Robust local slope estimation by deep learning

Abstract

In the seismic community, the local slope is essential for various applications, including structure prediction, seislet transform, trace interpolation and denoising. The most popular method to calculate slope is the plane-wave destruction, which assumes that the seismic data can be represented by local plane waves. However, the plane-wave destruction method fails when the seismic data become very noisy. Taking random noise into consideration, we adopt the deep learning method to calculate the local slope. In the deep learning architecture, the input is noisy data and the target is an accurate slope estimated from the clean data using plane-wave destruction. The deep learning architecture includes a convolutional layer, deconvolutional layer, normalization layer and activation layer which are suitable to suppress noise and learn the nonlinear relationship between the input and the target. After training, the network is applied to other test examples to calculate local slopes. In addition, we implement the seislet transform and structure prediction applications based on the estimated slope. Both the estimated slope and the related applications indicate that the proposed method can robustly obtain the local slope from noisy seismic data, compared with the planewave destruction method.

Outlier Denoising Using a Novel Statistics-Based Mask Strategy for Compressive Sensing

“…Irregular noise occurs randomly, and has no typical frequency band and propagation direction. To improve the SNR of seismic data, many methods have been proposed, such as median filters [3,4], predictive filters [5][6][7][8], sparse transformation [9][10][11][12], rank reduction theory [13,14], empirical mode decomposition [15][16][17][18][19][20], and deep learning assistant methods [21][22][23][24][25]. Although a lot of tests demonstrate that these methods can effectively suppress random and linear noise, they are difficult to completely remove outlier noise.…”