A machine learning approach for high-resolution fuel type mapping in Sardinia, Italy, using Sentinel-2 time series

Forest fires play a crucial role in shaping the Mediterranean biome while posing a significant threat to highly fire-prone Southern European countries. Factors such as prolonged dry periods and fuel accumulation increase the frequency and intensity of wildfires. Fire risk in the Mediterranean Basin is exacerbated by climate change and future climate projections highlight the need for advanced fire risk mapping, monitoring, and management strategies.

The EU-funded FirEUrisk project addresses these challenges through an integrated approach to wildfire risk monitoring, leveraging Earth Observation (EO) data and remote sensing techniques. EO data, particularly from satellites, can effectively monitor fire risk parameters over extensive areas in a resource-efficient manner. Fuel type and model mapping are essential for wildland fire risk monitoring and emergency management, yet existing global and continental-scale thematic maps lack sufficient spatial detail, particularly in regions with heterogeneous vegetation.

This study focuses on developing a methodology to classify fuel types in Sardinia, Italy, a fire-prone pilot site for the FirEUrisk project. Sardinia, the second-largest island in the Mediterranean Basin, experiences prolonged wildfire seasons triggered by human activities and sustained by intense droughts.

Fuel type classification was achieved using machine learning (ML) models trained on Sentinel-2 time series. Input datasets included digital terrain models, vegetation indices, and canopy height estimates. Training and testing samples were collected via an on purpose developed web application, enabling experts to label 10 m x 10 m pixels using orthophotos, Google Street View, and vegetation indices time series. The ML models were trained with 80% of the dataset and tested with 20% and performance metrics such as precision, recall, and F1-score were computed.

This study demonstrated the feasibility of producing high-resolution (10m) fuel type maps for Sardinia using Sentinel-2 time series. However, the classification task remains challenging due to the structural complexity of vegetation in Mediterranean regions, leading to diverse fire behaviours and impacts. Future improvements include additional training samples collection, validation of the resulting classification, and the integration of vertical vegetation structure data, such as RADAR or LiDAR.