Deep learning application in time-frequency analysis for noise attenuation

Time- frequency analysis of seismic data has been proven to be a reliable method for noise attenuation. Differentiation between the noise and signal energy is traditionally done by thresholding the coefficients, keeping the stronger values as they are assumed to be representing the signals, while zeroing out the noise energy. Despite its effectiveness, this is a timely procedure and needs the user's interaction for each seismic survey. Considering the growing success of the deep learning in image and signal processing, it is used in this paper to estimate the signal coefficients of the trace in the time-frequency plane. As the continuous wavelet transform is a successful method for time-frequency analysis of the seismic data, it has been used in this study to calculate the two dimensional input for the deep learning. To enable the network to focus on attenuation of the random noise and not simply modeling the signals in a specific geological setting, traces from three different datasets have been used for training the network. Testing the final network with traces not used for training shows promising results as it can conveniently attenuate the noise energy compared to the traditional method. © EAGE 2019.