Journal Paper

Wildfires pose significant threats to ecosystems worldwide, emphasizing the urgent need for rapid firefighting responses. It is well recognized that firefighters encounter substantial challenges when combating advanced-stage fires, underscoring the critical importance of early detection and suppression. This study proposes a novel approach employing deep learning algorithms to automate the detection of rural roads from aerial imagery. A manually annotated dataset was constructed from orthorectified aerial images of wildfire-prone areas, aiming to realistically capture the dynamic characteristics of rural road networks. The proposed system comprises a two-step process: a segmentation model to identify rural roads, followed by post-processing techniques designed to refine the segmentation results. Four deep learning architectures—VGG16 U-Net, VGG19 U-Net, ResNet50 U-Net, and MobileNet U-Net—were evaluated with extensive hyperparameter tuning, both before and after data augmentation. In the post-processing phase, two methods were implemented: one to remove small connected components, and another to reconnect segmented roads disrupted by occlusions caused by trees and shadows. Data augmentation yielded significant performance improvements across the models, which were further enhanced by post-processing. The highest performance achieved included a precision of 85.1%, a recall of 81.9%, and a Jaccard index of 71.7%, with comparable results observed for the other architectures. While post-processing generally contributed to improved outcomes, the percentage gains in evaluation metrics were modest. Notably, MobileNet U-Net exhibited the most substantial improvements after post-processing, with increases of 1.5% in Precision, 0.4% in Recall, and 1.2% in the Jaccard index. Overall, the proposed system demonstrates considerable potential for the efficient creation of rural road networks to support wildfire management.