CAIG: Advancing Wildfire Science, Prediction, and Management with Machine Learning

As the impacts of wildfires on humans and ecosystems continue to escalate, there is an increasing need for researchers and managers to understand and predict wildfire behavior. Traditional physics-based fire spread models are often limited by the high computational cost of running these models and the challenge of systematically verifying model predictions.

Machine learning (ML) advances provide an opportunity to gain greater insight into wildfire processes and improve forecasts of fire activity. However, researchers currently lack large-scale, open-access data on which to train and test ML models for improved wildfire prediction. In this project, a team of researchers at the University of California Irvine and Pyregence, a consortium of researchers and software engineers advancing scientific knowledge of wildfires and building next-generation forecasting tools, will assemble a new dataset of fire observations that combines weather, topography, and fuel information with observations of sub-daily fire spread.

The team will use this dataset to create new ML models that more accurately predict fire spread and the placement of fuel breaks in complex landscapes. These modeling advances will provide the scientific foundation for developing next-generation fire spread models used by wildfire managers, helping them limit fire damage to ecosystems and communities.

The project team will also host a summer school on ML where diverse early career scientists from across the country will gain hands-on experience in computational methods. The team will develop a new theme for this course on fire prediction and integrate perspectives from fire managers.

This project will enhance understanding, prediction, and management of wildfires by addressing the following three objectives: 1) develop a large new public dataset of fire-related environmental observations to support large-scale ML and reproducible research on wildfire spread modeling, 2) advance innovative spatiotemporal ML models for understanding and predicting wildfire spread and systematically compare these ML models to parameter-optimized physics-based models, and 3) develop novel network-based frameworks for optimizing the placement of fuel treatments that appropriately characterize uncertainty and risk. Together, the optimized physics-based and ML fire spread models will be used with graph theory to structure the optimal size, shape, and placement of fuel breaks.

In a set of hypothetical scenarios, the research team will re-run model simulations of known fires in the fire database but include different levels of fuel treatment within each domain. The proposed machine learning models offer a promising avenue for improving wildfire forecasting and mitigation strategies.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.