Abstract: Tight channel sandstone reservoirs represent significant reservoir types in continental basins and serve as favorable locations for hydrocarbon accumulation. However, conventional methods often fall short in accurately characterizing the 3D spatial distribution of such channels, due to their multi-phase development, complex sand-body stacking patterns, and rapid lateral variations. To overcome these challenges, an automated channel identification method based on an improved deep learning approach is proposed. First, guided by seismic sedimentology principles, Wheeler transformation is applied to time-domain seismic data, incorporating sedimentary cycle characteristics to identify sandstone stacking relationships and obtain high-quality training samples .Secondly, the cascaded dilated convolution module and attention mechanism are integrated into a U-net network architecture. These enhancements strengthen the network's capacity to extract multi-scale channel features, particularly improving the delineation of narrow, thin, and overlapping channel boundaries. Thirdly, the data augmentation methods suitable for channel characteristics are employed to automatically generate numerous training samples, followed by model training and testing. Application in a real-field case demonstrates that the improved U-net method significantly increases the accuracy of boundary identification for multi-stage superimposed channels. It successfully achieves 3D spatial characterization and temporal staging of channel systems. This approach offers critical technical support for the evaluation of tight channel sandstone reservoirs and the optimization of exploration strategies.