Multi-condition seismic data denoising using feature-expanded gradient penalty generative adversarial network

Abstract High-quality seismic data serve as a critical foundation for imaging and interpretation. However, field seismic data are contaminated by noise, leading to numerous spurious seismic signals that adversely affect subsequent processing and interpretation. Currently, supervised deep learning methods face challenges due to the scarcity of authentic labels for 3D seismic images, while unsupervised approaches suffer from a lack of guidance during training, resulting in blurred denoising outcomes. To deal with these matters, it is essential to propose an adaptive semi-supervised random noise attenuation method that leverages pseudo-labels, which are easier to generate, to guide the model effectively. In this work, we put forward a semi-supervised gradient-penalized generative adversarial network with feature expansion (FEWGAN-SGP). The network first employs a global residual structure to ensure overall convergence during training. Subsequently, a channel attention mechanism with feature expansion is adopted to extract finer details. Finally, a smooth gradient penalty term is applied to enhance the distinction between field and generated images, thereby promoting deeper model learning. Experiments were conducted on both 2D and 3D datasets, and the results indicate that, compared with mainstream traditional methods and unsupervised learning approaches, the proposed method demonstrates excellent signal preservation and noise attenuation capabilities in practical applications.