Abstract: Because of its multi-scale and multi-direction characteristics, 2D curvelet transform is widely used in multiple suppression. The multiple attenuation in 2D curvelet domain is better than the conventional method, but it is prone to "over-matching" the primaries and difficult to meet the requirements of higher separation fidelity. To solve this problem, a multiple amplitude-preserving attenuation method based on 3D curvelet transform is proposed. In this method, the seismic data and multiples predicted are arranged according to the 3D shot gather (X axis represents the shot point direction, Y axis represents the offset direction, Z axis represents time) and transformed into the 3D curvelet domain. They are divided into multi-scale and multi-angle components more finely, and then adaptive matching and soft threshold separation of primaries and multiples at different scales and angles are carried out respectively. Finally, the seismic data without multiples are obtained by inverse 3D curvelet transform. In order to improve the separability between primaries and multiples, NMO correction of seismic data and multiples is processed by using velocity information before 3D curve transform, which can effectively avoid primary-multiple "over-matching". The results of model data and actual seismic data processing show that the multiple attenuation method in 3D curvelet domain can effectively eliminate multiples, demonstrating higher separation fidelity and higher signal-to-noise ratio.