Predicting wildfire burns from big geodata using deep learning

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