Seismic Impedance Inversion Using Conditional Generative Adversarial Networks Assisted by Spectral Reconstruction-Based Data Augmentation

Deep learning-based seismic inversion methods can obtain promising results for reservoir characterization. However, training these models is difficult, especially when the labels are insufficient. To overcome this issue, we propose a spectral reconstruction-based data augmentation (SRDA) approach integrated with a semi-supervised deep learning method for seismic impedance inversion. First, we suggest an SRDA method to expand the training dataset, which only needs impedance log data. Our SRDA method is implemented by reconstructing the high-frequency spectrum of the impedance log while preserving its low-frequency information. We propose using the Dirichlet Process (DP) and Markov Chain Monte Carlo (MCMC) methods to reconstruct the high-frequency spectrum, which is then combined with the extracted low-frequency information to generate the reconstructed impedance data. Then, we propose an impedance inversion model based on conditional generative adversarial networks (cGANII), which can utilize conditional input to guide the generator in producing inversion results that better adhere to physical laws, thereby improving the accuracy of the inversion. Furthermore, we introduce a loss function based on Wasserstein divergence to stabilize the training of the cGAN. The numerical results demonstrate that our method can predict seismic impedance with higher accuracy than traditional model-based inversion and some commonly used deep learning models.