Seismic Impedance Inversion Based on Improved Cycle-WGAN

Abstract

The application of deep learning methods for seismic impedance inversion usually requires a large amount of labeled data to train the network, while labeled data available in practical applications is often limited, which affects the effectiveness of the relevant methods. In order to address this problem, this paper proposes one kind of deep learning method of a closed-loop cycle Wasserstein generative adversarial network (Cycle-WGAN) for seismic impedance inversion based on the combination of “data-driven and model-driven”. The method uses a small amount of labeled data and unlabeled seismic data generated by the forward modeling. It constitutes a bidirectional cycle of inversion and forward modeling through the generative and adversary networks for the inversion and generative and adversary networks for the forward modeling by convolution model, which improves the conventional GAN networks. The proposed method also introduces the Wasserstein loss function to improve the neural network’s training stability. Tested on the Marmousi model with complex structure, the proposed Cycle-WGAN network can effectively obtain the seismic impedance inversion results. Moreover, it is highly robust when seismic data are noisy and has higher accuracy than the inversion results from some conventional neural networks, such as RNN (Recurrent Neural Network), TCN (Temporal Convolutional Network, WGAN (Wasserstein Generative Adversarial Network), etc.