Discriminating Stratigraphic and Acquisition Discontinuities From Natural Fractures Using Multi-Attributes Neural Networks

Abstract

Application of Neural Networks analyses has recently attracted more attention among geoscientists trying to apply mathematical knowledge to gain more geological information from seismic data. Neural networks are one of the most efficient ways to recombine multiple input attributes and achieve a high quality extraction of a target feature or rock property from seismic data. This paper presents a new application of neural network method to combine curvature attributes, mainly most positive curvature, with different step-outs for discontinuity analysis using examples of North Sea 3D dataset (F3 block), Western Canada Sedimentary Basin, and Kandkhot (Sui Main Limestone) block, Pakistan. Curvature attribute analysis is often used to separate structure and/or stratigraphic discontinuities such as faults below seismic resolution, shear zones, channel edges and strand-plains. Classification process also allows the suppression of acquisition artifacts of the seismic data. In the curvature analysis, a critical aspect is the spatial extent of each curvature calculation. A smaller spatial extent will reveal features while the larger spatial calculations will reveal a different set of geological details. Usually, the curvature attribute highlights the acquisition footprints while the other lineaments are obscured by these curvatures. The interest of the interpreter is to have all these curvature information classified so they could be separated to investigate each type of lineament. The Most Positive Curvature attribute provides the maximum correlation to the different geological features of available 3D dataset. We generated Most Positive Curvature attributes using step-outs of 5×5, 7×7, 9×9, 11×11 and 15×15 samples in the XY domain. The afore mentioned five volumes were then classified together using un-supervised neural network. The process starts with choosing 5000 random samples between two horizons covering a slice of about 100 ms thick. These 5000 samples represent a set of various curvature values from all five volumes. The neural network approach is given 10 classes. This volume can also be known as a hybrid volume and it generates a classified output. The classified volume or Most Positive Classfied Curvature Cube (MPC3) can differentiate lineaments from the background such as acquisition footprints.

AAPG Datapages/Search and Discovery Article #90217 © 2015 International Conference & Exhibition, Melbourne, Australia, September 13-16, 2015