Seismic Attribute Redundancy Reduction Using Statistical Feature Extraction Technique

Combining seismic attributes and building a meta-attribute section is a useful stage for seismic object detection (fault, gas chimney and …), seismic facies analysis and many other applications. Many seismic attributes used in classification tasks are redundant, which yield a useless higher-dimensional feature space. In this paper, a criterion based on RDA with two sub-optimal forward and backward algorithms is presented to accomplish this problem. Then, classification based on four different MLP neural networks is done and the average error is plotted versus the ranked feature space. As a flexible method, the algorithm introduces dissimilar ranks as the training set changed by the user. This makes it possible to introduce several seismic attributes in input stage of the classification. Based on the proposed training set, the interpreter can find the most important seismic attributes for the classification.