A Bayesian-optimised network interdiction heuristic for firebreak placement under stochastic wildfire spread
6 affiliations
- 1Forest Science and Technology Centre of Catalonia (CTFC), Solsona, Spain; Department of Industrial Engineering, Facultad de Ingeniería, Universidad de Talca, Sede Curicó, Curicó, Chile; Corresponding author at: Forest Science and Technology Centre of Catalonia (CTFC), Solsona, Spain.
- 2Department of Industrial Engineering, Facultad de Ingeniería, Universidad de Talca, Sede Curicó, Curicó, Chile; Instituto Sistemas Complejos de Ingeniería (ISCI), Chile; Instituto de Ecología y Biodiversidad (IEB), Chile
- 3Forest Science and Technology Centre of Catalonia (CTFC), Solsona, Spain
- 4JRU CTFC-AGROTECNIO, Ctra. Sant Llorenç de Morunys, Km 2, 25280 Solsona (Lleida), Spain
- 5Instituto Sistemas Complejos de Ingeniería (ISCI), Chile; Industrial Engineering Department, University of Chile, Santiago, Chile
- 6Departamento de Industria, Facultad de Ingeniería, Universidad Tecnológica Metropolitana, Santiago, Chile
Abstract
Optimal firebreak placement is a key component of preventive wildfire risk management, enabling the design of fire-resilient landscapes that mitigate expected wildfire damage. However, balancing solution quality and computational tractability at operational scales remains challenging, as the problem requires minimising expected losses over spatially explicit fire scenarios, yielding a large-scale stochastic combinatorial optimisation problem with complex spatial interactions. We formulate the firebreak placement problem as a stochastic network interdiction problem, where a decision-maker seeks to disrupt fire propagation by removing edges from a graph constructed from ensembles of simulated wildfire scenarios. We propose the Dynamic Critical Propagation Index (DCPI), a novel hybrid edge-ranking metric that integrates simulation-derived fire behaviour characteristics with dynamically updated edge betweenness centrality, capturing both empirical propagation risk and topological criticality. This metric is embedded within a beam search heuristic that efficiently explores the combinatorial solution space, while Bayesian optimisation is used to calibrate the trade-off between fire behaviour and topological information in a data-driven manner. Computational experiments on three landscape instances from Alberta, Canada (400 to 10,000 cells) show that DCPI-guided interdiction yields a lower expected burned area than two state-of-the-art methods on the instances where they could be compared, with relative improvements of 10%–12% on the 400-cell instance and 33%–39% on the 1600-cell instance. The experiments also show that the value of the fire-behaviour component and of the beam-search look-ahead depends on scale: on the 10,000-cell instance, a purely topological greedy rule performs best, indicating that at landscape scale firebreak effectiveness under random ignitions is driven mainly by compartmentalising the fuel network. Run times range from seconds to a few hours on a desktop server, which is compatible with annual or multi-year strategic planning. Our results highlight the potential of combining network interdiction, simulation-based optimisation, and data-driven calibration for effective wildfire management.
Related
- High-Molecular-Weight Genomic DNA Extraction from Recalcitrant Australian Plants: An Optimised CTAB Protocol for Anigozanthos
- Optimised fluorescence-activated nuclei sorting for epigenomic analysis of cortical cell types
- CellPatch: a flexible and efficient framework for single-cell foundation models empowered by heuristic gene patching
- Coverage Geometry and Heuristic Discovery Times in Protein Sequence Space
- Mice adapt their learning rate to stochasticity and volatility through a simple heuristic
BibTeX
@article{ulloafierro2026bayesianoptimised,
title = {{A Bayesian-optimised network interdiction heuristic for firebreak placement under stochastic wildfire spread}},
author = {Felipe Ulloa-Fierro and Eduardo Álvarez-Miranda and José Salgado-Rojas and José Ramón González-Olabarria and Jordi Garcia-Gonzalo and Camilo Matus-Olivares and Andrés Weintraub and Jaime Carrasco-Barra},
journal = {Ecological Informatics},
year = {2026},
month = oct,
doi = {10.1016/j.ecoinf.2026.104104},
url = {https://doi.org/10.1016/j.ecoinf.2026.104104},
note = {CC BY},
}
RIS
TY - JOUR
TI - A Bayesian-optimised network interdiction heuristic for firebreak placement under stochastic wildfire spread
AU - Felipe Ulloa-Fierro
AU - Eduardo Álvarez-Miranda
AU - José Salgado-Rojas
AU - José Ramón González-Olabarria
AU - Jordi Garcia-Gonzalo
AU - Camilo Matus-Olivares
AU - Andrés Weintraub
AU - Jaime Carrasco-Barra
PY - 2026
DA - 2026/10/30
JO - Ecological Informatics
DO - 10.1016/j.ecoinf.2026.104104
UR - https://doi.org/10.1016/j.ecoinf.2026.104104
AB - Optimal firebreak placement is a key component of preventive wildfire risk management, enabling the design of fire-resilient landscapes that mitigate expected wildfire damage. However, balancing solution quality and computational tractability at operational scales remains challenging, as the problem requires minimising expected losses over spatially explicit fire scenarios, yielding a large-scale stochastic combinatorial optimisation problem with complex spatial interactions. We formulate the firebreak placement problem as a stochastic network interdiction problem, where a decision-maker seeks to disrupt fire propagation by removing edges from a graph constructed from ensembles of simulated wildfire scenarios. We propose the Dynamic Critical Propagation Index (DCPI), a novel hybrid edge-ranking metric that integrates simulation-derived fire behaviour characteristics with dynamically updated edge betweenness centrality, capturing both empirical propagation risk and topological criticality. This metric is embedded within a beam search heuristic that efficiently explores the combinatorial solution space, while Bayesian optimisation is used to calibrate the trade-off between fire behaviour and topological information in a data-driven manner. Computational experiments on three landscape instances from Alberta, Canada (400 to 10,000 cells) show that DCPI-guided interdiction yields a lower expected burned area than two state-of-the-art methods on the instances where they could be compared, with relative improvements of 10%–12% on the 400-cell instance and 33%–39% on the 1600-cell instance. The experiments also show that the value of the fire-behaviour component and of the beam-search look-ahead depends on scale: on the 10,000-cell instance, a purely topological greedy rule performs best, indicating that at landscape scale firebreak effectiveness under random ignitions is driven mainly by compartmentalising the fuel network. Run times range from seconds to a few hours on a desktop server, which is compatible with annual or multi-year strategic planning. Our results highlight the potential of combining network interdiction, simulation-based optimisation, and data-driven calibration for effective wildfire management.
SN - 1574-9541
SN - 1878-0512
ER -