Abstract
As wildfires become more frequent and intense, it is essential to develop sophisticated techniques for precise detection and damage evaluation. This research examines a Double-Step Deep Learning Framework using several U-Net models, including MultiResUNet, to identify burned areas and estimate severity. Using satellite images, the study explores the effect of different severity levels within mask output, focusing on both 4 and 5 level severity classifications. Additionally, the Mask R-CNN model was evaluated independently for image segmentation, revealing challenges due to its reliance on pretrained weights and limited spectral input. The comparative analysis illustrates how changes in the granularity of severity intervals influence model performance, providing insights into the benefits of more nuanced severity segmentation for wildfire assessment. This approach has the potential to improve the precision of damage assessments and support more informed decision-making in the management and response of wildfires