Wildfire dynamics from ECOSTRESS data and machine learning: The case of
South-Eastern Australia’s black summer
Abstract
In 2019–20 Australia was devastated by the worst wildfires observed in
decades. NASA’s ECOsystem Space-borne Thermal Radiometer Experiment
on Space Station (ECOSTRESS) mission, launched in 2018, captured many
dynamics of the fires at high resolution, including ecosystem stress
prior to the fires. We aimed to determine the predictive capacity of
ECOSTRESS observations for fire occurrence and intensity in Southeast
Australia. We found that ECOSTRESS data (evaporative stress index and
water use efficiency) were highly predictive of fire dynamics (25-65%
occurrence prediction accuracy for ESI; and, 40-95% occurrence
prediction for WUE > 1 gCkg-1H2O alone, depending
on their levels) with the ESI coefficient averaging approximately three
times stronger than general topographic variables or meteorological
variables. Our results, based on a logistic regression model, had an
overall predictive accuracy of 83%, suggesting high potential of using
ECOSTRESS data to project and examine fires in Australia and other
similar regions of the world.