Drill core images provide information of the texture, structure, and mineralogy of the rock, which enables the use of drill core images to optimise downstream processes. The impact on downstream processes from particles of similar composition and mineralogy but different textures has been examined by many researchers. The application of local binary pattern (LBP) and gray-level co-occurrence matrices (GLCM) has been extensively studied by many scholars, and they have been demonstrated to be effective in lithology classification through machine learning techniques. More recently, convolutional neural networks (CNN) with transfer learning for feature extraction have gained attention and have also been proven as an effective tool. Currently, imaging-based data is primarily used in mining for inspection by geologists to support qualitative descriptions. There exists considerable scope for the use of multivariate statistical process control (MSPC) to evaluate imaging data, and an opportunity for the use of image-based data for the detection of subtle changes in rock texture and lithology.

This study builds upon the study of Gu et al.(this volume) and seeks to compare the performance of three widely used feature extraction methods: LBP, GLCM and CNN, for change detection in optical drill core images using MSPC. The results for the GLCM and CNN are consistent with the LBP analysis, and similarly infer that there may exist a major change in lithology mid-way along the length of the drill core. The existence of these two main regimes is also supported by the K-means clustering analysis which is independent of the location of each image. Moreover, the GLCM features are less sensitive to the reference data used to construct the control charts. However, these control charts do show a peak around 160 meters, like the one shown in the control charts for CNN features.