<p>The size of the receptive field limits traditional deep convolutional networks and needs help to fully capture the detailed semantic information of targets in complex remote sensing images. At the same time, the attributes of boundary pixels are ambiguous due to many external factors, such as illumination and imaging technology. To solve these problems, we propose a spatiotemporal relationship-guided graph convolutional network for image segmentation. We aim to explore the global and contextual semantics of targets in images and judge the attribute information of boundary pixels. On the one hand, a particular contextual attention module is embedded in the feature extractor to improve the feature extractor’s global and local detail representation. On the other hand, the graph convolution layer, enhanced by the recursive network, aggregates the high-level and low-level features of nodes to strengthen the interactivity between semantic information at different levels. At the same time, long-term dependencies are established between boundary pixels and target center pixels to constrain boundary pixels and improve their attribute decision-making ability. It is worth noting that we designed a weighted loss function to supervise, adjust, and optimize different modules separately. Finally, experimental results on the indian driving dataset, uestc all-day scenery datasets and deepglobe land cover classification challenge datasets show that the proposed framework is robust and has good segmentation performance, with F1 of 89.03%, 98.38% and 85.46%, respectively.</p>