Forest Fire Segmentation Metrics and Algorithms

In this paper, we present a comprehensive comparative analysis of algorithms for forest fire image segmentation. Specifically, the CNN-based BiSeNet, I2I-CNN, and PID-Net region segmentation algorithms were applied on Unmanned Aerial Vehicle (UAV) fire images for segmenting fire regions. To this end, we propose the implementation of novel region segmentation metrics tailored specifically for forest fire and segmentation applications. They are designed to augment the interpretability of CNN outputs without the need of ground truth fire images, furnishing end-users with enhanced insights into CNN model performance, beyond the one offered by conventional metrics, such as mean Intersection over Union (IoU). By offering additional contextual information regarding forest fires, these metrics empower stakeholders to make more informed decisions regarding the utilization and deployment of CNN models in disaster management contexts. Experimentation on forest fire image data basis suggests that the PID-Net is the most appropriate algorithm for this task.