<p>Safe route selection has always been a challenge for maritime transportation. This paper proposes a dynamic route-aware spatial–temporal graph convolutional network (Dr-STN) to predict the navigation safety of routes in terms of path-segment and time-segment. Specifically, Dr-STN can capture the complex spatial and temporal dynamics between the safety of waypoints and sea conditions during navigation. In the spatial dimension, we design a dual-branch graph convolutional module, which not only captures the local impacts of neighboring seas but also focuses on the global impacts on waypoints from critical seas that are in extreme oceanic conditions. In the temporal dimension, we propose a decomposition-coupled gated unit, which individually captures trends and periodicities of each waypoint through a multiscale decomposition process, and then fuses the spatial–temporal dependencies among all waypoints on the whole route via a flexible gated recurrent unit. Meanwhile, to overcome the influence of limited labeled data on the prediction accuracy of Dr-STN, pre-training and fine-tuning techniques are employed. Subsequently, we propose a metric for comprehensively evaluating the navigation safety of a route, which focuses on the position and sequential order of safe or dangerous waypoints along the route, as well as the number of risk sub-routes. Furthermore, we design a route selection algorithm that prioritizes safety while considering minimum distance. Extensive experiments demonstrate that our methods are feasible and effective.</p>