LG-Umer: UNet-like network integrate local–global feature with novel attention for road extraction from remote sensing images

Road extraction from remote sensing images is a key research area in smart city development. While deep learning techniques have demonstrated remarkable effectiveness in this domain, existing approaches exhibit limitations: convolutional neural network (CNN)-based methods struggle to capture global contextual information for long-range road networks, vision transformer (ViT)-based methods fail to adequately extract multiscale local features, and hybrid CNN-ViT architectures overlook the synergistic guidance between local and global features. To address these challenges, we propose LG-Umer, a UNet-like network that integrates Local-Global features with a novel attention mechanism, combining the complementary strengths of CNNs and ViTs within an encoder-decoder framework. Specifically, the encoder employs a multiscale strip deformational module, which utilizes deformable convolutions to adaptively extract topological structures and variable-shaped local road features. In the decoder, a multistage gate unit module is introduced, incorporating a novel attention mechanism to model long-range dependencies by leveraging local features as attention operators for global feature refinement. Extensive experiments on three public benchmarks demonstrate the superiority of LG-Umer. It achieves IoU scores of 70.4%, 71.2%, and 68.7% on the Massachusetts Road, DeepGlobe Road, and CHN6-CUG datasets, respectively, surpassing recent state-of-the-art methods by 1.2%, 0.9%, and 1.1%. These results validate the effectiveness of our approach in balancing local detail preservation and global contextual modeling for road extraction tasks.