Seismic data denoising based on the convolutional neural network with an attention mechanism in the curvelet domain

Objective Noise in seismic data significantly affects the accurate interpretation of subsurface stratigraphic information. Given that effective signals with pronounced lateral correlations in seismic data are distributed in specific coefficients but random noise typically spreads uniformly over all coefficients in the curvelet domain, more effective separation of signals can be achieved. Methods The convolutional neural network based on the attention mechanism can adaptively extract key information by focusing on important features of images. Hence, this study proposed a noise attenuation method for seismic data using a convolutional neural network based on the curvelet transform and attention mechanism (Curvelet-AU-Net). First, the curvelet coefficients of noise-containing seismic data were obtained through curvelet transform to analyze the distributions of effective signals and noise in the curvelet domain. Second, a U-Net network with a convolutional block attention module (CBAM) was employed, with the curvelet coefficients of noise-containing seismic data as input data for training and the curvelet coefficients of noise-free seismic data as labels. Then, the parameters of the network were updated by comparing the loss function values of actual outputs and labels and backpropagating gradients layer by layer. The network training was completed as the loss function value reached its minimum. Finally, the test data were put into the trained network model. The denoising results of seismic data were obtained by performing inverse curvelet transform on the network output data. Results and Conclusions The processing results of simulation and actual data show that compared to conventional methods and ordinary convolutional networks, the method proposed in this study demonstrates superior attenuation effects on common noise (e.g., random noise) under different noise levels and scales, achieving higher signal-to-noise ratios and fidelity for seismic signals. This method, integrating the sparse representation of the Curvelet transform and the adaptability of deep learning models, provides a novel approach for the noise attenuation of seismic data.