[Paper Review] Deep graphical regression for jointly moderate and extreme Australian wildfires

Daniela Cisneros, J. Ian Richards+3 more|arXiv (Cornell University)|Aug 28, 2023

Fire effects on ecosystems10 citations

TL;DR

The paper develops a two-stage deep graphical regression model that jointly models the occurrence and the radial spread of Australian wildfires using graph convolutional neural networks and the extended generalized Pareto distribution on an irregular spatial domain.

ABSTRACT

Recent wildfires in Australia have led to considerable economic loss and property destruction, and there is increasing concern that climate change may exacerbate their intensity, duration, and frequency. Hazard quantification for extreme wildfires is an important component of wildfire management, as it facilitates efficient resource distribution, adverse effect mitigation, and recovery efforts. However, although extreme wildfires are typically the most impactful, both small and moderate fires can still be devastating to local communities and ecosystems. Therefore, it is imperative to develop robust statistical methods to reliably model the full distribution of wildfire spread. We do so for a novel dataset of Australian wildfires from 1999 to 2019, and analyse monthly spread over areas approximately corresponding to Statistical Areas Level~1 and~2 (SA1/SA2) regions. Given the complex nature of wildfire ignition and spread, we exploit recent advances in statistical deep learning and extreme value theory to construct a parametric regression model using graph convolutional neural networks and the extended generalized Pareto distribution, which allows us to model wildfire spread observed on an irregular spatial domain. We highlight the efficacy of our newly proposed model and perform a wildfire hazard assessment for Australia and population-dense communities, namely Tasmania, Sydney, Melbourne, and Perth.

Motivation & Objective

• Motivate robust hazard quantification for wildfires by modeling full distribution from moderate to extreme spread.
• Construct a novel irregular-spatial dataset of aggregated burnt areas (SA1/SA2) for Australia.
• Develop a hybrid model combining a GCNN-based occurrence predictor with an eGPD-based spread model.
• Provide continental hazard maps and assess hazards for population-dense communities.
• Identify key predictors driving wildfire frequency and severity across space and time.

Proposed method

• Model Y(s,t) as the square-root of monthly burnt area with a two-part distribution: occurrence probability p0(s,t) and conditional spread given occurrence F+(y).
• Use a deep logistic regression with a graph convolutional neural network to estimate logit p0(s,t).
• Model F+ with an extended generalized Pareto distribution eGPD(κ, σ, ξ), with σ(s,t) modeled as a function of predictors and σ(s,t) proportional to sqrt(area(s)).
• Represent spatial data as a graph with vertices as irregular regions and a weighted adjacency matrix A based on great-circle distance with a truncation Delta.
• Employ GCNN layers with skip connections to capture spatial dependencies, combining region-specific and neighbor information.
• Fit models via Adam optimization on negative log-likelihood, with temporal bootstrap for uncertainty and a robust cross-validated training/validation split.

Figure 1 : Maps of the monthly radial wildfire spread [ $\sqrt{\mbox{BA}}$ ; km] (top-left), polygon area [km 2 ] (top-right), monthly average air temperature [K] (bottom-left), and NDVI [unitless] (bottom-right) for January 2002. Grey regions in the top-left panel are Statistical Areas Level 1 with

Experimental results

Research questions

• RQ1 Can a joint model of wildfire occurrence and spread capture the full distribution from moderate to extreme burn areas on irregular spatial domains?
• RQ2 How well do GCNN-based predictors perform compared to dense neural networks when exploiting spatial structure in predictors?
• RQ3 What are the key covariates that drive both the frequency and severity of Australian wildfires across space and time?
• RQ4 How can an extended generalized Pareto-based framework be integrated with deep learning to model non-extreme and extreme wildfire spread?
• RQ5 What are the spatial trends and hazard characteristics in population-dense Australian communities using the proposed framework?

Key findings

• The framework enables joint modeling of zero versus positive burnt areas and the distribution of positive values across space and time.
• The eGPD-based component provides a flexible model for wildfire spread that covers both bulk and tail behaviors with asymptotically justified upper tails.
• GCNNs on irregular spatial graphs improve fit over standard dense networks by exploiting spatial structure in covariates.
• Hazard maps for Australia and major population centers (e.g., Sydney, Melbourne, Perth, Tasmania) reveal spatial and temporal trends in wildfire frequency and severity.
• The approach identifies important predictors that drive both occurrence and spread of wildfires across the domain.

Figure 2 : Top panels: region-wise $5\%$ (left) and $95\%$ (right) quantiles of the monthly radial wildfire spread [ $\sqrt{\mbox{BA}}$ ; km] pooled across the observation period. Grey regions are Statistical Areas Level 1 within the Northern Territory where the response, BA, is not available. Botto