Stanford’s AI can use satellite images to locate risk zones to help with the upcoming wildfire season

Over the past few years, the United States has been hit by devastating wildfires as global temperatures have risen and weather patterns have changed, making this natural phenomenon particularly unpredictable and serious. To help, Stanford researchers have found a way to track and predict dry high-risk areas using machine learning and satellite imagery.

Stanford's AI can use satellite images to locate risk zones to help with the upcoming wildfire season

The current method of testing the sensitivity of forests and thickets to wildfires is by manually collecting branches and leaves and testing their water content. This method is accurate and reliable, but obviously labor-intensive and difficult to scale. Fortunately, researchers have recently had other sources of data. ESA’s Sentinel and Land satellites have accumulated a large amount of images of the Earth’s surface, which, after careful analysis, could provide a second source for assessing wildfire risk.

This is not the first attempt to make such an observation from an orbital image, but previous work has relied heavily on “extremely specific location” visual measurements, which means that analytical methods vary widely from location to location and are difficult to scale. The advanced technology used by the Stanford team is the Sentinel satellite’s Synthetic Aperture Radar, which penetrates forest shadeand and images the surface below.

“One of our biggest breakthroughs was to study a newer set of satellites that used much longer wavelengths, which allowed the observations to be much more sensitive to moisture in the shade of forest trees and directly represented fuel moisture content,” Alexandra Konings, a Stanford ecologist and senior author of the paper, said in a press release.

Stanford's AI can use satellite images to locate risk zones to help with the upcoming wildfire season

The team “feedback” the new images, which have been collected regularly since 2016, along with the U.S. Forest Service’s manual measurements to a machine learning model. This allows the model to “learn” which specific features in the image are associated with ground-based measured data. They then tested the resulting artificial intelligence to make predictions based on old data. It is accurate, and the prediction of one of the most common biomes in the western United States and one of the most vulnerable to wildfires is accurate.

You can see the results of this project on this interactive map, showing the model’s drought predictions for different periods in the western United States. This is a validation for firefighters – but the same model, given the latest data, can predict the upcoming wildfire season, which could help authorities make more decisions about controlling combustion, hazardous areas and safety warnings.

The scientists’ findings were published in the journal Remote Sensing Of Environment.