A Machine-Learning Approach to Removing Highly Variable Surface-Waves for a Large Shallow-Water Survey Offshore Abu Dhabi

Abstract

A large, full-azimuth, shallow-water ocean-bottom node (OBN) survey offshore Abu Dhabi revealed an intriguing azimuth-dependent linear noise pattern. The mud rolls and guided waves appear to curve along surface and near-surface features. Although we have yet to fully understand the mechanism that causes this phenomenon, we need to accurately and effectively eliminate these linear noises to meet the tight processing deadline. In view of the huge amount of data involved, we implemented a data-driven machine learning (ML) technology. Mud rolls and guided waves are dispersive in nature. The standard way to remove them is to compute the dispersion curve (DC) (phase velocity as a function of frequency) from the data, model the corresponding mode, and subtract the model from the data in an adaptive way. In our case, the linear noise patterns are not only related to the azimuth angle of acquisition, but they also vary spatially, which makes removing them even more challenging. Since we cannot manually select the dispersion curves (DC) given the size of the data, we implemented an ML algorithm to classify the noise pattern and design the best noise removal operator for each category. The remaining outliers can then be resolved using customized solutions. In order to transform seismic data analysis into a pattern recognition problem, we need to generate attribute maps. In our case, one of the attribute map is the dispersion curve (DC), which has been calculated from the data. The maps are used as input to the ML stream which identifies classes by unsupervised K-means clustering analysis. Another attribute used is the root-mean- square (RMS) amplitude calculated over a fixed time window for each receiver gather. The corresponding maps are used as the input to the ML flow, which involves two-steps: first reduce the dimensionality of the input data by performing a principal component analysis (PCA), then identify classes through unsupervised K-mean clustering analysis. The generated class determines the common linear noise behavior and its spatial distribution group, thereby effectively driving the linear noise attenuation parameter setup.