Initial Results Continued - Early Detection

The first reported sign of fire is often a column of smoke rising above the distant landscape. It takes time to triangulate the location of the fire and more time to mobilise a fire team to the blaze. Infrared images from the Japanese Himiwari-8 satellite show bushfires as bright pixels or ‘hotspots’. However, the high geostationary orbit of the satellite results in a relatively low resolution image with pixels corresponding to a kilometre on a side.

The team applied two algorithms to Himawari-8 data, building a novel detection pipeline that they applied to the 2020 Orroral Valley Fire, near Canberra. The first was an image pre-processing method borrowed from astronomy - using image stacking to create and subtract a suitable cloud-free reference image. The ignition of the Orroral Valley Fire was detected at a level greater than 30 times the background variation. The second algorithm was a machine learning technique called 'super-resolution’. This allows the reconstruction of a higher resolution image from lower resolution imagery by ‘baking in’ information on likely structures being observed. The caveat here is that the model must be trained on the type of terrain it is being used to enhance. Investigations are ongoing but show great promise.

Top: The Japanese Meteorological Agency Himawari-8 satellite provides ‘full disk’ imagery of the Earth every 10 minutes. The Early Detection team applied an image stacking technique commonly used in astronomy to create a high signal-to-noise infrared image of the Orroral Valley Fire near Canberra. Bottom-left: The curves show the signal levels across pixel-rows of the reference-subtracted image, with row-16 bisecting the hotspot. Bottom-right: A super-resolved detection image. This is a machine-learning technique that can produce high-resolution images by including extra information on the subject being observed.

Next post - Results from the Fuels II Team.