Ground-truth-free deep learning for 3D seismic denoising and reconstruction with channel attention mechanism

Seismic denoising methods using supervised training normally rely on a large number of high-quality paired training data sets to obtain satisfactory results. There are two ways to generate labels for network training: one is to create synthetic data using the wave equation and the other is to use denoised data obtained through prior denoisers. However, using either training data generation methods will limit the networks’ denoising performances when faced with complex field data. Here, we develop a ground-truth-free method for 3D seismic data denoising. To improve its denoising efficiency, we exploit efficient channel attention and a convolutional block attention module to adjust the response of different channels to capture their correlation and significance using only a few parameters. Our method contains three stages: training set extension with a 3D patching scheme, training samples denoising using our unsupervised network, and an unpatching approach to reconstruct the denoised data. Numerical experiments demonstrate that our method outperforms the benchmark approaches in terms of signal-to-noise ratio improvement and useful signal preservation. Moreover, because our method is trained in an unsupervised manner, it offers greater flexibility than supervised methods when faced with different types of noise.