The internal architecture of sand body controls reservoir heterogeneity and further controls the movement law of oil and water. Especially in the middle and late stage of oil and gas field development, fine interpretation of sand body architecture is the key to clarify the distribution law of remaining oil and gas and guide the further efficient development of oil and gas fields. Taking the No. 2 sand group of Eocene Pinghu Formation of X gas field in Xihu Sag, Donghai Shelf Basin as an example, based on few well conditions, we make a fine interpretation of sand body architecture of fluvial reservoir guided by the theory of seismic sedimentology and architecture analysis. In this process, we apply the method of combining the seismic attributes fusion based on cluster analysis and the seismic attributes fusion based on deep learning. This study shows that the correlation between single seismic attribute and well point sandstone thickness is poor. The fusion of multiple seismic attributes can improve the ability of sand body prediction to a certain extent, but the predicted sand body boundary is not clear. Based on the conventional attribute fusion, we increase the number of samples for deep learning by establishing some virtual wells in regular grid, and further apply the deep learning seismic attribute fusion method to fuse 15 kinds of seismic attributes without difference. The deep learning of seismic integrated attributes effectively improves the ability of sand body prediction and channel boundary characterization, and eliminates the abnormal high value areas in the seismic attributes fusion map based on cluster analysis. The plane distribution and superposition relationship of the fifth-order and fourth-order channel sand body architecture of No. 2 sand group of Pinghu Formation are defined. The single channel of the No. 2 sand group in X gas field is characterized by meandering distribution, with bending index of 1.63, width of 150-480 m and point dam span of 154-366 m. This set of ideas and methods can be effectively promoted in areas with wide coverage of seismic data, but few well and uneven distribution.