Seismic Facies Analysis Using Generative Topographic Mapping
Seismic facies analysis is commonly carried out by classifying seismic waveforms based on their shapes in an interval of interest. It is also carried out by using different seismic attributes, reducing the dimensionality of the input data volumes using Kohonen's self-organizing maps (SOM), and organizing it into clusters on a 2D map. Such methods are computationally fast and inexpensive. However, they have shortcomings in that there is no definite criteria for selection of a search radius and the learning rate, as these are parameters dependent on the input data. In addition, there is no cost function that is defined and optimized and so usually the method is deficient in providing a measure of confidence that could be assigned to the results. Generative topographic mapping (GTM) has been shown to address the shortcomings of the SOM method and has been suggested as an alternative to it. GTM analysis does a nonlinear dimension reduction in latent space, and provides probabilistic representation of the data. We demonstrate the application of GTM analysis to a 3D seismic volume from central Alberta, Canada, where we focus on the Mannville channels at a depth of 1150 to 1230 m that are filled with interbedded units of shale and sandstone. On the 3D seismic volume, these channels show up at a mean time of 1000 ms plus or minus 50 ms. We first generate different seismic attributes and then using the sweetness, GLCM-energy, GLCM-entropy, GLCM-homogeneity, peak frequency, peak magnitude, coherence and impedance attributes we derive GTM1 and GTM2 outputs. These attributes provided the cluster locations along the two axes in the latent space to be used in the crossplotting that follows. Breaking the 2D latent space into two components allows us to use modern interactive crossplotting tools. While GTM1 shows the definition of the edges very well for the channels, GTM2 exhibits the complete definition of the channels along with their fill in red and blue. We show that the performance of GTM analysis is more encouraging than the simplistic waveform classification or the SOM multiattribute approach. We expect that by using constrained GTM analysis with the help of well log data, the facies patterns we have derived using the unconstrained GTM method used here would be further tightened and made more distinct.
AAPG Datapages/Search and Discovery Article #90216 ©2015 AAPG Annual Convention and Exhibition, Denver, CO., May 31 - June 3, 2015