COHERENT NOISE SUPPRESSION VIA A SELF-SUPERVISED DEEP LEARNING SCHEME

Coherent noise attenuation is an essential step in seismic data processing to improve data quality and signal-to-noise ratio. The use of deep learning based approaches for noise suppression has grown throughout the last five years due to neural networks strength in pattern recognition tasks and their low computation cost, i.e. fast application during the inference stage. A limitation of the majority of such procedures is their requirement for noisy-clean pairs of data for training. Here, we propose the use of self-supervised procedure, namely, Structured Noise2Void, which has no such requirements. Through the inclusion of a noise mask, the coherency of noise is suppressed by randomising the noise, allowing the network to learn how to predict only the signal component of a sample's value. Numerical experiments on synthetic and field seismic data demonstrate that our method can effectively attenuate trace-wise coherent noise. In the synthetic example, noise was injected into ten random traces, which showed no notable indication of their previously noisy state after denoising. In the field data, some locations already exhibited trace-wise coherent noise. After application of the trained network, the noise on these traces was drastically reduced resulting in a notable continuation in the seismic wave's first arrival.