<p>This study presents a comprehensive approach for detecting channel features in 3D marine seismic data by integrating conventional seismic attributes with deep learning-based semantic segmentation. Initially, widely used seismic attributes, including similarity, curvature, and spectral decomposition, were applied to delineate channel boundaries and guide manual interpretation. These attributes provided valuable insight but showed limitations in identifying complete channel bodies and distinguishing them from other geological features. To overcome these limitations, convolutional neural networks (CNNs) were employed, specifically a standard UNet and an enhanced UNet architecture with Attention mechanism. A training dataset was generated by manually interpreting 15 representative seismic time slices, which were subsequently expanded to 3,000 samples through data augmentation using geometric transformations. Both models were trained using a combined Dice and Binary Cross-Entropy loss function to address class imbalance and improve boundary delineation. Comparative analysis demonstrates that the Attention UNet outperforms the standard UNet in capturing subtle and discontinuous channel features. Quantitative evaluation using IoU and Dice metrics confirms the effectiveness of the Attention UNet, offering a robust and automated alternative for seismic channel interpretation.</p>