Development of an Integrated Static Fire Risk Index for Cyprus Utilizing Tree-Based Ensemble Classifiers: A Soft-Voting Approach

Accurate wildfire risk assessment is essential for disaster mitigation and landscape management, particularly in Mediterranean ecosystems. A number of wildfire risk maps for Cyprus use expert-driven indices, single-model statistical methods, and data from remote sensing. However, there is currently no standardized, high-precision Fire Risk Index (FRI) that comprehensively considers multiple risk factors and provides accurate, consistent predictions across different areas. This study introduces an innovative multi-stage machine learning framework designed to develop a comprehensive Static Fire Risk Index (FRI) for Cyprus. The methodology consists of two primary phases: the creation of four thematic sub-indices and their subsequent integration through an ensemble meta-modeling approach. More specifically, a topographic risk index was derived from derivatives of an EU Digital Elevation Model (DEM) (25 m spatial resolution), namely slope, elevation, aspect, plane curvature, and classification of landforms. A vegetation-moisture risk index was generated using multi-temporal satellite imagery from Landsat 8 and 9 to calculate the Leaf Area Index (LAI), Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), and Normalized Difference Moisture Index (NDMI). Fuel flammability index was assessed using a comprehensive vegetation type map, while an anthropogenic risk index included factors such as population density, proximity to roads, transmitter stations, picnic sites, power lines, and built-up regions to address human-induced fire risks. The historical fire location data from 2015 to 2024 were extracted from VIIRS sensors to facilitate the development of machine learning models. Initially, four thematic fire risk indices were generated: Fuel Flammability, Vegetation Moisture, Topography, and Anthropogenic Risk. These indices were subsequently standardized into five ordinal fire danger classes, ranging from 1 (Very Low) to 5 (Very High).

To determine the most effective integration strategy, eight distinct machine learning architectures were benchmarked: Random Forest (RF), XGBoost, LightGBM (LGBM), Decision Trees (DT), Support Vector Machines (SVM), Artificial Neural Networks (ANN), K-Nearest Neighbors (KNN), and Logistic Regression (LR). Model bias and uncertainty were assessed using cross-validation with historical fire occurrences, along with an examination of prediction residuals and spatial error patterns. The performance evaluation, which focused on accuracy (83%) and Area Under the Curve (AUC) (0.87), revealed that tree-based ensemble models (RF, XGBoost, LGBM, and DT) significantly outperformed both baseline and kernel-based algorithms. Consequently, these four top-performing models were chosen for the final fusion stage.

A "Soft Voting" ensemble method was used to combine the predictions of the chosen models. This approach involved pixel-wise averaging of fire occurrence probabilities, which effectively minimized individual model bias and improved spatial stability. The resulting continuous probability map was then reclassified into five distinct threat classes using the Jenks Natural Breaks optimization method. Validation against historical fire data demonstrated that this consensus-based methodology provides superior predictive reliability in comparison to single-algorithm models. The final Fire Risk Index (FRI) map acts as a high-resolution decision-support tool, allowing fire management authorities to prioritize resources in high-vulnerability zones through a mathematically robust and standardized classification system.