Water pot ability refers to the safety of drinking water, ensuring it is free from harmful contaminants, essential for health and well-being, a crucial factor for environmental monitoring. Machine learning-based system can help us to classify water pot ability by analyzing data from historical samples to identify safe and unsafe drinking sources. Previously, much of the research concentrated on chlorophyll concentration to assess water potability. In contrast, this study examined the chemical and physical property of water to evaluate its safety for consumption. The feature importance, an explainable artificial intelligence technique, provides a quantitative understanding of how factors affecting water potability impact the performance of ensemble learning (EL) models and enhances the model’s practical applicability. Both boosting and bagging-based EL models are introduced to achieve the said purpose. The number of input variable for model construction increased sequentially from 1 to 9, starting from the variable with the highest feature importance to that with the lowest. The model performance plateaued after considering five or six important features indicating that model performance can be achieved with a reduced number of features. To consolidate our observation further, synthetic minority oversampling technique is employed to remove imbalance in the dataset and to see whether the same performance is repeated with reduced number of features. The model performance for all such experiment was evaluated using typical benchmark metric like accuracy, recall, precision, and F1 measure.