At the 2021 Geophysical Society of Houston-SEG Spring Symposium & Exhibition, to be held as a Webinar on April 27-28, Dr. Elive Menyoli, Emerson Business Development Manager P&I, will present a paper, Wavefield Separation Via Principal Component Analysis and Deep Learning in the Local Angle Domain (co-authored by Duane Dopkin).
The recorded seismic dataset is a composite of many wavefields. The standard seismic image volume is dominated by the high-energy specular data associated principally with reflectors and fault planes. Consequently, lower energy wavefields associated with stratigraphic pinchouts, reefs, karst edges and small faults are often lost in the standard processing and imaging process.
This presentation shows an evolution of full-azimuth imaging technology, performed in the Local Angle Domain, for characterizing subsurface features from migrated seismic data. The system decomposes the recorded seismic wavefield, in-situ at the subsurface image points, into full-azimuth reflectivity and directivity components comprised of thousands of dips and azimuths. In the directivity gathers, different traces may contain energy from different features in the subsurface. We will demonstrate the use of Principal Component Analysis (PCA) with its inherent data reduction, to derive the principal components of the different energies contained in the decomposed wavefield. PCA measures are performed in local windows around individual depth slices and all directivity bins within the directional gathers.
The next stage involves using the power of convolutional neural network (deep learning) to train and classify these principal component directivity wavefields into geological features, such as reflectors, point diffractors, faults, plus other identifiable components, such as ambient noise, acquisition footprint or coherent migration “smiles”. This is a reliable method for separating these components and produce targeted images from the decomposed wavefield. The training of the deep learning algorithm use a data library containing many examples of different geometrical features, therefore increasing the credibility of the network learning process. Deep learning algorithm consist of multiple layers, where each layer contains a set of learnable filters with a small visual field of the input image. During the training process, each filter is convolved across the width and height of the input image, computing the dot product between the entries of the filter and the input, and producing an activation map of that filter. As a result, the network learn filters that activate when it detects a specific type of geometric feature at some spatial position in the input. The results reveal superior high-resolution images over previous diffraction weighted stack filters. Additionally, the deep learning approach offers significantly better time-to-results.
Dr. Elive Menyoli has over 15 years of experience in the oil and gas industry. Prior to joining Emerson, he worked at Marathon Oil and Total E&P USA, in deep water projects. Dr. Menyoli holds a MSc degree in Physics from the University of Goettingen, Germany and a PhD in Geophysics from the University of Hamburg, Germany. He has authored numerous publications in seismic imaging and interpretation, with a recent emphasis on shale resource plays.