A temporal perspective on the reliability of wildfire hazard assessment based on machine learning and remote sensing data

Wildfires pose a significant natural hazard, particularly in Mediterranean regions where they cause lasting and often irreversible damage. Despite ongoing prevention efforts, large-scale fires continue to devastate vast areas of forest and agricultural land, a situation exacerbated by climate change and human activities. Proactive approaches are essential to mitigate this phenomenon, requiring reliable and robust susceptibility models. This study aims to evaluate the reliability of wildfire susceptibility models by incorporating a temporal perspective into the assessment of machine learning models. Unlike traditional approaches that focus solely on spatial validation, this research integrates temporal analysis to assess the accuracy of future predictions. Focusing on the province of Jijel, Algeria, the study employs six well established machine learning algorithms-K-Nearest Neighbors (KNN), Histogram-Based Gradient Boosting (HGB), Random Forest (RF), Support Vector Machine (SVM), Multi-Layer Perceptron (MLP), and Logistic Regression (LR)-to assess the wildfire exposure. The findings reveal that temporal sampling in wildfire susceptibility models decreases performance by 2-20% compared to spatial sampling, indicating the significant influence of temporal aspect on model reliability and overestimation due to data partitioning based solely on spatial sampling. Additionally, the study highlights a trade-off between performance and reliability: while HGB emerges as the best-performing model, its reliability is relatively low. In contrast, LR, despite being the least performant model, demonstrates the highest reliability and consistency across both spatial and temporal evaluations.