Enterprise AI Analysis

Model Performance Overview

The study rigorously evaluated four advanced models: Random Forest (RF), XGBoost, Deep Neural Network (DNN), and Kolmogorov-Arnold Networks (KAN). The performance was assessed using key metrics like Accuracy, F1-Score, PR-AUC, and ROC-AUC on a held-out test set.

Random Forest consistently demonstrated the highest predictive performance, achieving an impressive 90.7% ROC-AUC and 82.7% accuracy, making it the most reliable model for regional wildfire prediction in this study. XGBoost also performed very strongly, closely following RF. While DNN and KAN showed competitive performance, they exhibited slightly higher false-positive rates, indicating a less conservative classification boundary in complex environments.

Key Drivers of Wildfire Susceptibility

A SHAP (Shapley Additive Explanations) analysis was conducted to identify the most influential factors contributing to wildfire occurrence. This method provides clear insights into how each environmental and anthropogenic variable impacts the model's predictions.

The analysis conclusively identifies Global Horizontal Irradiation as the single most influential factor across all models, underscoring the critical role of solar energy input in drying vegetation and promoting fire spread. Elevation and Distance from Settlements rank as the second and third most important factors, respectively, highlighting the combined influence of topography and human activity. Precipitation and air temperature also play significant roles, while land use (rangelands and forests) and distance from water bodies have moderate to lower importance.

Regional Wildfire Risk Distribution

The ensemble model, integrating insights from all four algorithms, provides a balanced spatial distribution of wildfire susceptibility across Southeastern Europe. This map reveals distinct regional patterns influenced by diverse climatic, geomorphological, and land-use conditions.

Greece exhibits the highest overall wildfire susceptibility, with 10.5% of its territory in the 'very high' category, reflecting its Mediterranean climate. Albania and Montenegro also show significant vulnerability. Conversely, Slovenia and Moldova are identified as the least fire-prone countries, with very low percentages in high susceptibility classes due to cooler, more humid continental climates. This spatial gradient, increasing from northwest to southeast, underscores the diverse wildfire challenges across the region.

Unique Contributions of This Study

This research stands out by overcoming limitations in previous studies and offering several distinct advantages for enterprise application in wildfire management:

• First regional spatial susceptibility prediction for 11 countries using medium spatial resolution data (100 m).
• Creation of a large geospatial database containing 28,952 reliable fire events and 11 quantitative variables, serving as a foundation for other natural hazards.
• Comprehensive comparative analysis of both machine learning (RF, XGBoost) and deep learning (DNN, KAN) model performances.
• In-depth identification of influential factors contributing to fire occurrence through SHAP analysis.
• Detailed assessment of wildfire susceptibility for each country individually, offering granular insights.
• Counter-intuitive finding: wildfire susceptibility decreases with an increase in forested areas, attributed to cooler, more humid microclimates within dense forests.

These unique aspects provide a robust, data-driven foundation for developing targeted and effective wildfire protection strategies across Southeastern Europe.