High-resolution remote sensing image road extraction via dual-branch fusion of convolutional and graph convolutional network

Deep learning–based methods, particularly deep convolutional neural networks (DCNNs), have demonstrated exceptional performance in extracting roads from high-resolution remote sensing images. However, current DCNNs often struggle to accurately extract small roads or roads that are substantially occluded because of the loss of position and global context information, resulting in incomplete and fragmented outputs. Graph neural networks have shown promise in modeling long-range and cross-scale features, as well as representing the overall structure of irregular objects. We propose a dual-branch fusion of convolutional and graph convolutional network (DBCGCN) for road extraction from high-resolution remote sensing images. The DBCGCN comprises two branches—a graph convolutional network (GCN) and a convolutional network—to generate complementary high-level semantic features and low-level detail features at superpixel and pixel levels, respectively. Furthermore, we employ a mathematical strategy to facilitate the exchange of features between graph nodes and image pixel points, enabling seamless collaboration between GCN and DCNN within a network. Extensive experiments conducted on widely recognized road extraction datasets validated the effectiveness of the proposed method in terms of both quality and quantity. The proposed network can improve the accuracy and reliability of road extraction in various real-world applications.