Three-dimensional seismic denoising based on deep convolutional dictionary learning

Dictionary learning (DL) has been widely used for seismic data denoising. However, it is associated with the following challenges. First, learning a dictionary from one dataset cannot be applied to another dataset and requires setting learning and denoising parameters, which is not adaptive. Second, the DL method based on sparse constraints adds sparse regularization terms to the coefficients, while seismic data only has many coefficients close to zero, which can be approximated as sparse. To overcome these challenges, we propose a seismic data denoising approach using deep convolutional dictionary learning(DCDL) that integrates the explanatory power of DL with the robust learning capacity of deep neural networks. The proposed approach replaces sparse priors with coefficient priors learned from the training dataset and system learns adaptive dictionaries for each seismic datapoint to maintain the data structure. Synthetic and field data in the experiment demonstrate that our method effectively suppresses random noise and maintains seismic data events.