3D seismic waveform of channels extraction by artificial intelligence

Channels are important exploratory objectives in reflection seismology. Geologic bodies such as channels and point bar produce the laterally-inhomogeneous geological body (LGB) waveform response in seismic sections. Due to the limitation of the seismic signal's resolution, it always makes detailed interpretation challenging when channels are seriously affected by the signals of overlying strata or the underlying strata. We propose a method based on convolutional neural networks (CNN) to extract LGB waveform response from the raw seismic dataset. We divide raw seismic data volume into training data volume and testing data volume. Our CNN is trained with selected training data volume in which the LGB waveform response is clear. Then, we test the model by using the testing data volume in which the LGB waveform response is seriously covered. In order to utilize more information from different directions, we feed a three-dimensional (3D) training dataset to the network. Two 3D field seismic data examples verify the validity of our proposed method. After extracting LGB waveform response by using our method, the channel structure becomes clearer, which is very helpful to reduce the texture interpretation's uncertainty.