Comparing Machine Learning-Based Remote Sensing for Fuel Type Mapping: Case Studies in Portugal, And Greece

Accurate fuel mapping is vital for wildfire risk assessment and management. This study combines Remote Sensing (RS) data and Machine Learning (ML) to differentiate fire behavior fuel models. Three ML approaches - Random Forest (RF), Support Vector Machine (SVM), and Convolutional Neural Network (CNN) - are compared in terms of accuracy, recall, and F1 score. Employing Sentinel-2 imagery, the ML-based classification accurately categorizes fuel types into eight main classes: broadleaf, conifer, shrub, grass, bare soil, burned area, urban area, and water. Past results in a Sardinia test case were promising, with CNN achieving impressive metrics - accuracy, recall, and F1 score - each at 99%. Notably, the network exhibits high validation score in identifying classes in unseen pixels: broadleaf at 99%, conifer at 79%, shrub at 76%, and grass at 84%. Subclasses, aligned with the Standard Scott and Burgan (2005) system, were derived from the eight classes using Above Ground Biomass (AGB) and Bio-Climatic (BC) maps to refine fuel mapping. A significant enhancement of this work involves testing the proposed method in Portugal and Greece to validate its robustness across diverse geographical regions. Furthermore, the final fuel type maps are rigorously validated by comparing them with FirEUrisk’s pre-existing validated fuel maps, providing a benchmark to assess the accuracy and reliability of the new maps.