The raging wildfires occurring all over the world have caused enormous economic damage and loss of life. Knowing when and where a large-scale fire can occur in advance can improve fire prevention and resource allocation. However, available forecasting systems provide only limited information. Furthermore, they do not provide long enough lead times for useful regional details.
Scientists have now applied a deep learning algorithm to enhance prediction of the danger of wildfires in the western United States. Researchers from South Korea and the United States have developed a hybrid method that combines artificial intelligence techniques and weather forecasts to produce improved forecasts of severe fire danger for one week at finer scales (4 km x 4 km accuracy), increasing its usefulness in fire suppression and management.
“We have experimented with several approaches to integrate machine learning with traditional weather forecast models to improve wildfire risk predictions. This study is a huge step forward because it demonstrates the potential of such an effort to enhance fire risk prediction without the need for additional computing power,” Says lead author Dr. Rackhun Son, a recent Ph.D. from the Gwangju Institute of Science and Technology (GIST) in South Korea, who is currently at the Max Planck Institute for Biochemistry in Germany. “Fire hazard predictions can be further improved using continuous development in both Earth system models and recent AI developments,” He adds.
While data-driven AI methods have shown excellent capabilities to infer things, explaining why and how to come to conclusions remains a challenge. This has led to AI being classified as a black box. “But when artificial intelligence was combined with computer models based on physical principles, we could diagnose what was going on inside this black box,” Says co-author Professor Simon Wang from Utah State University, USA. “AI-based predictions of extreme levels of fire risk are well based on strong winds and specific geographic characteristics, including high mountains and valleys in Western United States that has traditionally been difficult to solve with coarser models. “
Computational efficiency is another major advantage of this method. Conventional methods for predicting fire hazards with precise spatial resolution, a process called “regional minimization,” are often computationally demanding, expensive and time-consuming. “Although comparable computational resources were required in the development phase, once the training task for the AI was completed, i.e. it was done once initially, it only took a few seconds to use this component with the weather forecasting model to produce forecasts for the rest of the season.” says co-author Professor Kyu Sun Lim at Kyungpook National University, Korea. Therefore, the newly developed AI-based method with the ability to make high-accuracy accurate predictions in a shorter time was more cost-effective than traditional forecasting systems.
“In this study, AI is only being tested to predict fire hazard in the western United States. In the future, it could be applied to other types of extreme weather events or in other parts of the world,” Co-author Dr Philip J. Rush from Pacific Northwest National Laboratory and University of Washington. “The flexibility of our AI approach can help predict any weather-related feature.”