REAL-TIME CARBONATE PETROGRAPHY WITH DEEP LEARNING

Deep Convolutional Neural Networks (DCNN) have been used extensively to automate identification and interpretation of geoscience images, including thin sections, subsurface core images and seismic facies. To date, however, no study has applied the DCNN method to the classification and identification of carbonate rocks constituents from petrographic images in real time. Here, we used over 13,000 individually labeled objects from ~4,000 carbonate petrographic images as a training dataset and compared two different object detector frameworks (one-stage and two-stage detectors). We find that DCNN-based object detection can successfully locate and classify different carbonate grains, cements, matrix and porosity in both image and real-time video streaming. Overall, the one-stage detectors (YOLOv4 and RetinaNet) perform up to 10 times faster than the two-stage detectors (Faster R-CNN and R-FCN) in conducting real-time petrographic analysis but the two-stage detectors are more precise. Although slower than the one-stage detectors, two-stage detectors are comparable to human interpretation of carbonate petrography in both speed and accuracy. Based on these results, DCNN-based object has the potential to replicate human performance in carbonate petrography with improved cost-efficiency, speed, and reproducibility than conventional petrographic analysis. Furthermore, the technique can be extended to real-time petrographic analysis, such as well-bore analysis in hydrocarbon exploration or Mars rover science investigation.