In hydrocarbon production, more often than not, oil is produced commingled with water. As long as the water production rate is below the economic level of water/oil ratio (WOR), no water shutoff treatment is needed. Problems arise when water production rate exceeds the WOR economic level, producing no or little oil with it. Oil and gas companies set aside a lot of resources for implementing strategies to effectively manage the production of the excessive water to minimize the environmental and economic impact of the produced water.However, due to lack of proper diagnostic techniques, the water shutoff technologies are not always proficiently applied. Most of the conventional techniques used for water diagnosis are only capable of identifying the existence of excess water and cannot pinpoint the exact type and cause of the water production. A common industrial practice is to monitor the trend of changes in WOR against time to identify two types of WPMs, namely coning and channelling. Although, in specific scenarios this approach may give reasonable results, it has been demonstrated that the WOR plots are not general and there are deficiencies in the current usage of these plots.Stepping away from traditional approach, we extracted predictive data points from plots of WOR against the oil recovery factor. We considered three different scenarios of pre-water production, post-water production with static reservoir characteristics and postwater without static reservoir characteristics for investigation. Next, we used tree-based ensemble classifiers to integrate the extracted data points with a range of basic reservoir characteristics and to unleash the predictive information hidden in the integrated data.Interpretability of the generated ensemble classifiers were improved by constructing a new dataset smeared from the original dataset, and generating a depictive tree for each ensemble using a combination of the new and original datasets. To generate the depictive tree we used a new class of tree classifiers called logistic model tree (LMT). LMT combines the linear logistic regression with the classification algorithm to overcome the disadvantages associated with either method.Our results show high prediction accuracy rates of at least 90%, 93% and 82% for the three considered scenarios and easy to implement workflow. Adoption of this methodology would lead to accurate and timely management of water production saving oil and gas companies considerable time and money.3