FLOOD PREDICTION USING MACHINE LEARNING

Floods are destructive and frequent natural disasters. Because of this, machine learning models have been developed in an attempt to predict flooding. Furthermore, this project aims to review a variety of methods such as Long Short-Term Memory (LSTM), LightGBM, Multilayer Perceptron, Support Vector Machine, and Random Forests in their ability to predict floods using a multivariate dataset of historical flood data from Bangladesh (1949-2014) and a time-series dataset for the Minnesota River (2019-2025). The performance metrics of interest for this project were accuracy, precision, recall, F1-Score, Mean Square Error (MSE) and its root (RMSE), Nash-Sutcliffe Efficiency (NSE), and Kling-Gupta Efficiency (KGE). In addition, confusion matrices and ROC curves were developed in order to judge model performance. From this project, the LightGBM model worked best for the Bangladesh data while the LSTM worked best for the time-series data. In addition, the most important features for the LightGBM model were rainfall, recording location, and year.