Abstract: The river channel sand detection can provide critical insights for hydrocarbon exploration, sediment transport, and paleofluvial geomorphology analysis. However, manual identification of river channels is time-consuming, cumbersome, and subjective, while traditional river channel detection methods based on attributes such as coherence cube and curvature are susceptible to noise and faults. In contrast, owing to powerful feature extraction and efficient feature representation capabilities, deep learning has been widely applied in seismic data processing. This paper proposed a river channel detection method using 3D seismic data and attention mechanisms in convolutional neural network (CNN). The method trained the constructed network using synthetic seismic data and river channel labels, achieving effective detection of river channels from test data and actual 3D seismic data. The proposed method is of great significance for achieving efficient and automated river channel detection during actual production.