Abstract: Curvelet transform can achieve sparse representation of common curves in seismic records. Based on the characteristics of the sparsity of effective signals and noise coefficients in the frequency domain, noise attenuation in seismic data can be realized through energy threshold screening. However, due to the complex wave field of actual seismic data and the insufficient robustness of non-Gaussian random noise in terms of noise energy distribution, it is difficult to achieve an ideal denoising effect merely by this method. Based on the characteristics of multi-scale decomposition of signals of the variational mode decomposition (VMD) algorithm, combined with those of the sparsity in the curvelet transform, we proposed a joint method suitable for suppressing random noise in seismic data. First, the seismic signal is decomposed on multiple scales by the VMD algorithm to obtain a set of intrinsic mode functions (IMF) with better sparsity, and the noise-dominant IMF containing random noise are screened out through signal-to-noise ratio analysis. Then, taking advantages of the good recognition of noise by curvelet transform in frequency domain, the extracted noisy IMF profile is processed via adaptive thresholding. Finally, the denoised profile is obtained by reconstructing the signal with the processed IMF components and the effective IMF components. By comparing the denoising results of the proposed joint method with those of the traditional curvelet transform on both synthetic and real seismic data, it is demonstrated that our method can achieve a superior suppression on seismic random noise.