Flood Prediction in the Musi Catchment Region of Telangana Using Machine Learning and Hybrid Models

Abstract

Flooding is a common natural disaster that causes severe damage and loss. The aim of flood prediction is to take preventive measures and reduce the damage caused by floods. Flood occurrence is influenced by multiple meteorological factors, rather than rainfall alone. This study used multiparameter weather data, including precipitation, humidity, temperature, surface wind, and sea level pressure, to understand the atmospheric conditions. Historical and forecast weather data for the Musi catchment region were collected from meteorological sources, such as the Coupled Model Intercomparison Project Phase 6 (CMIP6) and Aphrodite data repositories. The data from 2015 to 2049 included multiple environmental parameters.  Seven model configurations were developed as four ensemble classifiers (Random Forest, XGBoost, LightGBM, and CatBoost), a two-layer Long Short-Term Memory (LSTM) network, and two hybrid architectures combining LSTM temporal encoding with XGBoost and CatBoost. Hyperparameter optimization used the Optuna framework with Tree-structured Parzen Estimation (TPE), which was applied uniformly across all models. Random Forest achieved the highest accuracy of 95.59% with an F1-score of 85.41%, whereas both hybrid models exceeded 94% accuracy, substantially outperforming. The model performance was evaluated based on accuracy, precision, recall, F1-score, ROC-AUC, and per-class confusion matrices. It also supports area-wise flood risk mapping.