What question did this study set out to answer?

This research aims to enhance wildfire preparedness frameworks by integrating environmental and structural factors to predict building damage.

May 16, 2026Open Access

A graph neural network and decision tree modeling approach for predicting wildfire-induced building damage

Key Points

This research aims to enhance wildfire preparedness frameworks by integrating environmental and structural factors to predict building damage.
Developed GraphFire-X, a three-tiered ensemble framework for wildfire risk assessment.
Utilized a Graph Neural Network for environmental contagion mapping based on community interactions.
Employed XGBoost for analyzing structural resilience at asset-level in multiple wildfire events across varying geographies.
The GNN shows strong transferability in predicting wildfire impacts, even in less data-rich areas.
The structural analysis identifies key vulnerabilities, such as enclosed eaves, within communities.
Findings aid in directing resources effectively for community-scale fuel reduction and structural improvements.

Abstract

As wildfires increasingly evolve into urban conflagrations, there is an urgent need for robust preparedness frameworks at the building scale. Currently, a significant resolution mismatch exists, where large-scale environmental data has not been fully leveraged to inform the small-scale, deterministic realities of building ignition. Furthermore, traditional models often treat structures as isolated observations, overlooking the non-linear contagion dynamics that drive community-scale destruction. This research addresses these gaps by establishing GraphFire-X, a three-tiered ensemble framework that disentangles wildfire risk into environmental contagion and structural fragility. The framework integrates two specialized predictive streams: a Tier 1 environmental specialist and a Tier 2 structural specialist. The environmental specialist is implemented as a Graph Neural Network (GNN) that operationalizes the community as a directed contagion graph enriched with high-dimensional Google AlphaEarth Foundation embeddings. This tier explores three distinct configurations to capture complex spread mechanisms via physics-informed convection, radiation, and ember probabilities: static edge, feature driven edge prediction, and physics-regularized GNN. The Tier 2 structural specialist utilizes XGBoost to isolate granular asset-level resilience. This complementary design ensures that the Tier 3 ensemble stacking model captures the full complexity of wildfire risk. The framework was applied to six wildfire events across California, Colorado, and Canada. The results demonstrate the GNN has strong transferability potential, even in data-scarce regions. Additionally, the structural specialist efficiently isolates critical micro-scale vulnerabilities, such as enclosed eaves. This diagnostic triage identifies whether vulnerability stems from landscape exposure or structural flaws. By distinguishing these risk, it supports decision-makers in targeting resources, such as community-scale fuel reduction or structural hardening, across diverse fire regimes.

Read Full Paperexternally

Bookmark

View Full Paper

Cite This Study

Esparza et al. (Fri,) studied this question.

synapsesocial.com/papers/6a0809f1a487c87a6a40bcb5 https://doi.org/https://doi.org/10.1016/j.cacaie.2026.100085

Bookmark

View Full Paper