A Comparative Study of Undersampling and Oversampling Methods for Flood Forecasting in Bangladesh using Machine Learning
Md. Asifur Rahman, Afroja Akter, Fahmida Sultana Richi, Ahmed Shoud, Tanvir Ahmed · 2023
Catastrophic floods can inflict devastating destruction on human life, agricultural lands, buildings, and economies. Bangladesh is highly vulnerable to this natural calamity, with varying levels of destruction occurring year after year. Accurately predicting flood occurrence could save Bangladesh from this destruction, but due to imbalanced data, this is nearly impossible. Machine learning models become biased towards only one class because of imbalanced data, leading to incorrect predictions. To address this issue, we employed different undersampling and oversampling techniques for data balancing and then compared the effectiveness of these techniques for predicting flood occurrence in Bangladesh using various machine learning models. The results highlight the Random Forest Classifier with Random Oversampling as the most successful combination, achieving an accuracy rate of 97.62%, precision of 96.09%, recall of 99.34%, F1-score of 97.69%, and ROC score of 97.61%. Because of Random Forest Classifier’s ensemble learning ability, feature handling, and resistance to overfitting, it outperforms other techniques when paired with Random Oversampling to solve class imbalance.