In inland river waters characterized by heavy traffic and a complex navigation environment, ferries frequently need to cross navigational channels due to the nature of ship activity. These situations are particularly challenging in the face of complex and dynamically changing encounter scenarios, which involve significant risks and uncertainties, thereby posing substantial challenges to the safety of navigation in these waters. To address these challenges, this paper proposes a scenario-constrained graph convolutional network (SC-GCN) for ferry behavior prediction. The model comprises an encoder module, a scene-labeling attention module, and a decoder module. The scene-labeling module employs expert knowledge, such as proficient boating techniques, to recognize and classify ferry encounter scenarios, thereby constructing scene labels. The encoder module subsequently extracts spatiotemporal interaction features of the ferry using a combination of GCN layers, LSTM layers, and scene label information. The decoder module then retains contextual information and adjusts the output as predictions are generated sequentially. To validate the SC-GCN model, Automatic Identification System (AIS) data from the Wuhan section of the Yangtze River was used to classify ferry encounter scenarios. The results demonstrate that the proposed scene-constrained graph convolutional network effectively predicts ferry behavior in encounter scenarios.