Seismic buffer recognition using mutual information for selecting wavelet based features

One of the first operations in a seismic signal processing system is to distinguish between valid and invalid records. Since valid signals are characterized by a combination of their time and frequency properties, wavelets are natural candidates for describing seismic features in a compact way. This paper investigates an approach to seismic pattern recognition comprising wavelet based feature extraction, feature selection based on the mutual information criterion, and neural classification based on feedforward networks. The ability of the wavelet transform to capture discriminating information from seismic data in a small number of features is compared against alternative feature reduction techniques, including statistical moments. Firstly it was determined whether wavelets could capture more informative data in an equal number of features compared with other features derived from raw data. Secondly the ability of wavelet based features, to improve upon the selection of features based on prior knowledge about class differences, was evaluated. Thirdly it was determined whether the definition of features can be optimized in a closed loop, by feeding back the results of a neural classifier to improve on the initial definition of the features. Three different variations of the wavelet transform were used to extract features: the discrete wavelet transform, using all wavelets as features; the so-called single wavelet transform, consisting of one scale only of the continuous wavelet transform and the continuous wavelet transform, using the mutual information criterion to select a relatively small set of wavelets from the time-frequency grid.