Abstract: Tight channel sands represent a significant reservoir type with high potential for hydrocarbon accumulation in continental basins. However, conventional methods often fall short in accurately characterizing the 3D distribution of these channel sands due to their multi-phase development, complex stacking relationships, and rapid lateral variations. To overcome this challenge, this study proposes an automated channel identification method based on an improved U-Net deep learning network. Guided by seismic sedimentology, the first step involves applying Wheeler transformation to the time-domain seismic data to incorporate sedimentary cycle characteristics, which facilitates the identification of sandstone stacking relationships and yields high-quality training samples. This is followed by the integration of a cascaded dilated convolution module and a spatial attention mechanism into the U-Net architecture. This integration strengthens the network's capacity to extract multi-scale features and thus improves the delineation of narrow, thin, and superimposed channel boundaries. Finally, data augmentation methods tailored to channel characteristics are employed to automatically generate a large number of training samples for model training and testing. Field application results demonstrate that the improved U-Net significantly enhances the accuracy of boundary identification for multi-phase superimposed channels and achieves 3D characterization of single-phase channel systems. This approach offers reliable technical support for the evaluation of tight channel sandstone reservoirs and the optimization of exploration strategies.