ANALYSIS OF DATA-DRIVEN INTERNAL MULTIPLE PREDICTION

The internal multiple prediction (IMP) algorithm analyzed in this paper is almost entirely data-driven, requiring a convolution and a crosscorrelation of the input data and information about the main internal multiple generators. The generators or generating horizons are the reflectors where the internal multiples' energy was downward reflected. There are two common approaches to applying IMP: 1) The first is the layer-stripping approach in which internal multiples are predicted starting from the shallowest generator (top-down approach) and subtracted from the input data prior to attempting the prediction using the next horizon as generator. For each generator's prediction, there is a subtraction. 2) The second approach; referred to as the non-top-down approach, predicts the multiples using one horizon at a time, but does not remove the predicted multiples from the input data prior to running the IMP algorithm with the next horizon. The first approach is in agreement with the theory behind this algorithm. The second approach still provides value; however, the same internal multiple can be predicted more than once by different horizons. These predictions have different amplitude information and opposite polarity with respect to each other. Hence, it is not always easy to deal with these internal multiple models when attempting to subtract them from the input data. I provide an analysis of the prediction of internal multiples using IMP with the different approaches.