A comparison of unsupervised learning techniques for channel detection in 3-d seismic data acquired over the Strait of Hormuz

Summary In recent years, due to increasing the size of three-dimensional (3-D) seismic data and the number of seismic attributes, it is difficult for interpreters to examine every seismic line and time slice. Pattern recognition techniques as first-hand interpretation tools are used to both address the problem of large size of seismic data and to provide an initial guidance when working on a new seismic data where previous studies and data are limited. Different types of unsupervised learning techniques have recently been used for seismic facies clustering and object detection in seismic data. Among unsupervised learning techniques k-means, self-organizing maps (SOM), generative topographic mapping, and principal component analysis (PCA) are used for facies analysis. In this study, we have applied k-means, SOM, and PCA on a 3-D seismic data volume acquired over the Strait of Hormuz to detect buried channels. Not surprisingly, the most important parameter in this study is the choice of appropriate seismic attributes. Although the PCA is not a clustering technique, it can detect channels in 3-D seismic data more efficiently than the k-means and SOM. According to the dip of the structure, the detected channels are prolonged from the west to the east and the southeast where there is a mini basin within the Mishan Formation.
Introduction One important class of machine learning tasks is the unsupervised learning. In the unsupervised learning, no labels are given to the learning algorithm, leaving it on its own to find structure in its input. The main task of this learning method is data clustering, but some different tasks such as dimensionality reduction and density estimation are belonged to this category. PCA is a dimensionality reduction technique, which can be used for better visualization of data. After explaining the geology of the study area, we discuss the learning methods, and their workflows. In the next step, we present the chosen attributes, and the learning algorithms applied to the data.
Methodology and Approaches We have used OpendTect for attribute measurements. After preparing the data, we have applied three unsupervised learning techniques of k-means, SOM, and PCA, on attributes of the 3-D seismic data volume acquired over the Strait of Hormuz. The chosen attributes in this study are spectral decomposition, curvedness, and gray-level cooccurrence matrix (GLCM) homogeneity. First, we have applied the PCA to reduce the dimension of attribute data to 2-D, and then, the k-means and SOM are applied on the data. Next, we have presented the two first principal components of attributes to the RGB color system, and consequently, we found that this method is superior than the k-means and SOM in the illumination of the channels.
Results and Conclusions Although the PCA method is not a clustering technique, it can detect channels in 3-D seismic data more efficiently than the k-means and SOM clustering methods. According to the dip of structure, these channels are prolonged from the west to the east and to the southeast where there is a mini basin within the Mishan Formation.