GCGR-Net: Road Extraction Model Based on Graph-Convolution Global Reasoning Network

The task of road extraction in remote sensing is subjected to pronounced challenges, which arise from a variety of converging factors. To increase road extraction accuracy, road extraction model based on graph-convolution global reasoning network is introduced in this study. Firstly, by utilizing deformable convolution, the DEFB module is proposed to capture local road features, conduct feature extraction at different scales, integrate richer contextual information, and enhance the expression capability of local features. Subsequently, the spatial information of the road feature is extracted in the channel domain using depth-wise separable convolution. Furthermore, spatial features of road nodes are captured in a global context using graph convolution, so that the model obtains comprehensive road topology information. Finally, a global reasoning method is adopted to efficiently capture long-range dependencies between road semantic information. The proposed method outperforms the compared models in the mean intersection over union (mIoU) scores of 80.24% and 79.18% on the respective datasets. The proposed model has been proved to be effective in addressing the problems of inaccurate and incomplete road extraction that arise from object shadows, vegetation occlusions, as well as challenges "same-spectrum-different-object".