Effective suppression of random noise while preserving seismic signal has consistently been a key challenge in seismic data processing. The high dimensionality and complex spatial structure of three-dimensional (3D) seismic data often lead traditional denoising methods to fail to meet the requirements of noise suppression. To address this problem, we design a novel 3D seismic denoising method based on unsupervised deep learning, which integrates enhanced multi-scale features fusion and an efficient transformer (EMST). The basic framework of the EMST employs two main encoding-decoding paths to capture multi-scale features of 3D seismic data. An attention mechanism further enhances the expression of features at different scales, which are then integrated by the feature fusion layer. Subsequently, an efficient transformer is incorporated to strengthen the model’s global modeling capability. Moreover, a Monte Carlo patch selection method is employed to select representative patches for optimizing the training sets, thereby enhancing the training efficiency. Experimental results on multiple data indicate that the EMST improves denoising performances and reduces signal leakage compared to baseline methods.
Zhang et al. (Wed,) studied this question.