Climate driven dynamic fuel maps in wildfire management under climate change: an AI approach

Climate change has markedly increased the intensity and frequency of wildfires, emphasizing the need for predictive tools to inform adaptive management and mitigation strategies. This study presents a dynamic framework for assessing wildfire susceptibility, focusing on Southeastern Europe, a region particularly vulnerable due to diverse topographical and climatic conditions. By integrating machine learning (ML) with historical wildfire records and climate projections, the framework provides high-resolution susceptibility and fuel maps essential for informed decision-making.

The methodology incorporates data from the European Forest Fire Information System (EFFIS), CORINE land cover, and the Inter-Sectoral Impact Model Intercomparison Project (ISIMIP) climate projections. Climatic variables such as precipitation, wind speed, maximum and average daily temperatures, and consecutive dry/wet days were included in the predisposing factors of wildfire occurrence. These were combined with topographical and land cover information to train a supranational machine learning model capable of mapping annual wildfire susceptibility at a 100-meter resolution. The use of ISIMIP dataset (2008-2019) ensures using coherent datasets for historical and future time periods, allowing for dynamic projections under multiple climate scenarios (SSP126, SSP245, SSP585).

The susceptibility maps highlight regions where climatic and environmental conditions have historically facilitated wildfire occurrences. Susceptibility data were integrated with vegetation classifications, producing detailed wildfire hazard maps (or fuel maps). These maps categorize terrain into 12 classes based on a contingency matrix of susceptibility levels and fuel types, combining potential fire behavior in a worst-case scenario and its likelihood. As a sample case, areas classified as high susceptibility combined with coniferous forest cover represent hotspots where mitigation efforts should be concentrated. The possibility to generate future projected fuel maps leads to estimate the areas where wildfire hazard increases the most.

This study provides actionable insights for stakeholders by identifying critical zones for fuel management, ignition prevention, and adaptive planning. The dynamic nature of the model also allows for periodic updates as new data become available, ensuring its relevance under evolving climatic conditions. It establishes a foundation for risk assessment methodologies and potentially enables the estimation of annual losses and their temporal evolution in the next decades. This framework not only advances the scientific understanding of wildfire susceptibility but also supports practical applications in disaster risk reduction and land-use planning.

Keywords: Wildfire susceptibility, hazard mapping, machine learning, climate change, fuel type dynamics