Initial Results Continued - Fire Behaviours

 

Extreme fires can drive a plume of hot, turbulent air and smoke high into the atmosphere. The plume cools as it rises and can form a thunderstorm - a pyrocumulonimbus cloud- that is capable of igniting new fires down-wind of the original site. On the ground, these extreme fires can accelerate into a rapidly-spreading and very dangerous firestorm.

The behaviours team investigated the formation of pyrocumulonimbus clouds seen in Himawari-8 data of the 2020 Australian fire season. They developed a software pipeline to retrieve, re-format and re-project all 16 bands of the data into analysis-ready data products. They developed custom methods to distinguish between smoke haze and regular cloud, resulting in a time-series of well-segmented images. By applying dimensionality-reduction and clustering algorithms they identified the key bands containing information on smoke from extreme fires and pyrocumulonimbus formation during the Gospers Mountain Fire. They discovered tantalising hints that it might be possible to detect these events further in advance of what is currently possible.

The behaviour team were extremely well supported by a team of talented undergraduate students from Northwest Nazarene University in the USA. The students built an ID3 decision tree that also identified the most important bands for detecting extreme fire behaviour, complementing the results of the behaviour team.

Summary animation from the Behaviour Team. The top row shows the region of the Gospers Mountain Fire as seen by the Himawari-8 satellite, with smoke haze colour-coded in red. The middle-left panel shows the band-7 thermal infrared image (blue = cool…

Summary animation from the Behaviour Team. The top row shows the region of the Gospers Mountain Fire as seen by the Himawari-8 satellite, with smoke haze colour-coded in red. The middle-left panel shows the band-7 thermal infrared image (blue = cool, red = warm), which is often used to detect fire hotspots. The team ran a principal component analysis on the pixel values of all 16 bands and the three most significant components are plotted in the middle-right panel. The data clearly show an evolution over time. In particular, the green points are selected from a region containing a known pyrocumulonimbus event (green box at top-right) and many of these pixels show different spectral behaviour to their neighbours as the event progresses. The bottom row shows a complementary view of the same data: colour-colour diagrams between bands 7, 8, 12 and 14.

 
Emeline Paat-Dahlstrom