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

The reduction of seismic white noise is an important step in seismic data processing. Although machine learning algorithms can be successfully used to remove noise in seismic data, the preparation of "pure signals" and "pure noise" is one of challenges in seismic denoising with machine learning. To address this problem, we assume that the noise in the seismic data behaves according to a certain distribution, e.g. a normal distribution. We treat the field seismic data as "pure signals" and a simulated noise as "pure noise". By adding the simulated noise (pure noise) to the field seismic data (pure signals), new "noisy seismic data" are formed. Then we can train a model to distiguish the pure signal (learning signal) or the pure noise from the noisy seismic data. Subsequently, by applying the trained model to the field seismic data, the noise-reducted seismic data is obtained. To show which learning (either signal learning or noise learning) is more effective in seismic denoising by machine larning, we compare the denoised seismic data through comparing the commonly-used seismic geometric attributes (coherence and curvature) calculated from the denoised seismic data. The application shows that the signallearning is a better choice in seismic random noise reduction. We choose structure-oriented filtering (SOF) as a traditional denoising method to evaluate our method to enhance the stratigraphic and structural features of the seismic data. The results show that our method is superior to SOF in reducing random noise in seismic data.