Kincade Fire, seen by NOAA’s GOES West on Oct. 24, 2019, covering 10,000 acres
By Marc Kodack
In the western United States, climate change is increasing the risk of wildfires, such as the ongoing Kincade Fire that is devastating northern California. This is in part because increased temperatures driven by greenhouse gas emissions are drying out available fuels for fires. As we’re seeing today, when ignition occurs under these conditions, fires can become extraordinarily challenging for local communities and first responders to manage.
While it’s critically important to reduce current climate change trajectories so as to reduce the incidence and severity of wildfires, we’re likely to see an increase in such fires for the foreseeable future. In this context, Artificial Intelligence (AI) might help fight those fires. Enter a research effort by the Department of Defense (doD) Joint Artificial Intelligence Center and the California Air National Guard. The effort involves testing machine learning – a branch of AI based on the concept that artificial systems can learn from data – as a tool to delineate a fire’s perimeter.
Currently, full motion video of a fire is manually captured. It is then shared with multiple analysts who determine where the fire perimeter is located using geospatial mapping tools. The process can take hours, all the while a dynamic fire is moving and shifting. By some measures, “wildfires can spread at rates of 7-10 miles per hour in forested regions, and even more quickly in grasslands.” By introducing automated fire tracking support via machine learning, this process could potentially be improved significantly, through “near-real time updates of the fire perimeter,” increased accuracy of its’ location, and better information for first responders and local communities.
The machine learning model being utilized by this research partnership was constructed using annotated short videos (150 frames) of wildfires from around the U.S. Approximately 400,000 video frames were included in the data set, of which 100,000 frames had an active fire in the frame. The computer evaluates each frame using an infrared sensor because “the fire perimeter would rarely be visible in the standard RGB [red, green, blue] spectrum due to smoke and other factors.” The computer’s task is to find burning or burnt areas. The goal of the model is to balance speed and accuracy. People can then take the model’s output and use it in decision-making.
After all the raw data are analyzed, the model runs at roughly 20 frames per second with a 92 % accuracy for the perimeter’s location. Thus, live video that is captured by an aircrew can be analyzed almost instantaneously by others on-the ground. Such an improvement could be significant for both civilian populations, and the military – including in cases where wildfires affect one or more military installations simultaneously.
DoD and National Guard physical assets and personnel are critically important partners in regional wildfire fighting. Projects like this one can further enhance their effectiveness, and help save lives.
