MCMCNet: A Semi-Supervised Road Extraction Network for High-Resolution Remote Sensing Images via Multiple Consistency and Multitask Constraints

Influenced by deep learning, extracting roads from high-resolution remote sensing images has attracted extensive attention. However, most previous works have focused on fully supervised models relying on large amounts of annotated data and have not considered the characteristics of narrow and elongated roads. In order to alleviate the model’s dependency on labeled data, reduce annotation workload, and fully exploit road characteristics, we proposed a semi-supervised road extraction network via multiple consistency and multitask constraints (MCMCNet) that utilizes only minimal labeled data, while exploiting unlabeled data through the mining of pseudo-label information for constraint. Moreover, to ensure the generation of more accurate pseudo-labels, we incorporated a guided contrastive learning module (GCLM) into the model to increase interclass discriminability and enhance consistency constraints. In addition, to ensure the continuity of road extraction and integrity of the main roads, we added a road skeleton (road centerline) prediction head (RSPH) in addition to the original road segmentation prediction head. Finally, we introduced an adaptive road augment module (ARAM) to enhance linear road features and avoid learning redundant information by the use of local and global information adapted to road features. Extensive experiments demonstrated that MCMCNet achieved a 3%–5% improvement in <inline-formula> <tex-math notation="LaTeX">$F1$ </tex-math></inline-formula> and IoU across three benchmark datasets, compared to other classical semi-supervised road extraction models, and the visualization results confirmed that MCMCNet partially addressed challenges including road occlusion, foreground-background high-similarity regions at extremely low label rates. The code is available at <uri>https://github.com/zhouyiqingzz/MCMCNet</uri>.