Using Echo State Networks to Inform Physical Models for Fire Front Propagation

TL;DR

This paper introduces a hybrid model combining echo state networks with level set methods to improve wildfire front prediction, offering efficient, data-driven forecasts with uncertainty quantification, demonstrated on real California megafire data.

Contribution

The novel hybrid model integrates echo state networks with physical level set models for wildfire spread prediction, enabling data-driven learning and uncertainty quantification.

Findings

Accurate wildfire front forecasting demonstrated on real data.

Model provides calibrated uncertainty estimates.

Efficient computation suitable for real-time wildfire management.

Abstract

Wildfires can be devastating, causing significant damage to property, ecosystem disruption, and loss of life. Forecasting the evolution of wildfire boundaries is essential to real-time wildfire management. To this end, substantial attention in the wildifre literature has focused on the level set method, which effectively represents complicated boundaries and their change over time. Nevertheless, most of these approaches rely on a heavily-parameterized formulas for spread and fail to account for the uncertainty in the forecast. The rapid evolution of large wildfires and inhomogeneous environmental conditions across the domain of interest (e.g., varying land cover, fire-induced winds) give rise to a need for a model that enables efficient data-driven learning of fire spread and allows uncertainty quantification. Here, we present a novel hybrid model that nests an echo state network to learn nonlinear spatio-temporal evolving velocities (speed in the normal direction) within a physically-based level set model framework. This model is computationally efficient and includes calibrated uncertainty quantification. We show the forecasting performance of our model with simulations and two real data sets - the Haybress and Thomas megafires that started in California (USA) in 2017.