Deep-learning-based wildfire detection and transferability verification

This study focuses on the critical task of estimating wildfire-burned areas, utilizing remote sensing via cubesats to obtain high-resolution imagery for analysis. Using PlanetScope imagery and surface reflectance from visible and NIR bands, along with vegetation indices, the research mapped burned areas in two locations. A U-Net model based on CNN architecture was employed for estimations, with the data divided into training, testing, and validation sets. Model performance was assessed using metrics like IoU and F1-score, demonstrating the usefulness of vegetation indices combined with visible imagery for detecting burned areas, despite challenges from false alarms in snow and rocky terrains. The study also examined the model's transferability across different regions, indicating its adaptability despite variations in accuracy due to local conditions. This underscores the model's potential for broad geographic application in wildfire analysis.