GLTDNet: Cross-Domain Road Extraction Through Collaborative Optimization of Global-Local Feature Enhancement and Topological Decoupling

In the realm of cross-sensor and cross-resolution applications, remote sensing-based road extraction across diverse domains frequently encounters hurdles such as undetected road segments, erroneous identifications, and distortions in topological representation. To tackle these challenges, this study proposes an innovative cross-domain road extraction methodology designated as GLTDNet. This approach leverages a hybrid CNN-Transformer architecture and incorporates a Global-Local Feature Enhancement Unit (GLFEU) designed to effectively capture both the intricate local detail features and the overarching global topological structures of roads. Specifically, to enhance the extraction of domain-invariant road features from a topological perspective, we utilize road topology characterization to guide the alignment of features between the source and target domains during the domain adaptation procedure. Additionally, the methodology integrates pseudo-label refinement and self-training techniques tailored for the target domain. These advancements notably enhance the model's capacity to adapt and maintain robustness across diverse target domain scenarios. The results demonstrate that the proposed method surpasses current cross-domain road extraction techniques, which showcases an exceptional ability to extract interconnected road networks.