AUTOMATED GRAIN SIZE CLASSIFICATION FROM CORE PHOTOGRAPHS IN PEGMATITES

The economic importance of pegmatites is ever-increasing due to their potential for high concentrations of critical and rare-earth minerals. Understanding the processes involved with pegmatite formation provides insight into their economic potential. To form an understanding of pegmatite formation processes at a given deposit, analysis of chemical composition, color, texture, and grain size is performed on core samples acquired during drilling.

In the presented analysis, X-ray fluorescence (XRF) data and high-resolution photographs have been acquired on core samples from a deposit including lithium-bearing pegmatites. In conjunction with geologists’ logs from the deposit, the XRF data and photographs have been used to train machine learning models to perform automated classification of lithology and pegmatite grain size class, respectively. The image-based grain size classification model is the subject of this presentation.

The image-based grain size classification algorithm uses an image feature descriptor to analyze localized portions of the core images and measure occurrences of gray-scale gradients and orientations. The distributions of gray-scale gradients and orientations are then coupled with the geologists’ grain size classification logs, and a machine learning model trained to predict the pegmatite grain size class from the image gray-scale gradients and orientations directly. Model performance is good, with approximately 70% of predicted grain size classes matching those provided by the geologists.

It was noted that when predicted grain size classes did not match that provided by the geologists, it very often landed in an immediately adjacent class, indicating that most sample images do not exhibit a single grain size class, but rather a distribution over several adjoining classes, with the most dominant predicted by the model sometimes differing from what the expert geologist had selected for the same sample. When adjacent classes are considered as also being acceptable predictions, measured model performance increased to 94%.