Spatial assessment of gross vertical reservoir heterogeneity using geostatistics and GIS-based machine-learning classifiers: A case study from the Zubair Formation, Rumaila oil field, southern Iraq

The study of oil-field reservoir heterogeneity is an important task in the oil industry as it affects waterflooding, developing injection production systems, and optimizing hydrocarbon production. In this study, vertical reservoir heterogeneity was quantified using the Lorenz statistical index, empirical Bayesian kriging, and seven machine-learning classifiers (Classification and Regression Trees, Boosted Regression Trees, Random Forest, Naïve Bayes, Logistic Regression, K-Nearest Neighbors, and Support Vector Machine with three different kernels (linear, radial, and polynomial) under the geographic information system platform. The main pay zone of the Zubair Formation in the Rumaila oil field from southern Iraq was used as a case study. The degree of heterogeneity was first quantified using the Lorenz index, and a borehole-heterogeneity inventory location map was prepared according to the determined Lorenz index. Information about five factors influencing the heterogeneity, namely, porosity, permeability, volume of shale, reservoir-unit thickness, and depth to the top of reservoir unit, was collected based on available cores, nuclear magnetic resonance log, gamma-ray logs, and drilling-information logs. Factors from these sources were interpolated to show their spatial distribution using the empirical Bayesian kriging technique. The relationship between the borehole inventory map of vertical heterogeneity and the five factors was examined using the seven machine-learning classifiers. Two statistical-error measures, namely, accuracy and Cohen's kappa, were used to verify the performance of the classifiers in both training and testing stages. Results proved that Random Forest, Support Vector Machine with radial kernel function, and Logistic Regression were the best models. The probabilities of the best performance models were then interpolated and classified into five heterogeneity zones: Very low, low, moderate, high, and very high. The high-very high classes for each of these models approximately occupy 60% of the oil field and are mainly distributed in the middle and north of the field, whereas the other classes encompass about 40% of the field and mostly occur in the south. This distribution of classes is most likely related to the distribution and complexity of former depositional environments.