Follow Datanami:
January 10, 2020

Explaining the Australian Bushfires with Deep Learning

For months, Australia has been ravaged by bushfires. These massive fires, fueled by lengthy droughts and high temperatures, have caught the world’s attention as lives have been lost and entire towns have been evacuated. But even as the firefighters push back against the flames, another kind of flame war has emerged: the debate over whether or not changes in the world’s climate are responsible for the devastating scale of the bushfires. As it turns out, deep learning techniques may have already provided some crucial answers.

A few years ago, the seemingly increasing frequency of Australian bushfires had already caught the attention of a group of Tasmanian researchers. “Understanding such climatic change for Australian bushfire is limited and there is an urgent need of scientific research which is capable enough to contribute to Australian society,” they wrote. However, experimental or observational studies were expensive and difficult due to practicality concerns, as well as health and safety regulation – and even when those kinds of studies were conducted, they could only hope for “limited generalization.”

Instead, the researchers turned to deep learning techniques. They integrated over six years’ worth of satellite-derived NASA fire data, burned area data and data from the Australian Water Availability Project (AWAP). Together, these provided 26 different variables ranging from temperature and rainfall to soil moisture and surface runoff. The researchers integrated these variables using a training process for an unsupervised ensemble of deep neural networks, then fed the resulting weighted data into a supervised ensemble classification phase.

After all of this, the researchers were able to produce a map of predicted bushfire hotspots across Australia, coded according to the most probable level of fire intensity. This map was also divided into climatic zones. When validating the ensemble model, they found that the estimation accuracy was 91% with a false discovery rate of just 6%. Further, they found that bushfire frequency was largely dependent on weekly trends in soil moisture, solar irradiation, dry fuel and wind speed. 

Crucially, they also landed on another conclusion.

“Estimated average bushfire event frequency over the whole continent of Australia was 3,284 per week in 2007,” they wrote, “in comparison with the frequency of 4,595 events per week in 2013. This shows that weekly bushfire frequencies … have increased by 40% since 2007. In particular, a major increase in bushfire frequencies has been recorded … since June 2011 [indicating] a major climatic shift.”

Apart from its immediate implications for firefighting and climate research, the researchers hope that this research will lead the way for deep learning becoming a useful tool for Australian issues. “To the best of our knowledge,” they wrote, “this is the first attempt to create an ensemble of deep learning and applied to solve a very topical and complex climatic problem of continental Australia.”

About the research

The research discussed in this article was published in Volume 3, Issue 2 of Royal Society Open Science as “Big data integration shows Australian bush-fire frequency is increasing significantly.” The paper, which can be accessed here, was written by Ritaban Dutta, Aruneema Das and Jagannath Aryal.

Datanami