Low-frequency band-shaped swell noise with strong amplitude is common in marine seismic data. The conventional high-pass filtering algorithm widely used to suppress swell noise often results in serious damage of effective information. This paper introduces the residual learning strategy of denoising convolutional neural network (DnCNN) into a U-shaped convolutional neural network (U-Net) to develop a new U-Net with more generalization, which can eliminate low-frequency swell noise with high precision. The results of both model date tests and real data processing show that the new U-Net is capable of efficient learning and high-precision noise removal, and can avoid the overfitting problem which is very common in conventional neural network methods. This new U-Net can also be generalized to some extent and can effectively preserve low-frequency effective information. Compared with the conventional high-pass filtering method commonly used in the industry, the new U-Net can eliminate low-frequency swell noise with higher precision while effectively preserving low-frequency effective information, which is of great significance for subsequent processing such as amplitude-preserving imaging and full waveform inversion.

在海洋地震数据中，振幅较大的低频带状膨胀噪声很常见。广泛用于抑制膨胀噪声的常规高通滤波算法通常会严重破坏有效信息。本文将去卷积神经网络（DnCNN）的残差学习策略引入到U形卷积神经网络（U-Net）中，以开发出更具通用性的新U-Net，从而可以高精度消除低频膨胀噪声。模型数据测试和真实数据处理的结果表明，新的U-Net具有高效的学习能力和高精度的噪声消除能力，并且可以避免传统神经网络方法中非常普遍的过拟合问题。这种新的U-Net也可以在一定程度上得到推广，并可以有效地保留低频有效信息。与业界常用的常规高通滤波方法相比，新型U-Net可以更高精度地消除低频骤增噪声，同时有效保留低频有效信息，这对于幅度等后续处理具有重要意义。 -保留成像和全波形反演。