Boundary Loss 기반 경계 정보 강화를 통한 산악지역 도로 추출 정확도 개선

As South Korea has a high proportion of mountains in its territory, landslides caused by heavy rains often occur in the summer. Therefore, the damaged roads in mountainous areas caused by landslides are reported every year. In the event of a disaster, road extraction from satellite imagery can provide information to identify the location and extent of the damaged area rapidly and accurately and establish a recovery policy. Recently, deep learning approaches have been widely used for road extraction. However, there are some challenges, particularly in recognizing unclear boundaries. In this study, we applied the boundary loss, which considers boundary misalignment, to improve the accuracy of road boundaries in mountainous areas with the DeepLabv3+ network. We defined various combinations of loss functions by integrating boundary loss with binary cross entropy (BCE) loss and intersection over union (IoU) loss to consider both road surfaces and boundaries. Afterward, the model was trained separately using each defined loss function, and their performance was compared. As a result, the models trained with the loss functions that included boundary loss tended to exhibit higher accuracy than the others. The highest accuracy was achieved when IoU loss and boundary loss were combined with weights of 0.8 and 0.2, respectively. Particularly, in areas where road boundaries were obscured by surrounding vegetation or curved road sections. However, there are some limitations in extracting roads where the distinction between road and non-road is visually unclear. In future studies, we aim to address this issue by robustly enhancing the network to train for road connectivity and complex geometries.