Intelligent data-driven denoising based on texture complexity

Abstract

Abstract Denoising is considered to be the most frequently technique of seismic signal processing to improve signal-to-noise ratios measured and provide a precondition for subsequent processing. Deep learning (DL), a data-driven approach, is based on statistical techniques to extract features from large amounts of training datasets to learn extremely complex functions that effectively attenuate noise, rather than assumptions about linearity, low rank or sparsity. Most of the existing strategies use all available training datasets for network training, which ignores the contribution of different training data to the results. In this case, however, the generatively huge amount of training patches inevitably increases the memory and computational overburden of DL optimization. Targeting this issue, entropy sampling strategy based on the texture complexity of the seismic data is developed to extract training patches more efficiently, reducing the training datasets while guaranteeing prediction performance. Texture complexity equates to the validity of the seismic data, with regions of higher complexity representing areas of valid signal enrichment that are sampled with higher probability. Using the proposed strategy, we can use the training dataset more efficiently. We evaluate the performance of the proposed method for denoising based on the denoising convolutional auto-encoder network. Numerical illustrations on simulation models and two real surveys show that the proposed methodology has better denoising performance with higher accuracy and efficiency compared to the conventional methods.