A key question for fire managers is, ‘Can remote wildfires be detected sooner from orbit?’
Currently, fires are mostly ‘called in’ by observers on the ground. Satellite data is only used after the fire has grown and spread, but there is excellent potential to build an early-warning system using space-based platforms.
Scion rural fire researchers Ilze Pretorius and Kate Melnik teamed up with Alex Codoreanu and Jack White from Swinburne University of Technology in Melbourne to create an early fire detection method using geostationary satellite imagery, machine learning and image enhancing techniques borrowed from astrophysics.
Team members chose geostationary satellite data because it is the only type of satellite that rotates in-sync with the Earth and always has a view of the same landscape below. The satellite’s constant position means it can update imagery as frequently as every 10 minutes, making it ideal for an early fire detection system. However there are two major obstacles that needed to be overcome. Firstly, its low resolution imagery makes it hard to see fires in the early stages if they are smaller than the image pixel size (1 pixel = 4 km2). The second challenge is the possibility of false alarms, created by other hot fixtures in the landscape, such as large rocks.
By combining machine learning image enhancing techniques known as image stacking and super resolution, the team members were able to deal with these challenges. The method was successfully tested on data from the Orroral Valley fire that devastated Canberra in January 2020. (Check out the story on our website for more detail on how it works).
The team believe that their method will be able to detect fires sooner than current systems. This could help emergency services respond faster and more effectively to fires, better protecting people, property and biodiversity. There is also significant opportunity to apply this work in Scion’s own research developing real-time automated tools for fire detection, fire growth and smoke prediction tailored to New Zealand.
The development of this model was part of the inaugural Bushfire Data Quest. The Quest paired experts in machine learning with fire researchers, focusing on a single problem and trying to prove it is solvable using AI and machine learning. Researchers spent one intense week (and several weeks of preparation) working together, and then presented their results online. Tara Strand was also involved in the quest’s science advisory panel. Tara’s role was to advise a team looking to predict when a wildfire suddenly spreads rapidly – a key element needed in fire behaviour prediction.
You can also watch Ilze, Kate, Jack and Alex present their findings online here (begins at 58min 33sec).