading models together. These models are enhanced through the use of machine learning algorithms, trained on the features of past wildfires (such as the moisture content of vegetation, weather conditions, and satellite imagery). The model then compares data on a current fire to the past fire data in order to make predictions on how that fire might spread. The software also allows the user to create simulations based on how different variables like weather conditions change. The tool correctly predicted that the CZU Lightning Complex Fire would move toward the town of Felton, enabling firefighting crews to arrive early and save many structures that may not have been saved otherwise.
Meanwhile, firefighting departments in southern California are making use of a different fire tracking and prediction system called FireMap, developed by the Wifire Lab. FireMap uses both aerial and ground data from cameras, as well as climate conditions, wind conditions, the moisture content in vegetation, and more, to predict where fires will spread.
As more AI-driven fire detection and prediction platforms are created, drones will likely become increasingly important. Satellites are extremely useful, but they have limitations regarding the type and volume of data that they can collect. Two types of satellites are used to collect data: polar orbiter satellites and geosynchronous satellites. Polar orbiter satellites are capable of taking high-resolution images, but the images are only captured twice a day. In contrast, images gathered by geosynchronous images are gathered more frequently, typically every 5 minutes or so. However, geosynchronous satellites have to fly approximately 22,000 miles above the Earth’s surface in order to stay in sync with the Earth’s orbit. As a result, these images contain far less detail than the polar orbital satellites. Drones can help fill in the gaps in the data, getting more constant, detailed images of an area of interest.