Deep learning based identification of carbonate rock components in core images

Identification of constituent grains in carbonate rocks is primarily a qualitative skill requiring specialist experience. A carbonate sedimentologist must be able to distinguish between various grains of different ages, preserved in differing alteration stages, and cut in random orientations across core sections. Recent studies have demonstrated the effectiveness of machine learning in classifying lithofacies from thin section, core and seismic images, with faster analysis times and reduction of natural biases.  In this study, we explore the application and limitations of convolutional neural network (CNN) based object detection frameworks to identify and quantify multiple types of carbonate grains within close-up core images. Nearly 400 images of carbonate cores we compiled of high-resolution core images from three ODP and IODP expeditions. Over 9,000 individual carbonate components of 11 different classes were manually labelled from this dataset. Using transfer learning, we evaluate one-stage (YOLO v3) and two-stage (Faster R-CNN) detectors under different feature extractors (Darknet and Inception-ResNet-v2). Despite the current popularity of one-stage detectors, our results show Faster R-CNN with Inception-ResNet-v2 backbone provides the most robust performance, achieving nearly 0.8 mean average precision (mAP). Furthermore, we extend the approach by deploying the trained model to ODP Leg 194 Sites 1196 and 1190, developing a performance comparison with human interpretation.