Prediction of flood using ensemble machine learning methods in Bangladesh

Abstract

Flooding occurs when a water body engulfs the mainland, disrupting the regular lives of the inhabitants. Bangladesh faces a persistent risk of flooding due to its geographical location and the impacts of climate change. Annual floods consistently disrupt the ordinary rhythm of life in the country, causing the impoverished residents to lose their homes, crops, and, tragically, loved ones. This ongoing threat renders these individuals more susceptible. In our study, we gathered a comprehensive set of climate data spanning 74 years, ranging from 1949 to 2022, from the Bangladesh Meteorological Department. This research aims to alleviate the destructive impacts of flooding by employing ensemble machine learning techniques. We have used two ensemble approach Bagging and Stacking and utilize two models for each approach. Different combination of six algorithms namely Decision Tree, Random Forest, Xtreme Gradient Boosting, AdaBoost, Support Vector Machine and Logistic Regression are used to develop those four models. All our four model demonstrate robust performance for predicting flood in different regions of Bangladesh. Our highest accuracy is obtained 97.22% with the Bagging approach model where we have used Decision Tree, Random Forest, Xtreme Gradient Boosting as base classifier. The precision, recall, F1-score, and ROC-AUC for this model were respectively 0.92, 0.91, 0.92 and 0.95. We have also evaluated Matthews Correlation Coefficient (MCC) and Brier Score. Those are respectively 0.9 and 0.028. These results signify the potential of our model to play a significant role in flood forecasting, showcasing its effectiveness in predicting and mitigating the impact of floods