Self-supervised learning for low frequency extension of seismic data

Conventional towed streamer data usually do not have reliable signals below 3 Hz, which brings huge challenges to seismic imaging algorithms such as full waveform inversion (FWI). We propose a scheme using convolutional neural networks (CNN) to extend seismic data's frequency band toward increased low frequencies. Due to a lack of annotated data with desired low frequencies for training our CNN, we propose an implementation scheme based on self-supervised learning. First, we train a neural network on real seismic data on relatively higher frequencies labeled with and without the relatively low frequency components. Then we apply the trained network model on the down-sampled seismic data to generate lower frequency components. The major advantage of our scheme is that we have adequate realistic labeled data for some frequencies by applying different sampling intervals and frequency ranges for the training dataset and testing dataset. Both synthetic and real data examples demonstrated the effectiveness and robustness of the method. Presentation Date: Tuesday, October 13, 2020 Session Start Time: 8:30 AM Presentation Time: 10:35 AM Location: 351F Presentation Type: Oral