Co-occurrence matrix in simulated seismic pattern inspection

Don and Fu (1984) had successfully inspected the metal surface using pattern classification techniques. Due to the similarity between the seismic data images and the metal surface images, several pattern classification techniques are applied to inspect the seismic patterns. Based on the co-occurrence matrix (Haralick 1973), 12 effective features, contrasts, are extracted and used to the seismic pattern recognition. Eight analyzed seismic patterns are chaotic reflections, few reflections, many reflections, bright-spot, flat-spot, pinch-out, gradual sealevel fall, and gradual sealevel rise. The classification techniques include four clustering algorithms (unsupervised), two linear classifiers, and two piecewise linear classifiers (supervised). The supervised classification is to work on all the 8 classes (32 training samples and 32 test samples). For the minimum distance classifier, four test samples are misclassified. This yields 87.5 % correct rate (28 out of 32). Linear classifier results in 84.4 % correct rate (27 out of 32). Supervised K-means plus linear classifier yields 78.1 % correct rate (25 out of 32). The local optimal piecewise linear classifier (Lee and Richard, 1984) achieves 84.4 %. The unsupervised classification results are also presented. The results are quite good in the seismic pattern inspection.