Abstract
Background
Although soil burn severity (SBS) is traditionally mapped post-fire, pre-fire predictive approaches remain limited despite their potential to identify areas at risk of soil degradation, runoff and erosion.
Aim
To develop a large-scale model for predicting post-fire SBS from pre-fire conditions at ~30 m spatial resolution.
Methods
A harmonized dataset of 62 wildfires (2018–2024) was used to build machine learning (ML) frameworks based on Random Forest, eXtreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM), using pre-fire vegetation, climate, soil–terrain and anthropogenic predictors. Performance was evaluated using five-fold cross-validation (CV), leave-one-fire-out (LOFO) validation and independent testing on 2025 fires.
Key Results
All models showed similar CV overall accuracy (OA ~0.58), but LOFO performance varied. LightGBM maintained more consistent OA (0.51–0.58) across fire size classes, whereas XGBoost and Random Forest exhibited comparable performance. Key predictors included vegetation biomass (Normalized Difference Moisture Index, Enhanced Vegetation Index), fuel stress (Evaporative Stress Index, Evapotranspiration), elevation and soil moisture. High SBS was associated with dry conditions and high biomass availability, with region-specific controls modulating these patterns.
Conclusion
ML models using pre-fire environmental data enable proactive SBS prediction.
Implications
This framework supports fuel management planning, potential post-fire hydrological impact assessment and the prediction of SBS from pre-fire conditions.