Super Real-Time Forecast of Wildland Fire Spread by A Dual-Model Deep Learning Method

Abstract

Driven by climate change, more frequent and extreme wildfires have brought a greater threat to humans globally. Fastspreading wildfires endanger the safety of residents in the wildland-urban interface. To mitigate the hazards of wildfires and facilitate early evacuation, a rapid and accurate forecast of wildfire spread is critical in emergency response. This study proposes a novel dualmodel deep learning approach to achieve a super real-time forecast of 2-dimensional wildfire spread in different scenarios. The first model utilizes the U-Net technique to predict the burnt area up to 5 hours in advance. The second model incorporates ConvLSTM layers to refine the forecasted results based on real-time updated input data. To evaluate the effectiveness of this methodology, we applied it to Sunshine Island, Hong Kong, and generated a numerical database consisting of 210 cases (12,600 samples) to train the deep learning models. The simulated wildfire spread database has a fine resolution of 5 m and a time step of 5 minutes. Results show that both models achieve an overall agreement of over 90% between numerical simulation and AI forecast. The real-time wildfire forecasts by AI only take a few seconds, which is 102 ~ 104 times faster than direct simulations. Our findings demonstrate the potential of AI in offering fast and high-resolution forecasts of wildfire spread, and the novel contribution is to leverage two models which can work in tandem and be utilized at various stages of wildfire management.