[Paper Review] Spatiotemporal wildfire modeling through point processes with moderate and extreme marks

Jonathan Koh, François Pimont+2 more|arXiv (Cornell University)|May 17, 2021

Fire effects on ecosystems4 citations

TL;DR

This paper proposes a novel Bayesian hierarchical spatiotemporal point process model that jointly models wildfire ignition intensity and burnt area size—distinguishing moderate from extreme fires—using shared random effects to improve estimation efficiency and predictive accuracy. The model integrates extreme-value theory with log-Gaussian Cox processes and captures nonlinear, seasonal covariate effects, revealing regional disparities in fire risk across the French Mediterranean basin (1995–2018).

ABSTRACT

Accurate spatiotemporal modeling of conditions leading to moderate and large wildfires provides better understanding of mechanisms driving fire-prone ecosystems and improves risk management. We here develop a joint model for the occurrence intensity and the wildfire size distribution by combining extreme-value theory and point processes within a novel Bayesian hierarchical model, and use it to study daily summer wildfire data for the French Mediterranean basin during 1995--2018. The occurrence component models wildfire ignitions as a spatiotemporal log-Gaussian Cox process. Burnt areas are numerical marks attached to points and are considered as extreme if they exceed a high threshold. The size component is a two-component mixture varying in space and time that jointly models moderate and extreme fires. We capture non-linear influence of covariates (Fire Weather Index, forest cover) through component-specific smooth functions, which may vary with season. We propose estimating shared random effects between model components to reveal and interpret common drivers of different aspects of wildfire activity. This leads to increased parsimony and reduced estimation uncertainty with better predictions. Specific stratified subsampling of zero counts is implemented to cope with large observation vectors. We compare and validate models through predictive scores and visual diagnostics. Our methodology provides a holistic approach to explaining and predicting the drivers of wildfire activity and associated uncertainties.

Motivation & Objective

• To develop a unified statistical framework that models both the spatiotemporal occurrence of wildfires and their size distribution, particularly distinguishing moderate from extreme events.
• To improve predictive accuracy and reduce estimation uncertainty by sharing latent spatial random effects between the occurrence and size components of the model.
• To capture nonlinear and seasonal influences of covariates such as the Fire Weather Index and forest cover on wildfire activity.
• To address data sparsity and zero-inflation in large wildfire datasets through stratified subsampling of zero counts.
• To provide actionable insights for wildfire risk management by identifying common drivers of ignition and extreme fire size across regions.

Proposed method

• Models wildfire ignitions as a spatiotemporal log-Gaussian Cox process to capture complex spatial and temporal dependence in ignition patterns.
• Assigns burnt area as a continuous numerical mark to each ignition point, with extreme fires defined as those exceeding a high threshold (79 ha).
• Uses a two-component finite mixture model for fire sizes that separately captures moderate and extreme fire distributions, with component-specific smooth functions of covariates.
• Incorporates shared random effects between the occurrence and size components to borrow strength across model parts and reduce uncertainty.
• Applies a stratified subsampling strategy to manage large datasets with abundant zero counts, improving computational efficiency.
• Employs full Bayesian inference via MCMC with hierarchical priors to estimate model parameters and quantify uncertainty.

Experimental results

Research questions

• RQ1 How can we jointly model the spatiotemporal occurrence of wildfires and their size distribution, particularly distinguishing moderate from extreme events?
• RQ2 What is the impact of shared random effects on estimation uncertainty and predictive performance in a joint wildfire modeling framework?
• RQ3 How do covariates such as the Fire Weather Index and forest cover influence wildfire occurrence and size distribution in a nonlinear and seasonally varying manner?
• RQ4 To what extent do regional differences in landscape structure (e.g., urbanization, fragmentation) affect the likelihood of extreme fires?
• RQ5 Can a Bayesian hierarchical model with marked point processes and extreme-value theory improve risk prediction and driver identification compared to separate models?

Key findings

• The inclusion of shared random effects significantly reduces estimation uncertainty and improves predictive performance, especially in regions with sparse extreme fire data.
• Corsica exhibits a highly positive COX-BETA sharing effect, indicating that moderately large fires are more likely to become extreme there, likely due to delayed fire suppression and large contiguous forest areas.
• Landscape fragmentation and human activity in densely populated or rural areas increase ignition frequency but reduce the likelihood of large fires due to fuel discontinuity.
• The model reveals strong nonlinear and seasonal effects of the Fire Weather Index on wildfire risk, suggesting that current fire danger rating systems require careful interpretation.
• Threshold exceedance probabilities for extreme fires vary significantly across regions, with higher probabilities in large, contiguous forested areas and lower ones in fragmented or urbanized landscapes.
• Without shared effects, the extreme fire size component would suffer from wide credible intervals due to data sparsity, limiting practical utility.