Wildfires are a growing problem worldwide and cause significant devastation to communities. This year alone, there have been catastrophic wildfires in Greece, Maui, and Canada. The reasons behind the increase in wildfires are complex, including factors such as climate change and land development. However, technology, particularly artificial intelligence (AI), can play a role in addressing these challenges.
Google Research has been investing in various climate adaptation efforts, including the use of machine learning (ML) for wildfire prevention and providing information during these events. One example is the wildfire boundary tracker, which uses ML models and satellite imagery to map large fires in near real-time with updates every 15 minutes.
To advance their research efforts, Google Research has partnered with the US Forest Service (USFS) to improve fire modeling tools and fire spread prediction algorithms. They are using ML to significantly reduce computation times, allowing the model to be employed in near real-time. The new model can also incorporate localized fuel characteristics, such as fuel type and distribution, in its predictions.
Currently, the most widely used fire behavior models are based on the Rothermel fire model developed in the 1970s. These models consider factors like wind, terrain slope, moisture level, and fuel load. While they have been widely used in fire management, there are limitations, such as the simplification of underlying physical processes and the need for expert adjustments.
To overcome these limitations, USFS researchers have been working on a new model that improves the physical fidelity of fire behavior prediction. However, the computational cost and inference time make it impractical for real-time applications. This is where ML comes in.
Google Research and the USFS have collaborated to apply ML to decrease computation times for fire models. They used deep learning techniques to train a model that can capture complex fire behavior. This model has significantly lower computational costs compared to the original physics-based models and can provide fire spread estimates much faster.
In addition to training, the new model is also useful for operational fire prediction. It can capture detailed fire behavior changes due to differences in fuel structures. To fully leverage this capability, high-resolution satellite imagery and geo-information are being integrated into ML models to allow for fuel-specific mapping at scale.
Moving forward, there are still many challenges in fire management that can benefit from more accurate fire models. These models can consider 3D flow interactions, fluid dynamics, thermodynamics, and combustion physics. Google Research has developed a high-fidelity 3D fire simulator that can be run on their TPUs (Tensor Processing Units). This simulator can be used to generate insights on extreme fire development scenarios and build ML classification models.
In conclusion, AI and ML have the potential to revolutionize wildfire prevention and management. By leveraging these technologies, researchers and agencies can better predict fire behavior, map fire boundaries, and make informed decisions to protect communities from the devastating effects of wildfires. Continued collaboration between technology companies like Google and government agencies like the USFS is crucial for advancing these efforts.