Attenuation surface noise with autoencoders
Ivan Sanchez, Daniel O. Trad, William M. Agudelo, Daniel Sierra
Surface wave attenuation is one of the essential stages in data processing for land seismic exploration. Conventionally, this stage involves $f-k$ filters that can separate the surface waves in a specific area of frequency and wavenumber domain. However, these filters must be tuned manually shot by shot according to the surface wave behaviour in the acquisition zone. Thus the quality of the filtered data depends on human expertise, and processing times increase according to the number of shots. Moreover, the $f-k$ filter requires seismic data with uniform spatial sampling, which is not possible in some complex land area acquisitions. Therefore, we propose to use a convolutional autoencoder to predict the Radon model of seismic data without surface wave noise. The Radon model allows working with seismic data with irregular spatial sampling. We train the autoencoder with synthetic data generated by elastic wave modelling with several 2D earth models. The results show that the trained model can accurately attenuate the surface noise with a performance similar to well-tuned f-k filters.