Road extraction network based on global–local perspective and morphological adaptation

Road extraction from remote sensing images is vital for urban planning and environmental monitoring. Roads often display complex tubular structures amidst intricate landscapes, making accurate extraction challenging. Existing methods, in their efforts to capture long-range dependencies, often result in spatial information loss, insufficient exploitation of global information, and the presence of low-quality redundant features. To address these challenges, we propose a road extraction network with a global–local perspective and morphological adaptation, called global–local morphological adaptation network (GLMA-Net). First, a dual-branch encoder, combining convolutional neural networks with axial attention mechanisms in parallel, is designed to efficiently extract global and local features. Specifically, morphological adaptive convolutions are introduced to provide a more flexible receptive field, focusing on the tubular structure of roads while suppressing the interference from complex environments. In addition, a cross-scale information flow module is integrated after the encoder to transmit multiscale features in a cross-fusion manner, effectively balancing global and local information. Finally, an iterative detail optimization module is constructed at the network's final stage to maximize exploitation of spatial information from specific modalities, further enhancing the fine details of the road extraction results. On the challenging BJRoad dataset, GLMA-Net achieves 76.47% F1 and 61.90% IoU, outperforming the state-of-the-art method by 0.86 and 1.12 percentage points, respectively—demonstrating its superior accuracy and continuity in road extraction.