A Spatio-Temporal Neural Network Forecasting Approach for Emulation of Firefront Models

Andrew Bolt, Carolyn Huston, Petra Kuhnert, Joel Janek Dabrowski, James Hilton, Conrad Sanderson

Abstract: Computational simulations of wildfire spread typically employ empirical rate-of-spread calculations under various conditions (such as terrain, fuel type, weather). Small perturbations in conditions can often lead to significant changes in fire spread (such as speed and direction), necessitating a computationally expensive large set of simulations to quantify uncertainty. Model emulation seeks alternative representations of physical models using machine learning, aiming to provide more efficient and/or simplified surrogate models. We propose a dedicated spatio-temporal neural network based framework for model emulation, able to capture the complex behaviour of fire spread models. The proposed approach can approximate forecasts at fine spatial and temporal resolutions that are often challenging for neural network based approaches. Furthermore, the proposed approach is robust even with small training sets, due to novel data augmentation methods. Empirical experiments show good agreement between simulated and emulated firefronts, with an average Jaccard score of 0.76.

Introduction

Predicting wildfire propagation is a critical, time-sensitive task. Traditional physics-based simulators, such as those utilizing the level-set method, offer high accuracy but are computationally intensive. This latency limits the ability of fire managers to run large ensembles—hundreds of simulations with varying parameters—to account for uncertainties in wind, weather, and fuel conditions.

Machine learning (ML) surrogates, or emulators, offer a potential solution by providing rapid approximations of these complex models. However, existing ML approaches often struggle with maintaining high spatial resolution, capturing hour-by-hour temporal dynamics, and performing effectively with the limited datasets common in wildfire science. This research presents a spatio-temporal neural network designed to address these challenges, providing high-resolution, time-stepping forecasts of firefront evolution.

Methodology

The proposed framework utilizes a modular architecture designed to handle high-resolution spatial data while maintaining computational efficiency. The system is divided into three primary components:

  1. The Autoencoder: This component compresses high-resolution fire states into a lower-dimensional latent representation. By operating in this compressed space, the model can process large spatial extents without the memory overhead typically associated with high-resolution grids.
  2. The Outer Network: This serves as a feature engineering layer, processing static spatial inputs (terrain gradients and fuel types) and dynamic meteorological data (wind, temperature, humidity). These features are scaled into a format compatible with the latent representation of the fire state.
  3. The Inner Network (Dynamics Engine): At the core is a specialized U-Net that predicts the evolution of the fire. Crucially, the model employs residual learning; rather than predicting the entire fire state from scratch, the U-Net predicts the change (the update) for the next time step. This skip-connection approach prevents the model from falling into a "lazy" state where it simply replicates the previous input.
The proposed spatio-temporal neural network architecture, featuring an autoencoder for spatial compression and a U-Net-based inner network for predicting state updates.

Data Augmentation and Training

To overcome data scarcity, the researchers implemented a novel "cropping" strategy. Instead of training on entire map extents, the model was trained on thousands of smaller windows cropped around the active fire perimeter. This approach effectively multiplied the 155 training simulations into a robust dataset. Additionally, a custom log-based Mean Absolute Error (MAE) loss function was introduced to ensure the model remains sensitive to errors in both small, emerging fires and large, established ones.

Experimental Setup

The model was trained and validated using data from the SPARK Simulator, specifically 195 simulated fires based on South Australian conditions. Input features included 30-meter resolution heightmaps, land classification maps, and meteorological time series polled every 30 minutes. The model also incorporated "forcing terms" such as drought and curing factors to account for long-term fuel moisture conditions.

Results and Analysis

The emulator demonstrated high fidelity to the physics-based simulations, achieving an average Jaccard score (Intersection over Union) of 0.76. When weighted by total burned area—prioritizing the accuracy of larger, more impactful fires—the score improved to 0.83.

Qualitative comparison between predicted and target firefronts, including a 'Predicted Minus Target' heatmap highlighting false positives (red) and false negatives (blue).

Qualitative analysis shows that the model successfully captures complex firefront behaviors, such as "fingering" patterns caused by shifting winds and rugged terrain. The "Predicted Minus Target" heatmaps reveal that while the model is highly accurate, it occasionally struggles with very small geographic features, such as minor water bodies, which it may occasionally "jump" over.

Conclusion

This spatio-temporal neural network framework provides a viable path toward real-time ensemble forecasting for wildfires. By combining spatial compression with residual dynamics, the model achieves high-resolution emulation at a fraction of the cost of traditional simulators. Future work will focus on mitigating compounding errors in long-term recursive forecasts and fine-tuning the model using real-world satellite and aerial observation data to transition from a simulator-emulator to a tactical field tool.

Summary updated

Abstract

Computational simulations of wildfire spread typically employ empirical rate-of-spread calculations under various conditions (such as terrain, fuel type, weather). Small perturbations in conditions can often lead to significant changes in fire spread (such as speed and direction), necessitating a computationally expensive large set of simulations to quantify uncertainty. Model emulation seeks alternative representations of physical models using machine learning, aiming to provide more efficient and/or simplified surrogate models. We propose a dedicated spatio-temporal neural network based framework for model emulation, able to capture the complex behaviour of fire spread models. The proposed approach can approximate forecasts at fine spatial and temporal resolutions that are often challenging for neural network based approaches. Furthermore, the proposed approach is robust even with small training sets, due to novel data augmentation methods. Empirical experiments show good agreement between simulated and emulated firefronts, with an average Jaccard score of 0.76.
BibTeX
@misc{bolt2022spatio,
  author = {Andrew Bolt and Carolyn Huston and Petra Kuhnert and Joel Janek Dabrowski and James Hilton and Conrad Sanderson},
  title = {{A Spatio-Temporal Neural Network Forecasting Approach for Emulation of Firefront Models}},
  year = {2022},
  eprint = {2206.08523},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2206.08523},
}
Figure