Initial Results Continued - Fuels II

Fuels II Team:

The Fuels II Team focused their attention on predicting the area burnt by fire spreading from a particular ignition site. This is a difficult problem as the spread of a fire is influenced by a myriad of factors: weather conditions (especially wind), slope and aspects of terrain, fuel-load in the landscape and fuel moisture levels. The current best fire spread predictions come from physics-based models that can require significant time and computing power to run. However, machine learning models can potentially predict fire spread in just a few seconds on laptop hardware. Imagine fire chiefs being able to make predictions on-site instantly, with the latest information. 

The machine learning model created by the team is trained on historic multispectral satellite images from Sentinel-2, maps of historic fire extent, high-resolution elevation data and three different weather variables. This data feeds into a semantic segmentation model that has successfully predicted the fire scar of an unseen test region. The model also outputs a ‘burnability’ map encoding the risk of fire in the landscape. The team anticipates that the model could be used as a tool to prioritise deployment of fire teams to the most at-risk areas.

This three panel figure shows initial results of the segmentation model created by the Fuel II team. On the left is the Sentinel-2 visible-light image of a region west of Sydney prior to the 2019/2020 fire season. The middle panel shows the same map overlaid by the burn probability predicted by the model. Areas of high flammability are coloured bright orange and low flammability purple or black. The actual burn scar can be seen in the right panel as a distinctive brown area spreading over the image, which was taken after the fire passed through.

Next post - Results from the Fire Behaviour Team.