Abstract: Extracting roads from satellite images through semantic segmentation algorithm has become the mainstream solution for RS-based road monitoring tasks. However, due to complex features and changeable textures of roads in satellite imagery which derive from various geographical environments and the high cost of pixel-level road labeling, it is unaffordable to acquire a substantial dataset with pixel-level road annotation to train semantic segmentation models. To solve the above problems, a stagewise weakly supervised road extraction algorithm based on road centerline scribbles is proposed. The feature of road centerline scribbles is learned in a weakly supervised way, and the road segmentation model is trained by stages. In addition, the pseudo mask update strategy and the hybrid training strategy are proposed, and the loss functions for road foreground and road background are designed. The results show that compared with other weak supervision methods based on road centerline, the proposed algorithm achieves superior performance in road segmentation task. Ablation studies are also conducted to verify the effectiveness of the proposed training strategy.

Key words: road extraction, image segmentation, weakly supervised learning, remote sensing imagery