Large-scale road extraction from high-resolution remote sensing images based on a weakly-supervised structural and orientational consistency constraint network

Abstract

Fully supervised road segmentation neural networks from remote sensing images rely on a large number of densely labeled road samples, limiting their potential in large-scale applications. In this paper, a weakly-supervised road segmentation network based on structural and orientational consistency (SOC-RoadNet) is proposed to learn road surface from open-source road maps, thus supporting high-quality large-scale road extraction. The open-source road maps can be made with scribble-labels to train the weakly-supervised SOC-RoadNet. However, these scribble-labels that represent the approximate location of road centerlines with only one-pixel width cannot provide supervision for road boundaries, leading to low road surface segmentation accuracy using scribble-supervised methods. Existing weakly-supervised road extraction networks deploy complex scribble-label propagation procedures to generate pseudo labels to train models and exploit image edge maps to refine road boundaries. Nevertheless, road edges in images are blurry due to occlusions and spectral-confusing backgrounds, limiting reliability when processing complex scenes. The proposed SOC-RoadNet is designed as a one-stage end-to-end architecture that directly learns road surface features from scribble-labels without extra label propagation procedures. Instead of using unreliable edge maps to regularize road boundaries, a structural consistency loss function is introduced to evaluate the structural similarity between road surface segmentation results and the input images for improved road boundary accuracy. Considering that roads are elongated targets with local directional consistency, orientation learning is integrated to SOC-RoadNet to improve road integrity. A road refining module was designed to aggregate the road surface features and road orientation features to further refine the road extraction results. The reference road orientation maps are generated from the scribble labels; thus, the results of orientation learning are supervised without extra input. Extensive experimental results on the public Deep Globe Road Dataset and four city-scale images confirm the effectiveness of our SOC-RoadNet in large-scale road extraction tasks.