A Similarity Contrast Framework for Unsupervised Geological Segmentation on Heterogeneous Structures

Image segmentation is an essential step in many geological applications. Before physical analyses, it is necessary to identify spatial structure from raw images. The researches in the field of petrophysics have highlighted the demand for extracting pore and grain from the micro-Computed Tomography (mCT) data. Moreover, detecting flow and dry areas plays an important role in understanding a flume system associated with autogenic dynamics. However, one challenge within existing segmentation methods is the generalization ability. Traditional image processing techniques, such as Ostu and watershed thresholding, have a limited ability to address geometrically complex structures. On the other hand, the deep learning method is heavily dependent on training images. It is time-demanding to collect sufficient labels in a specified case. The segmentation task becomes more challenging in the case of heterogeneity and non-stationarity. In this work, we present a similarity contrast framework to achieve unsupervised geological segmentation. Heterogeneity and anisotropy play an important role in our segmentation. The main idea is that two matching structures in the raw image are supposed to exhibit intensive similarity in the segmented space. First, we employ a descriptor referred to as local binary pattern (LBP) to extract geometrical features from the raw image. Based on LBP histogram and Jensen-Shannon divergence, our program creates a distance matrix between raw images. Second, a set of segmented images are analyzed by LBP. We characterize spatial structures across different scales. Third, our program carries out Mantel test to calculate the correlation between two distance matrices. The segmentation method that has a positive effect on correlation improvement is given top priority. As the first application, a three-dimensional shale model with intensive heterogeneity is employed to test our method. Geometrical similarities are computed in grayscale as well as binary images. Our method exhibits competitive performance since the key connectivity is properly preserved in the selected realization. Further application focuses on a flume experiment of a braided river system. The computing results indicate that our similarity contrast framework provides a valuable way to extract geological structures.