<p>Flooding has become a persistent and intensifying threat in Bangladesh, causing widespread damage to infrastructure and affecting large portions of the population each year. The increasing frequency and intensity of these events underscore the need for advanced methods to assess and predict flood severity effectively. This study aims to develop a robust machine learning model for accurately classifying flood severity in both multi-class and binary formats, specifically addressing imbalanced data challenges by utilizing historical weather and water level data. A systematic approach was employed, beginning with extensive data preprocessing to ensure quality and consistency. The dataset was then prepared in multiple formats (multi-class and binary) to capture different aspects of flood severity classification. To tackle class imbalance, the Synthetic Minority Oversampling Technique (SMOTE) was applied to each format, enhancing model reliability. Multiple classification models were evaluated, including individual classifiers and ensemble techniques, with the stacking ensemble emerging as the top performer. This model achieved accuracies of 98.62% for multi-class and 98.87% for binary classification before SMOTE, improving to 99.89% and 99.14%, respectively, after applying SMOTE. These findings demonstrate the model's potential as an effective tool for flood severity prediction, with significant implications for enhanced risk management and disaster response strategies. Future research will focus on deploying this model for real-time flood alerts, aiming to bolster resilience and preparedness in flood-prone communities.</p>