RoadSeg-CD: A Network With Connectivity Array and Direction Map for Road Extraction From SAR Images

Road extraction from synthetic aperture radar (SAR) images has attracted much attention in the field of remote sensing image processing. General road algorithms, affected by shadows buildings and trees, are prone to producing fragmented segments. To improve accuracy completeness extraction, we propose a neural network-based algorithm which takes connectivity direction features roads into consideration, named RoadSeg-CD. It consists two branches: one is main branch for segmentation; other auxiliary learning directions. In branch, array designed utilize local contextual information construct loss based on predicted probabilities neighboring pixels. proposed novel map, used directions roads. The branches connected specific feature fusion process, output taken as result. Experiments real implemented validate effectiveness our method. experimental results demonstrate that method can obtain more continuous complete than several state-of-the-art algorithms.