Submeter-Level Global-Scale Road Extraction Based on Limited Labeled Data From Optical Remote Sensing Images

Accurate and regularly updated global road maps have many applications in intelligent navigation and urban planning. Deep learning algorithms have recently shown promising road extraction results using various Earth observation data. However, limitations in sample annotation and domain discrepancies between different regions and sensors pose challenges to the applicability of existing road extraction methods for large-scale tasks. In this paper, we propose a novel GlobalRoadMapper scheme to extract global-scale roads at the submeter level from optical remote sensing images that only require limited labeled samples. The proposed GlobalRoadMapper is designed as a two-stage method integrating Supervised Domain Incremental and Unsupervised Domain Adaptation. The supervised domain incremental stage learns road features from the source domain. To deal with catastrophic forgetting when deep models learn new knowledge from new domains, a strategy that couples pre-domain replay and mean-teacher is proposed. The unsupervised domain adaptation stage expands knowledge learned from the source domain to unseen scenes. Multi-level domain adaptation is proposed at the image, feature, and prediction levels in the second stage to address the issue of weak cross-domain generalization. The effectiveness of GlobalRoadMapper was tested at 43 sites worldwide. Qualitative and quantitative results demonstrate that GlobalRoadMapper outperforms the existing methods for global-scale road extraction tasks. Furthermore, city-scale sub-meter road mapping was conducted in six cities from different regions worldwide. GlobalRoadMapper achieved visual results comparable to the manually annotated OpenStreetMap road networks. Overall, GlobalRoadMapper holds great potential for large-scale road extraction and can be adapted to create easily updatable road maps globally.