Unsupervised CNN Based on Self-similarity for Seismic Data Denoising

Convolutional neural network (CNN)-based methods are powerful tools for seismic data denoising. Most methods adopt a supervised learning strategy, which requires noise-free labels to construct an objective function to guide the training of network parameters; however, it is impossible to obtain true noise-free field data. We propose a novel unsupervised random-noise-suppression method that can train a network directly on noisy target data without noise-free labels. The proposed method is inspired by the simple denoising idea of averaging multiple noisy observations, and it requires noise to satisfy two assumptions: it should be zero-mean and independent of the signal. In this method, multiple observations can be used as labels, and the loss function is constructed as the mean square error expected between the network output and these observations. The trained network estimates the expected value of these noise observations (i.e., the clean signal). This unsupervised method theoretically requires multiple noisy observations. Considering the good nonlocal self-similarity of seismic data, we used self-similar blocks to rearrange the data to construct multiple pseudo-observations and finally realize unsupervised training. The proposed method was compared with the traditional f-x deconvolution, curvelet, and generator CNN method on synthetic and field data, and the experimental results verified the effectiveness of the proposed method.