DLCSIA-Net: dataset-level context and spatial information aggregation network for road extraction from remote sensing images

Accurate and efficient road segmentation in high spatial resolution remote sensing images (HSRRSIs) is crucial for automated driving, intelligent transportation, and disaster relief. However, road extraction from HSRRSIs remains challenging because roads have few visual features but strong structural priors and a distribution that extends beyond that of a single sample in the training dataset. Unfortunately, established approaches for extracting roads only capture global–local visual and local structural features in a single sample and are still susceptible to occlusions and noise from the background, leading to false and fragmented results. In this work, we propose a modified external attention (MEA) and dataset-level spatial information aggregation (DLSIA) module-based LinkNet for road extraction from HSRRSIs. Due to their ability to model the dataset-level context, MEA modules are introduced to the last two encoding stages to capture dataset-level features and contexts for the DLSIA module. The DLSIA module aggregates dataset-level topological information and guides the MEA module to extract reliable features. Extensive experiments with the Massachusetts and Deep Globe road datasets demonstrate that our approach effectively enhances the accuracy and continuity of segmented roads with the mutual reinforcement of the MEA and DLSIA modules.