WFNet: A hierarchical convolutional neural network for wildfire spread prediction

Abstract

Pattern analysis in wildfire spread behaviors is crucial for rescue actions and disaster reduction. Deep learning methods have the potential to model the wildfire spread despite problems such as continuous time prediction and multimodal environmental encoding. Therefore, we present a novel hierarchical convolutional neural network (CNN) denoted as WFNet to model the spread pattern of wildfires. The core of WFNet is defining the spread spatiotemporal distribution field (SSDF) to describe the process of wildfire spread, enabling global optimization and end-to-end prediction. Then, a hierarchical State-Condition mechanism is implemented to progressively and efficiently encode high-order features pertaining to multimodal elements. The experimental results demonstrate that WFNet has a competitive performance to existing models in computation time and model accuracy. More interestingly, WFNet shows excellent robustness when input fire state is in an uncertain moment, enabling investigators to quickly backward the ignition from the fire perimeter.