To improve the denoising performance of seismic data contaminated with random and coherent noises, a hybrid denoising scheme is proposed in this paper. It aims to whiten the random noise and identify the coherent noise for the preserved or prominent seismic features. Using the wavelet and curvelet basis functions in curvelets alternately, the hybrid denoising scheme utilises the representation of edges and singularities along curves. Then it adapts the wavelet-based higherorder correlative stacking denoising method from seismic exploration sequentially. With regard to seismic records for bedrock surface detection after the artificial backfill, the noisy data are significantly improved both in terms of denoising and improving fidelity. To illustrate the advantage of the hybrid denoising scheme, a comparison of the performances between the different individual denoising methods, including the higher-order correlative stacking method and curvelets with wavelet and curvelet basis functions, has been presented for the complex seismic records contaminated with different noises. Numerical case studies and a field data analysis have been used to show that the proposed hybrid denoising scheme is more effective for seismic data containing complex features than the individual denoising methods. deal with the complicated events of curved and conflicting dips in a single scale (Shan et al. 2009; Warden et al. 2012; Chen and Ma 2014).Correlative statistics in seismic records is introduced early in the theory of second-order stacking (Robinson 1970) for incoherent noise denoising. However, the seismic data contaminated with coherent noises cannot be improved much under the second-order stacking scheme. Therefore, higherorder correlative statistics in spectral domain are designed for seismic data processing (Zhang 1996;Zhang and Xu 2006). One kind of the spectral transform uses the wavelets. It allows the so-called "sparse representation" of the original information, in which the significance is only attached if a high correlation between the target event and the chosen wavelet is found in some characteristic frequency content (Saracco et al. 2007;Ma and Plonka 2010;Mauri et al. 2012). In seismic exploration, a known fact is is that each trace in the records can be formulated using a dilated and translated basic wavelet in one dimension (1D). The selection of such a wavelet leads to representing most energy of the signal using a limited number of chosen coefficients. Therefore, 1D wavelet is only suited for singularity detection and individual trace processing (Roueff, Chanussot and Mars 2006;Shan et al. 2009), and the 2D wavelet representation is only significant in horizontal,