Reverse time migration (RTM) is used to obtain complex structural images of subsurface media. The RTM can be expressed as a zero-lag cross-correlation between the source and receiver wave fields. As imaging conditions can be calculated based on the pre-stored source wavefield and the received wavefield generated during backward modelling, only the source wavefield must be stored in the memory (or disc). Therefore, reducing source-wavefield storage requirements can improve memory efficiency. High-performance computing based on graphic processing units (GPUs) is being developed to reduce the computational time in wave-propagation modelling. Accordingly, GPU-based RTM technology has the potential to improve the computational efficiency of RTM. Storage of the source wavefield wholly in GPU video random-access memory (VRAM) may further improve computational efficiency. In this paper, we present a new algorithm for a three-dimensional (3D) GPU-based RTM that can enable efficient storage of the source wavefield in VRAM using both Nyquist sampling and discrete cosine transform (DCT) compression. A numerical example employing a modified SEG/EAGE 3D overthrust model presented in this study verifies that the proposed algorithm requires only 2% of the memory usage of conventional RTM while producing similar results.

Full waveform inversion (FWI) in the time domain has limitations due to large computing time and memory requirements. Some studies have addressed this problem by using machine learning techniques. Most FWI studies using machine learning directly estimate the subsurface velocity structure by training the seismic data generated through various synthetic models to obtain the subsurface velocity structure. In this study, we propose a method to convert the common midpoint (CMP) gather to zero-offset data at a CMP location using a convolutional neural network (CNN) to increase the computing efficiency for the FWI. As the training data, we use synthetic data generated by the seismic exploration geometry and the source signature of field data. Since the proposed method performs FWI using the converted zero-offset data, it can be performed more efficiently than FWI using the existing multichannel data. However, it is difficult to apply a seismic exploration geometry and a source signature that has not been used for training. To verify the proposed method, it is applied to a synthetic model not used for learning as well as focused field data. It is confirmed that a proper subsurface velocity structure was obtained.

We propose seismic elastic full-waveform inversion (EFWI) using both the multiparametric approximate Hessian and the discrete cosine transform (DCT). EFWI is a promising technology for obtaining physical parameters. EFWI using the multiparametric approximate Hessian can suppress crosstalk artifacts for each physical parameter, but the computational burden increases as the number of physical parameters considered increases. In this study, to reduce the computational burden, DCT was used to compress the model parameters and accelerate EFWI in the truncated DCT domain. EFWI using DCT is possible to increase the calculation efficiency by reducing the total number of unknown variables. We implemented EFWI using monoparametric Hessian and multiparametric Hessian, which reduced the computational burden using DCT and conducted tests through numerical experiments using each Hessian. In numerical result, despite using a small number of DCT coefficients, we confirmed that contamination due to crosstalk artifacts was suppressed when EFWI was performed using the multiparametric approximate Hessian.

Full-waveform inversion (FWI) is a useful tool for estimating the properties of subsurface media, but it has the disadvantage of incurring considerable computing costs. Therefore, we develop a method for reducing the computing memory requirement in the time domain of FWI that uses a new approach with the discrete cosine transform (DCT) and the Nyquist sampling theorem. Our method enables accurate reconstruction of a wavefield with several optimal DCT coefficients derived by considering the maximum wavenumber component of the virtual source wavefield and the maximum frequency of the source wavelet. Consequently, virtual source wavefields can be stored in a computer’s memory rather than on a disk. Our FWI algorithm is compared with existing compression methods that use DCT and wavelet transform in a numerical example. The results of this comparison demonstrate that our algorithm provides significantly improved computational efficiency.

This study aims to detect the bioacoustics signal in the underwater soundscape, specifically those produced by snapping shrimp, using adaptive iterative transfer learning. The proposed network is initially trained with pre-classified snapping shrimp sounds and Gaussian noise, then applied to classify and remove snapping-free noise from field data. This separated ambient noise is subsequently used for transfer learning. This process was iterated to distinguish more effectively between ambient noise and snapping shrimp sounds characteristics, resulting in improved classification. Through iterative transfer learning, significant improvements in precision and recall were observed. The application to field data confirmed that the trained network could detect signals that were difficult to identify using existing threshold classification methods. Furthermore, it was found that the rate of false detection decreased, and detection probability improved with each stage. This research demonstrates that incorporating the noise characteristics of field data into the trained network via iterative transfer learning can generate more realistic training data. The proposed network can successfully detect signals that are challenging to identify using existing threshold classification methods.