A strategy for preparing training data for machine learning for seismic noise reduction

Seismic white noise reduction is an important step in seismic signal processing. Recently, supervised deep learning algorithms have been successfully used to remove noise in seismic data. Preparing “pure signals” and “pure noise” is one of the main challenges in seismic denoising with supervised deep learning. To tackle this issue, we assume that the noise in the seismic data behaves in a certain distribution, such as normal distribution. We treat the field seismic data as the “pure signals” and a simulated noise as “pure noise”. A new “noisy seismic data” is formed by adding the simulated noise (pure noise) to the field seismic data (pure signals). Then, we can train a model to differentiate pure signal (learning signal) or pure noise from noisy seismic data. Next, the noise-reduction seismic data is obtained by applying a trained model to the field seismic data. To show which learning is more effective in seismic denoising, we compare the denoised seismic data and the most used seismic geometric attributes (coherence and curvature) computed from the denoised seismic data. The application shows that the learning signal is the better choice in seismic random noise reduction. We choose structure-oriented filtering (SOF) as the traditional denoising method to evaluate our method in enhancing the stratigraphic and structural features of the seismic data. The results demonstrate that our method is superior to SOF in random noise reduction of seismic data. • This paper develop a strategy of deep learning to denoise the seismic image without requiring labels.