自由表面多次波压制是海洋地震数据处理中的难题之一。本文利用地震信号的非高斯性，从不同方法的联合使用、多次波自适应相减中3D匹配滤波器的求解、多次波自适应相减的计算效率提升这三个方面，研究自由表面多次波的压制方法，并在模型和实际数据上验证所提方法的有效性。本文的研究成果分为以下三部分：提出了一种结合Radon滤波和Radon域多次波自适应相减的多次波压制方法。所提方法将原始数据的抛物线Radon域剖面分为三个区域，切除一次波映射的Radon区域，并保留未被有效预测的多次波映射的Radon区域；在第三个区域，2D卷积混合盲分离方法用来将预测多次波与原始数据进行匹配，通过求解匹配滤波器来估计多次波。将这三个Radon区域合并后变换到时间-空间域，并从原始数据中直接减去得到一次波估计结果。所提方法在压制未被有效预测的自由表面多次波的同时，能有效地分离具有相近剩余时差的一次波和多次波。提出了一种基于3D卷积混合盲分离的多次波自适应相减方法。所提方法将多次波自适应相减表征为一个3D卷积混合盲分离问题，通过求解3D匹配滤波器来消除2D预测多次波和真实3D多次波之间的差异。为避免3D滤波器产生的过匹配问题，所提方法采用相同的3D匹配滤波器来同时拟合多个原始数据道集。另外，所提方法对一次波施加非高斯最大化约束，并利用迭代重加权最小二乘算法求解3D匹配滤波器。相对于2D卷积混合盲分离方法和基于一次波能量最小化的方法，所提方法在压制多次波的同时，能更好地保护一次波。提出了一种GPU加速实现的基于迭代收缩阈值算法的多次波自适应相减方法。GPU并行计算首先用来对迭代重加权最小二乘算法进行加速。然后，迭代收缩阈值算法被引入到2D/3D卷积混合盲分离方法中。迭代收缩阈值算法采用收缩阈值算子促进一次波的非高斯性，并迭代求解匹配滤波器。相比于迭代重加权最小二乘算法，迭代收缩阈值方法只需进行一次矩阵-矩阵相乘和基于克莱斯基分解算法进行一次矩阵求逆。最后，所提方法采用GPU并行计算对迭代收缩阈值算法进行加速。所提方法在保持多次波自适应相减效果的同时，能有效地提高2D/3D卷积混合盲分离方法的计算效率。

Surface-related multiple attenuation is one of the challenging problems in the data processing of ocean seismic exploration. This thesis uses the non-Gaussianity of seismic signal to research the surface-related multiple attenuation methods in terms of combining different methods for surface-related multiple attenuation, estimating the 3D matching filter for adaptive multiple subtraction and improving computational efficiency of adaptive multiple subtraction. Three contributions of this thesis are summarized in the following.This thesis proposes the demultiple method combining Radon filtering and Radon domain adaptive multiple subtraction. The proposed method divides the Radon model of the original data into three areas, mute the Radon area where primaries map into and keep the Radon area where ineffectively predicted surface-related multiples map into. In the third area the method based on 2D blind separation of convolved mixtures is used to make the predicted multiples match with the original data and the multiples can be estimated by solving the matching filter. The three Radon areas are blended and then transformed into the time-space domain to model multiples. The modeled multiples are subtracted directly from the original data to estimate primaries. The proposed demultiple method can not only attenuate the ineffectively predicted surface-related multiples, but also separate primaries and multiples with close residual moveout effectively.This thesis proposes the adaptive multiple subtraction method based on 3D blind separation of convolved mixtures. The proposed method expresses the adaptive multiple subtraction as a problem of 3D blind separation of convolved mixtures, and eliminates the mismatches between 2D predicted multiples and true 3D multiples by solving the 3D matching filter. To avoid the possible over-fitting problem caused by the 3D matching filter, the proposed method employs the same 3D matching filter to fit several original data gathers. In addition, this thesis introduces the non-Gaussian maximization constraint of primaries into the proposed method and solves the 3D matching filter with the iterative re-weighted least-squares algorithm. Compared with the adaptive multiple subtraction based on 2D blind separation of convolved mixtures and adaptive multiple subtraction based on the energy minimization of primaries, the proposed method can better preserve primaries while suppressing multiples effectively. This thesis proposes the adaptive multiple subtraction method based on the iterative shrinkage thresholding algorithm with GPU acceleration. GPU parallel computation is first used to accelerate the iterative re-weighted least-squares algorithm. Then the iterative shrinkage thresholding algorithm is used to estimate the 2D/3D matching filter for the adaptive multiple subtraction based on 2D/3D blind separation of convolved mixtures. The iterative shrinkage thresholding algorithm employs the shrinkage thresholding operator to promote the non-Gaussianity of primaries and estimates the matching filter with iterative steps. Compared with the iterative re-weighted least-squares algorithm, the iterative shrinkage thresholding algorithm only needs matrix-matrix multiplication and matrix inversion based on the Cholesky decomposition algorithm once. Finally the proposed method uses the GPU parallel computation to accelerate the iterative shrinkage thresholding algorithm. While keeping the effectiveness of multiple attenuation, the proposed method can improve the computational efficiency of the adaptive multiple subtraction based on 2D/3D blind separation of convolved mixtures effectively.