Predictive analytics for floods in Bangladesh: A comparative exploration of machine learning and deep learning classifiers

Abstract

Flood prediction in Bangladesh, vital for safeguarding communities and resources, faces formidable challenges due to the region's recurrent inundations. In this study, we revolutionize flood forecasting by utilizing advanced machine learning and deep learning approaches. Following rigorous feature selection, which extracts crucial factors from the enormous dataset and improves the accuracy of our models. A variety of classifiers, such as Logistic Regression, RandomForestClassifier, XGBoostClassifier, MLP, AdaBoost, ExtraTreesClassifier, SVM, LightGBM, TabNet, and TabPFN, are built on these carefully selected characteristics. LightGBM and RandomForestClassifier, with staggering accuracies of 97.76% and 97.71% respectively, emerged as the cornerstones of predictive analytics. Navigating 65 years of historical weather data, our methodology meticulously curated features and employed a diverse ensemble of classifiers, including the emerging deep learning models TabPFN and TabNet. It is strengthened by our normalization and feature selection procedures. Our research sheds light on the critical importance of precise data preparation and deliberate model choice in addition to predictive capacity.