Examples of multi-attribute, neural network-based seismic object detection
P. De Groot, H. Ligtenberg, T. Oldenziel, D. Connolly, P. Meldahl, 2004. "Examples of multi-attribute, neural network-based seismic object detection", 3D Seismic Technology: Application to the Exploration of Sedimentary Basins, Richard J. Davies, Joseph A. Cartwright, Simon A. Stewart, Mark Lappin, John R. Underhill
Download citation file:
Certain seismic objects, like faults and gas chimneys, are often difficult to delineate using conventional attribute analysis. Many attributes contain useful information about the target object but each new attribute provides a new and different view of the data. The challenge is to find the optimal attribute for a specific interpretation. In this paper the optimal attribute is found with a pattern recognition approach based on multi-dimensional/multi-attributes and neural network modelling. Multi-dimensional attributes, as opposed to point attributes, can provide the spatial information on the seismic objects. The role of the neural network is to classify the input attributes into two or more output classes. Neural networks are trained on seismic attributes extracted at representative example locations that are manually picked by a seismic interpreter. This approach is a form of supervised learning in which the network learns to recognize certain seismic responses associated with the identified target objects. Application of the trained network yields an ‘object probability’ cube for the target object. Essentially, the neural network can target any seismic or geological feature requiring detailed analysis. In this paper the method is described and examples are shown of gas chimneys, faults, salt domes and 4D anomalies. Some interpretation aspects are discussed.