Permeability Prediction of Carbonate Rocks Based on Digital Image Analysis and Rock Typing Using Random Forest Algorithm

The diversity and multiscale characteristics of pore types in carbonate rocks usually result in extremely complex permeability–porosity relationships. Clarifying the main controlling factors of permeability and their response mechanisms is essential for improving permeability prediction. In this study, a digital image analysis (DIA) framework was developed to reveal the variation trend of permeability with pore structure parameters, and a rock-typing method consisting of flow zone indicator (FZI) and discrete rock type clustering technology (DRT) was applied to the carbonate samples to establish the binary permeability–porosity model. Results showed that permeability was approximately positively correlated with pore equivalent diameter, dominant pore size, and γ and negatively correlated with perimeter over area and fractal dimension. The commonly used multivariate linear regression (MLR) method failed to describe the response relationship between permeability and pore structure parameters. However, different discrete rock types exhibited special permeability–porosity relationships. Analysis showed that different discrete rock types were mainly controlled by the product of shape factor and the square of tortuosity. Further, three permeability prediction schemes were designed using the BP neural network and random forest algorithm based on the pore structure parameters derived from the thin sections. Compared with the direct prediction models trained by the optimized BP neural network and random forest algorithm, the indirect permeability prediction method based on FZI prediction showed better generalization ability with the highest R 2 of 0.936, demonstrating that the combination of digital image analysis, rock typing, and random forest algorithm is a robust and reliable method to realize the permeability prediction of carbonate rocks based on thin-section images.