Predicting wildfire burns from big geodata using deep learning

• Published in: Safety science Vol. 140; p. 105276
• Main Authors: Bergado, John Ray, Persello, Claudio, Reinke, Karin, Stein, Alfred
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier Ltd 01.08.2021; Elsevier BV
• Subjects: Big geodata; Burns; Convolutional networks; Crop fields; Datasets; Deep learning; Feature extraction; Forest & brush fires; Ground-based observation; Input statistical importance; Land cover; Land surface temperature; Land use; Lightning flashes; Multilayer perceptrons; Power lines; Regression coefficients; Remote sensing; Remote sensors; Risk management; Spatial data; Statistical analysis; Wildfire burn prediction; Wildfires; Wind direction; Woodlands; Deep learning; Big geodata; Convolutional networks; Wildfire burn prediction; Input statistical importance
• ISSN: 0925-7535; 1879-1042
• DOI: 10.1016/j.ssci.2021.105276

•Convolutional network for predicting daily maps of the probability of a wildfire burn.•Convolutional networks demonstrate higher predictive accuracy and map quality.•Exploratory feature statistical importance metrics improves model transparency. Wildfire continues to be a major environmental problem in the world. To help land and fire management agencies manage and mitigate wildfire-related risks, we need to develop tools for mapping those risks. Big geodata—in the form of remotely sensed images, ground-based sensor observations, and topographical datasets—can help us characterize the dynamics of wildfire related events. In this study, we design a deep fully convolutional network, called AllConvNet, to produce daily maps of the probability of a wildfire burn over the next 7 days. We applied it to burns in Victoria, Australia for the period of 2006–2017. Fifteen factors that were extracted from six different datasets and resulted into 29 quantitative features, were selected as input to the network. We compared it with three baseline methods: SegNet, multilayer perceptron, and logistic regression. AllConvNet outperforms the other three baseline methods in four of the six quantitative metrics considered. AllConvNet and SegNet provide smoother and more regularized predicted maps, with SegNet providing greater sensitivity in dificriminating less wildfire-prone locations. Input feature statistical importance was measured for all the networks and compared against logistic regression coefficients. Total precipitation, lightning flash density, and land surface temperature occur to be consistently highly weighted by all models while terrain aspect components, wind direction components, certain land cover classes (such as crop field and woodland), and distance from power lines are ranked on the lower end. We conclude that wild-fire burn prediction methods based on deep learning present quantitative and qualitative gains.