Ensemble Machine Learning for Flood Prediction in Northern Nigeria: Comparing Baseline Classifiers and Ensemble Strategies Using Satellite-Derived Hydrometeorological Data

Abstract

Floods constitute one of the most devastating natural calamities in the world, which result in massive loss of life, destruction of infrastructure and gross economic impact. The states of Adamawa, Borno and Gombe are especially prone to regular flooding in Nigeria, but there are few reliable and data driven prediction systems in the area. The paper designed and tested an ensemble machine learning model that predicts floods based on hydrometeorological variables collected by the Nigerian Bureau of Statistics and the satellite-based Power API of NASA, which runs between 2019 and 2025. The dataset consisted of 4,383 records containing 13 original climate and environmental variables, which had been increased to 86 features with the help of temporal, lag, interaction, and polynomial feature engineering. Five machine learning classifiers such as Support Vector Machine (SVM), Random Forest, Artificial Neural Network (ANN), XGBoost, and LightGBM were trained and compared with four ensemble strategies including hard voting, soft voting, bagging and stacking. Randomized search was used as the hyperparameter optimization tool and 5-fold cross validation was used to evaluate the model performance on various metrics such as Accuracy, Precision, Recall, F1-score, AUC-ROC, and RMSE. LightGBM showed the superior overall performance with a cross validation F1-score of 0.1131 ± 0.0281. The results confirm the feasibility of satellite derived climate information to operational flood forecasting and the necessity to combat class imbalance with methods like Synthetic Minority Oversampling Technique (SMOTE) or cost sensitive learning.