Robust Prediction of Wildfire Spread in Australia

Wildfires can have devastating effects on urban infrastructure and natural ecosystems, making wildfire management an important, but yet complex and difficult task. The systematic collection of data, increased computing power, and the development of physical models made it possible to get an understanding of the dynamics of wildfire spread. As exact computational simulations of wildfires are not feasible yet, several subtasks such as the estimation of the spread rate were analyzed with various methods in the literature. In this paper, different types of predictive models are evaluated to forecast the spread of wildfires on a daily and weekly basis in a comparative study. These models are tested on real-world data of wildfires from the seven Australian regions New South Wales, Northern Territory, Queensland, South Australia, Tasmania, Victoria, and Western Australia from 2005 to 2020, including weather, vegetation, and land cover class data, in a univariate and multivariate setting. Furthermore, relevant features are identified and discussed which can have an important influence on wildfire spread. We find that robust models, which are less sensitive to outliers, capture the dynamics of wildfire spread most accurately.