Sparse Seismic Deconvolution Via Recurrent Neural Network

We propose a new efficient method to perform deconvolution of post-stacking post-migration seismic data. We employ recurrent neural networks (RNNs) to obtain super resolution reflectivity images. The network is designed and trained to take into account time and space relations. The robustness of the proposed method is experimentally validated for both synthetic data and real data with challenging structures and difficult signal-to-noise ratio (SNR) environment. We explore the system's behavior for different training and testing scenarios and discuss potential problems for future research. We show that training with synthetic data of simple structures solely can yield enhanced and detailed real data inversion results. The proposed method can be applied to large volumes of three-dimensional (3D) seismic data.