Seismic waveform classification based on Kohonen 3D neural networks with RGB visualization

Seismic trace segments in a 3D volume over a reservoir interval are classified into seismic facies units via neural network trace analysis. Unsupervised classification is carried out in two stages: 1) Training, whereby typical (average) objects of each class are estimated and 2) Classification stage whereby all study objects are assigned to a certain class, based on a minimum similarity to a typical object of this class. Input parameters for the algorithm are: the number of classes, the size of the vertical segment, the investigated time window and the colour scheme applied. Unsupervised classification is fairly rapid and several software packages are available for this purpose. In contrast, a supervised workflow is more demanding yet facilitates interpretation of results. In addition, supervised classification and calibration permit probabilistic uncertainty analysis. An example of a non-supervised classification scheme is shown and the main advantages of supervised partitioning are discussed.