Auto-Detection Method Using Convolution Neural Network for Bottom-Simulating Reflectors

In studies on gas hydrate, bottom-simulating reflectors (BSR) are used to determine the potential hydrate-bearing sedimentary layers. Usually, BSR detection is performed manually by experienced interpreters. Therefore, a method for implementing an automatic BSR detection process should be established. In this study, we develop a novel architecture for BSR characterization using the convolutional neural network (CNN) technique. We propose the use of Stokes' transform (ST) to obtain a time-frequency spectrum for the input of CNN. ST fully uses the frequency content of the seismic data, and a part of the 3D seismic data collected from the Blake Ridge is utilized to train the CNN. Synthetic seismic records with variable signal-to-noise ratios (SNR), as well as Blake Ridge seismic data, were used to validate the detection effect of the CNN. Results show that the CNN trained by this method exhibits excellent performance in noise-resistant testing and achieves an accuracy of more than 89% in field seismic data detection.