Seismic Denoising and Stacking Using Non‐Linear Principal Component Analysis

ABSTRACT Stacking seismic data is the simplest way of improving the signal‐to‐noise ratio and is accomplished by summing the seismic traces from different shot records that have a common reflection point to form a single trace. The weighted stacking methods are found to be superior to conventional stacking and result in improved‐quality seismic images. Further, weights calculated from the principal component analysis (PCA) give more robust results than the other conventional weights. PCA is widely used in seismic data processing and interpretation as it has the ability to extract useful signals based on a low‐rank decomposition method by solving an optimization problem with a low‐rank constraint. However, the PCA is a linear method, and it can map the data to linear principal components (PCs) only, causing an information loss in the seismic stack if the data are not exactly flattened during NMO correction. In order to address this problem, we propose an improved method of seismic stacking using auto‐associative neural network‐based non‐linear principal component analysis (NLPCA) method, wherein the PCs are generalized from straight lines to curves. In this method, the seismic gathers are mapped into NLPCAs using the encoder network, and then they are mapped back into the data domain by the decoder. We applied our method to synthetic data, offshore field data (case study I) and onshore field data (case study II). The NLPCA approach not only enhances seismic imaging but also reduces processing time compared to PCA‐based methods, for the same training dataset.