agglomerative hierarchical clustering

Seismic data interpretation methods provide useful information about underground structures. Since many years ago, several methods have been developed to aim this goal. Seismic facies analysis is one of the new methods in seismic interpretations. This method can produce a classified section using reflection seismic data and/or seismic attributes. Classified sections can reveal lateral changes in seismic facies which may relate to geological facies changes. Using different pattern recognition methods, several seismic facies analysis methods have been developed in recent years. However, in this study, an agglomerative hierarchical clustering algorithm has been utilized to produce classified sections. Seismic facies is a group of data whose attributes are different from those of neighbor groups. Each attribute can extract additional information about underground. Using a single attribute makes it difficult to get more information. However, by combining several attributes in a hierarchical clustering algorithm, it is possible to interpret seismic data in a more appropriate way. In hierarchical clustering, all time samples are divided into similar clusters. At first, each sample is assigned to one cluster. Dissimilarity matrix is constructed based on a distance definition such as Euclidean distance between samples. This matrix is then used to cluster all samples in a hierarchical procedure. In each step, more similar clusters merge into a new cluster and the dissimilarity matrix is updated. Finally, all samples merge into one cluster. Before clustering it is common to perform a principal components analysis, PCA. PCA is a statistical technique to perform dimension reduction. Using PCA, we can find the directions in data with the highest variation and reduce the dimensionality of a large data set with interrelated variables without considerable loss of information. In this study, the PCA was utilized to attenuate the redundant and random noisy data. Prior to the PCA, it is necessary to normalize the data. Clustering algorithm in this study was applied to three synthetic models as well as 2D and 3D real seismic data of an oilfield, Southwest of Iran. The first model was a horizontal-layer one with lateral changes in facies. The second model was a horizontal-layer one with a normal fault which caused a movement of layers. The third model was an anticline one with lateral changes at the top of the anticline. Real seismic data from an oilfield in the Southwest of Iran was used for this study. Nine seismic attributes were calculated using the Paradigm software to extract more information from migrated seismic data. These nine attributes and the primary seismic data were normalized and entered into the PCA. Seven principal components were selected based on the PCA. These data were used to apply to clustering algorithm. Our results showed that the seismic facies analysis can provide useful information about the underground structures and lateral changes. In the cases of the first and second models, lateral facies changes were revealed for signal-to-noise ratios of up to 4 dB. Regarding the third model, the results were acceptable for signal-to-noise ratios of up to 8 dB. In addition, it was shown that defining more number of clusters could not lead to better results. By comparing 2D and 3D data clustering, it is concluded that the resolution of seismic facies in 3D clustering is quite related to 2D one.