Improving reservoir rock classification in heterogeneous carbonates using boosting and bagging strategies: A case study of early Triassic carbonates of coastal Fars, south Iran

An accurate reservoir characterization is a crucial task for the development of quantitative geological models and reservoir simulation. In the present research work, a novel view is presented on the reservoir characterization using the advantages of thin section image analysis and intelligent classification algorithms. The proposed methodology comprises three main steps. First, four classes of reservoir intervals are defined using a limited number of porosity and permeability values obtained from the core plugs of Kangan and Dalan formations. Then seven micro-scale features including distribution of pore types (interparticle, interaparticle, moldic, and vuggy), pore complexity, and cement distribution as well as textural characteristics are extracted from thin section images. Finally, the features extracted from each photomicrograph and its corresponding reservoir class are used as the training data for several intelligent classifiers including decision trees, discriminant analysis functions, support vector machines, K-nearest neighbor models and two ensemble algorithms, named bagging and boosting. The relationship between the micro-scale features and the reservoir classes was studied. Performance of all classifiers is evaluated using the concepts of accuracy, precision, recall, and harmonic average. The results obtained showed that the bagging decision tree delivered the best performance among the models and improved the accuracy of simple models up to 7.7% compared with the best single classifier.