Salt domes are one of the seismic patterns with exploration significance. This study focuses
on supervised and unsupervised machine learning (ML) and feature selection methods using
North Sea F3 block seismic data. Nine 3D dip-steered seismic attributes sensitive to chaotic
features, including geometric, edge-detection, and two GLCM (gray level co-occurrence
matrix) texture, were selected. A pickset consisting of 10,402 samples was gathered and
normalized. Unsupervised self-organizing map (SOM), supervised support vector machine
(SVM), and multi-layer perceptron (MLP) methods were applied through two different
workflows. The dimensionality reduction techniques that were involved in the workflows
included cross-plots, neighborhood components analysis (NCA), bagged decision tree, and
Laplacian score (LS). SOM was trained by selected sections and the picked dataset. The
latter elevated its performance and guided the neurons, differentiated salt, and noticeably
reduced the misclassified samples. Learning curves were plotted to show the influences of
different data populations on mean squared error (MSE). The results showed stability of
SVM performance around 0.001 MSE for varying representation set size and fluctuation of
MSE for the MLP method. For training the SVM/MLP, 35% of the pickset data was used.
MLP and SVM with 99.79% and 99.9% accuracy, respectively, on the test dataset were
applied to the sections. The supervised methods exploited the multi-attribute input and
differentiated salt from the background. This study showed the importance of feature
selection procedures and their resultant improvements in ML techniques. The workflows
implemented here can be used in the automatic detection of seismic interpretation targets.