Data-Driven Fire Modeling: Learning First Arrival Times and Model Parameters with Neural Networks

TL;DR

This study addresses learning fire-spread dynamics from simulations by jointly modeling the forward map $F:\mathbb{R}^p \to \mathbb{R}^N$ and the inverse map $G:\mathbb{R}^N \to \mathbb{R}^p$, where $p=5$ and $N$ is the number of raster cells. It compares three forward-network architectures—image-based CNN, image-based U-Net, and a scalar-input FC-UNet—and finds FC-UNet offers the best efficiency-accuracy trade-off, with U-Net producing sharper first-arrival fronts. For the inverse problem, a CNN-FC network achieves about a $10\%$ average relative error in estimating the five parameters, albeit with sensitivity to burn pattern and edge effects. The results demonstrate the potential of data-driven approaches to augment physics-based fire models, while highlighting dataset size, overfitting, and scenario-dependence as key considerations for reliable application to real data.

Abstract

Data-driven techniques are being increasingly applied to complement physics-based models in fire science. However, the lack of sufficiently large datasets continues to hinder the application of certain machine learning techniques. In this paper, we use simulated data to investigate the ability of neural networks to parameterize dynamics in fire science. In particular, we investigate neural networks that map five key parameters in fire spread to the first arrival time, and the corresponding inverse problem. By using simulated data, we are able to characterize the error, the required dataset size, and the convergence properties of these neural networks. For the inverse problem, we quantify the network's sensitivity in estimating each of the key parameters. The findings demonstrate the potential of machine learning in fire science, highlight the challenges associated with limited dataset sizes, and quantify the sensitivity of neural networks to estimate key parameters governing fire spread dynamics.

Figures (17)

• Figure 1: The computational domain represents a 200 m $\times$ 200 m burn unit, and it is discretized into cells that are 1 m $\times$ 1 m. The ignition pattern (red line) is a straight line with varying angles $\theta$ relative to the northerly wind direction. An angle of $\theta = 0$ is a head fire, while an angle of $\theta = \frac{\pi}{2}$ is a flank fire. To minimize edge effects, a 20 m region around the fuels (green) is initially burnt out (black).
• Figure 2: The objective of the neural networks being applied to the forward problem is to learn and predict the first arrival time based on key input parameters. Three networks are described in Sections \ref{['sec:fwd-CNN']}--\ref{['sec:fwdparam']}.
• Figure 3: The image-based CNN used for the forward problem. The network includes three convolution layers (red), three max pooling layers (blue), one fully-connected layer (grey), and one transpose convolution layer (orange). The size of the output and the number of channels of each layer are reported.
• Figure 4: (a) The average RMSE over the test set versus the number of epochs for the CNN. Initially, the RMSE quickly decreases (see inset) and then it decreases at a much slower rate and plateaus around $35.8$ s. A (b) simulated and (c) learned first arrival time from the test set. The ignition pattern is a straight line through the middle of the black region, and the other parameters are a background wind speed of $5.08$ m/s, sink strength of $0.51$ s$^{-1}$, flame out time of $18$ s, and diffusive ignition probability of $0.63$.
• Figure 5: The image-based U-Net used for the forward problem. Each blue box corresponds to a multi-channel feature map. The number of channels is denoted on top of the boxes. The size is provided at the left side of each row. The light blue boxes represent copied feature maps. The arrows denote the operations defined in the legend.