A Deep Learning Approach for Automating the Thin-section Petrography of the Carbonate Rocks for Effective Reservoir Characterization

This paper summarizes the work undertaken to automate the process of reservoir characterization with the help of deep learning. The general objective of the work is to design a semiautomated process for effective rock characterization of the carbonate rocks with the help of image segmentation models. The detailed process of rock sample selection, field, and laboratory work to prepare thin sections from cores cut out of the rock samples is outlined. The problem statement is approached using two ways: classification approach using convolutional neural network for identifying the pore spaces in thin-section images and segmentation approach using U-Net for identification of spatial location of the pore spaces and to quantify them. Results of both models are compared and validated using standard models, namely transfer learning models and Trainable Weka Segmentation. Models built using both approaches yielded excellent results achieving training and testing accuracy of more than 97%. The limitations of the models include inability to effectively separate the pore edges from the background and tendency of the models to misclassify the non-pore objects having similar pixel intensity as that of pores.