Home | Wildfires | Contact Info

Wildfire Surrogate Modeling

Overview

As part of my MSc in ACSE at Imperial College London, I undertook a research project to develop a foundational deep learning surrogate model for wildfire spread. The model was based on the SCALED model developed by Yueyan Li at Imperial College London, a UNet-backed DDIM diffusion model. Following the methodology of Burge et al. (2023), data was generated using the empirical fire spread model, ELMFIRE.

The model was trained to predict the next state of the wildfire spread given environmental conditions. The visualization above shows a GIF comparing the predicted wildfire spread with the ground truth through 12 hours of autoregressive rollout, at each step using the model's previous prediction as input for the next prediction.

Wildfire Simulation GIF

Region of Interest

10,000 ELMFIRE simulations were generated with real-world fuel model data for the purposes of training and evaluating the model's ability to generalize to unseen fuels and topography. The training set was generated from a region spanning parts of eastern California and western Nevada, USA, while the test set was generated from a region of western and northern California.

Work is ongoing to publish the datasets used for this project to a public repository for further use.

Ignition Areas for Real-World Fuel Model Dataset

Future Work

Work is ongoing to improve the model's accuracy and robustness, including generating more comprehensive datasets and refining the loss function.