The potential of blended seismic acquisition to improve acquisition efficiency and cut acquisition costs is still open, particularly with efficient deblending algorithms to provide accurate deblended data for subsequent processing procedures. In recent years, deep learning algorithms, particularly supervised algorithms, have drawn much attention over conventional deblending algorithms due to their ability to nonlinearly characterize seismic data and achieve more accurate deblended results. Supervised algorithms require large amounts of labeled data for training, yet accurate labels are rarely accessible in field cases. We present a self-supervised multistep deblending framework that does not require clean labels and can characterize the decreasing blending noise level quantitatively in a flexible multistep manner. To achieve this, we leverage the coherence similarity of the common shot gathers (CSGs) and the common receiver gathers (CRGs) after pseudo-deblending. The CSGs are used to construct the training data adaptively, where the raw CSGs are regarded as the label with the corresponding artificially pseudo-deblended data as the initial training input. We employ different networks to quantitatively characterize decreasing blending noise levels in multiple steps for accurate deblending with the help of a blending noise estimation–subtraction strategy. The training of one network can be efficiently initialized by transfer learning from the optimized parameters of the previous network. The optimized parameters trained on CSGs are used to deblend all CRGs of the raw pseudo-deblended data in a multistep manner. Tests on synthetic and field data validate the proposed self-supervised multistep deblending algorithm, which outperforms the multilevel blending noise strategy.

混合地震采集在提高采集效率和降低采集成本方面的潜力仍然存在，特别是通过高效的去混合算法为后续处理程序提供准确的去混合数据。近年来，深度学习算法，特别是监督算法，由于能够非线性表征地震数据并获得更准确的去混合结果，因此比传统的去混合算法受到了更多关注。监督算法需要大量标记数据进行训练，但在现场案例中很难获得准确的标签。我们提出了一种自监督的多步骤去混合框架，不需要干净的标签，并且可以以灵活的多步骤方式定量地表征降低的混合噪声水平。为了实现这一点，我们利用伪去混合后公共炮集（CSG）和公共接收器集（CRG）的相干相似性。 CSG 用于自适应地构造训练数据，其中原始 CSG 被视为标签，相应的人工伪去混合数据作为初始训练输入。我们采用不同的网络来定量表征多个步骤中降低的混合噪声水平，以便借助混合噪声估计减法策略进行精确的去混合。通过从前一个网络的优化参数进行迁移学习，可以有效地初始化一个网络的训练。在 CSG 上训练的优化参数用于以多步方式对原始伪去混合数据的所有 CRG 进行去混合。对合成数据和现场数据的测试验证了所提出的自监督多步去混合算法，该算法优于多级混合噪声策略。