Local and global feature fusion with boundary-guided network for road extraction from remote sensing imagery

Road extraction in remote sensing imagery is of great importance. Because of the complex shape, narrowness, and high span of road, the results are often unsatisfactory. To address these problems, this study proposes a local and global feature fusion with boundary-guided network named Road-LGB. Road-LGB takes remote sensing imagery as input and produces binary road masks in an end-to-end manner. During feature encoding, both convolutional neural network and Transformer are employed to capture local and global information, respectively. In the decoding path, the size of feature is gradually restored to obtain the results using convolution, upsampling, and skip connection. Notably, a lightweight boundary refinement and guidance module is introduced to enhance feature learning. Extensive experiments conducted on the public Massachusetts road dataset demonstrate the superior performance of the proposed method compared to several state-ofthe-art deep learning-based methods. Furthermore, ablation studies further validate the positive contribution of each module to the overall performance improvement.