Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

TL;DR

The paper tackles the computational burden of forecasting wildfire spread and the need for uncertainty quantification by proposing a spatio-temporal neural emulator that mirrors firefront dynamics at high resolution. It introduces a three-component architecture (autoencoder, outer feature encoder, inner dynamics module) that updates a latent fire state via a shallow U-Net and employs data augmentation to robustify training on small datasets. Empirical results on SPARK-simulated fires achieve an average Jaccard score of 0.76, demonstrating faithful replication of firefronts across varying resolutions and extents. The approach is flexible, scalable, and relevant for operational planning, with potential extensions to uncertainty quantification and adaptation to other geo-spatial spread phenomena.

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.

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

TL;DR

The paper tackles the computational burden of forecasting wildfire spread and the need for uncertainty quantification by proposing a spatio-temporal neural emulator that mirrors firefront dynamics at high resolution. It introduces a three-component architecture (autoencoder, outer feature encoder, inner dynamics module) that updates a latent fire state via a shallow U-Net and employs data augmentation to robustify training on small datasets. Empirical results on SPARK-simulated fires achieve an average Jaccard score of 0.76, demonstrating faithful replication of firefronts across varying resolutions and extents. The approach is flexible, scalable, and relevant for operational planning, with potential extensions to uncertainty quantification and adaptation to other geo-spatial spread phenomena.

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.
Paper Structure (11 sections, 1 equation, 5 figures, 1 table)

This paper contains 11 sections, 1 equation, 5 figures, 1 table.

Table of Contents

  1. Introduction
  2. Emulator Architecture
  3. Outer Network
  4. Autoencoder
  5. Inner Network
  6. Design Features
  7. Empirical Evaluation
  8. Dataset
  9. Metrics and Loss
  10. Evaluation
  11. Conclusion

Figures (5)

  • Figure 1: The proposed emulator architecture, comprising three main components. The autoencoder (blue) encodes and decodes the fire input state. The outer component (red) incorporates the autoencoder, as well as encoding spatial, forcing, and weather features. The inner component (orange) handles the dynamics of the emulator. The latent fire state is updated using information from the spatial and forcing layers, as well as weather feature inputs. These layers are concatenated and passed through a shallow U-Net structure. The sum of the U-Net output and the input latent fire state produce a new latent fire state estimate.
  • Figure 2: Evolution of firefront contours for a trial, shown over four panels (left-to-right, top-to-bottom). Emulator (red), simulation (blue) and ignition point (green) are overlaid over land classes. Dominant land classes are grassland (yellow), mallee-heath shrubland (orange), and water (blue). The wind initially drives the fire south, before turning west. Map size is 46.1 km $\times$ 46.1 km, 30 meter resolution.
  • Figure 3: The difference between predicted and target fire arrival times (measured in 30 minute intervals) for the same test sample as illustrated in Fig. \ref{['fig:quad']}. Positive values (red) indicate false-positives while negative values (blue) represent false-negatives. The Jaccard score for this trial is 0.81.
  • Figure 4: Evolution of firefront contours for a trial, shown over four panels (left-to-right, top-to-bottom). Emulator (red), simulation (blue) and ignition point (green) are overlaid over land classes. Dominant land classes are grassland (yellow), mallee-heath shrubland (orange), and water (blue). The wind initially drives the fire south-east, before turning north. Map size is 46.1 km $\times$ 38.4 km, 30 meter resolution.
  • Figure 5: The difference between predicted and target fire arrival times (measured in 30 minute intervals) for the same test sample as illustrated in Fig. \ref{['fig:quad_b']}. Positive values (red) indicate false-positives while negative values (blue) represent false-negatives. The Jaccard score for this trial is 0.90.