Machine Learning techniques with X-ray microcomputed tomography to segment mineral phases in a Marcellus and a Mancos shale.

Abstract:

Mineral segmentation of X-ray computed tomography images (X-ray CT) of porous rock is increasingly desired to enhance understanding of petrophysical and hydrological properties of porous media, particularly in reactive systems. This analysis is favorable over traditional means of characterization due to its non-destructive process and its ability to analyze the three-dimensional interior of a sample. This study investigates the potential of combining machine learning (ML) methods with different image filtering techniques for accurate mineral segmentation in X-ray CT images of a Marcellus and a Mancos shale, using Scanning Electron Microscopy (SEM) Energy Dispersive X-Ray Spectroscopy (EDS) and backscattered electrons (BSE) images as ground truth. This work first labels known components in the X-ray CT images from the segmented SEM images where 80% of data were selected randomly for training and the remaining 20% were considered as testing data. Next, the image modalities produced from applying image filtering techniques to X-ray CT images (e.g. canny edge, sobel) were evaluated before post-processing of data including a cleaning data phase using non-local mean denoising to better train and evaluate the ML algorithms in the last step. Several different ML algorithms such as K-means, Fuzzy C-means (FCM), Self-Organized Map (SOM), Feed Forward Artificial Neural Networks (FNN), Random Forest (RF) and Convolutional Neural Network (CNN) will be applied on the stacked extracted features. Their respective clustering and classification performance were compared. The preliminary results show the segmented image obtained from non-local mean denoising following with K-mean clustering has higher accuracy compared to a manually segmented image. It is anticipated RF will have the best accuracy among all applied methods due to its capability to handle imbalance datasets and data scarcity. In comparation, CNN is promising when abundant data is available. The resulting CT-driven ML-based mineral mapper approach will identify minerals using only X-ray CT data. The value of this approach comes from the fact that using X-ray CT data enables non-destructive analysis and it is relatively fast to collect and interpret whereas the SEM data are inherently destructive, may take longer to collect and require longer to interpret.