Fuel Type Mapping from Sentinel-2 Multispectral Imagery Using a Convolutional Neural Network (CNN) Classification Framework: A Case Study on Mudge Island, British Columbia

The final output is a 28-class fuel map that aligned with the Scott and Burgan (2005) Fire Behavior Fuel Model (FBFM). Model performance differed across evaluation approaches. Internal validation against a high-confidence reference layer indicated robust classification performance, achieving an overall accuracy of 86.9%, a mean producer's accuracy of 87.9%, a mean user's accuracy of 86.9%, and an F1 score (a balance between precision and recall) of 87.3%.

In contrast, external validation using four independent datasets showed lower levels of agreement, with most accuracy metrics falling below 70%. Comparing the model with the baseline fuel map revealed significant differences: conifer fuels achieved an accuracy rate of 97.9%, while the accuracy rate for other classes ranged from 0% to 18.2%. Despite these discrepancies, the framework demonstrated the potential of multispectral data and deep learning for scalable fuel mapping, offering invaluable insights to inform community-scale, small-island wildfire risk assessment, fuel treatment prioritization, and fire management planning.