Predicting forest fire refugia using machine learning: The role of topography and microclimatic variables

Abstract

Forest fire refugia are areas within fire-prone landscapes that remain fire-free or experience lower fire frequency and severity. These refugia are crucial for biodiversity, supporting species, maintaining mature vegetation, and aiding post-fire recovery. They enhance forest resilience by preserving genetic diversity and facilitating regeneration.Understanding how topography influences fire behaviour is key to identifying and conserving refugia, informing forest management to protect them from logging and disturbances. Machine learning models using wind patterns and altitude can effectively predict refugia in localized areas, aiding conservation efforts and promoting ecosystem resilience in fire-prone regions. Forest fire refugia can be accurately predicted using aspect, surface wind direction and speed (derived from computational fluid dynamics), topographic roughness, and temperature in machine learning algorithms (Random Forest, XGBoost; 2 ensembles models) and K-Nearest Neighbour (all run with and without ADASYN over-sampling). Six iterations were run per algorithm to assess the impact of leaving variables out.Among these variables, aspect is the most influential across both feature importance (Random Forest) and feature gain (XGBoost), as it aligns fire refugia with the leeward slopes of prevailing fire winds. Surface wind speed and direction, and global irradiation are also key predictors, with significant drops in model accuracy when these features are excluded. Temperature and topographic roughness show context-dependent importance. Temperature was significant in XGBoost but diminished after ADASYN oversampling, while topographic roughness increased in importance when elevation was excluded. Elevation did not significantly enhance model performance, and its exclusion had minimal impact on predictive accuracy. Ensemble models consistently produced the most accurate results, although accuracy metrics across all experiments where high and averaged 0.96 (±0.2), indicating robust predictive performance. These findings highlight the importance of topographic and micro climatic variables in fire refugia prediction, with machine learning providing reliable forecasting frameworks.