With the increasing popularity of deep learning, the Convolutional Neural Network (CNN) has been extensively studied and shown to yield promising results in image analysis, as CNNs process data with a grid-based topology in order to keep spatial information. Unlike conventional neural network structures, where only one-dimensional data is accepted as input, the CNN is designed to process multidimensional data. It consists of an input layer, output layer and hidden layers. The hidden layers of the network consist of convolutional layer, pooling layers, fully connected layers and normalization layers. Using different functionalities of the layers, the CNN not only keeps the shape of input data, but also powerfully extracts the features that can present the image.
Absolute Imaging has developed technology for the application of machine-learning to the problem of velocity analysis. The process involves the compilation from historical projects of training sets consisting of semblance images together with the corresponding expert-picked velocity functions. Training semblances are separated into overlapping strips or patches and each patch is labelled with a velocity value obtained from the associated expert picks. These labelled patches are used to ‘train’, that is ‘iteratively to refine the parameters of’ a CNN model that can ultimately be used in predicting optimal NMO velocities for any semblance image given to it. Many such training patches are needed as a CNN typically involves dozens of layers and millions of initially unknown parameters. Individual network layers may consist of anything from classifiers to some form of image-processing filter. The proposed model uses the very successful VGG16 network and further improves it for the application.
With a well-established training set in place, the program can analyze new semblance images for fast and reliable velocity picking based on its past experience. Early results have shown promise for use of the technology in obtaining useable preliminary velocity fields. With continual addition to the training sets, we expect the power of the tool to increase both in reliability and scope of application.
In the future, the study can be extended by taking other machine learning schemes into account to improve other aspects of processing workflow. Remote sensing data such as seismic and well logs is usually analyzed visually by geophysicists in order to characterize geological and structural properties. This type of pattern recognition fits well into the Machine Learning (ML) and computer vision field. For example, deep learning can be used to discover and correlate stratigraphic units across multiple wells by recognizing patterns, textures, and similarities in the data. ML methods, specifically Support Vector Regression (SVR), can be developed for reconstructing seismic data from under-sampled or missing traces to facilitate 5D interpolation. Seismic noise attenuation is another common obstacle that can benefit from the strength of deep learning in adaptive denoising. Absolute Imaging is working on ﬁnding new ways of using data analysis and machine learning in current and future applications.