NBASNet: a network-based approach with super token sampling for accurate road extraction in ambiguous and obscured environments

Abstract

Road extraction remains a challenging task due to irregular shapes, varying widths, occlusions, and complex terrain. Existing models often fail to address key issues, such as extraction interruptions caused by trees, buildings, or shadows, and the difficulty of distinguishing road edges from surrounding features under dense occlusion and complex lighting conditions. To address these challenges, we propose NBASNet, a novel network architecture for improving road extraction in complex scenes. NBASNet integrates several key innovative modules: the SMC module captures texture features of occluded roads and contextual information; the STCM module captures global dependencies via super tokens; the CMSHA and DMLP modules enhance multi-scale feature extraction and local information; and the CGF module optimizes multi-scale feature representation. Experiments across two datasets show that NBASNet outperforms existing methods across all evaluation metrics, and visualizations further verify its effectiveness and reliability in addressing the aforementioned challenges.