Flooding in Nigeria has evolved into a recurrent humanitarian and economic crisis, with the 2024 floods alone displacing over 641, 000 individuals, destroying infrastructure, and severely impacting food security and livelihoods. In regions around the Benue River, where the Lagdo Dam has impacts, there is a need for an accurate flood risk analysis and water levels’ predictive model. The optimal capacity of the Lagdo dam basin is set at 210 m, and the designed level at 216 m, while the overall capacity and area of the submerged zone measure 7. 7 10⁹\, m3 and 586\, km2, respectively. However, over the past few decades, sediment accumulation has significantly affected the effective capacity of the dam and further decreased the remaining capacity to about 1. 6 10⁹\, m3 in 2021. To address these challenges, this study develops an integrated machine learning–based predictive framework capable of jointly forecasting flood severity, water level variations, and associated economic impacts. The proposed approach employs a multi-task learning (MTL) architecture and benchmarks its performance against single-task machine learning models, including Random Forest, XGBoost, and Logistic Regression. To enhance physical interpretability, we engineered causally informed features, including lagged hydrological drivers, monthly stressor aggregates, and exposure and resilience indices. Time-aware models were trained using historical hydrological (inflow, dam discharge, reservoir level), meteorological (daily rainfall), and socioeconomic (persons affected, agricultural and infrastructure damage) data covering the period 1982–2024. The flood severity classification component achieves a mean cross-validated accuracy of 67. 5% (±1. 2%) across five folds, with a weighted-average F1-score of 0. 662, reflecting balanced performance across severity classes. In water level forecasting, the model achieved a mean R2 score of 0. 90, signifying strong alignment between predictions and actual water levels, while the cross-validation results show a MAE of 0. 15 with minimal variation (0. 01). Similarly, in the economic impact prediction, the model achieved a MAE of 0. 16 and a R2 score of 0. 88, reflecting its strong predictive accuracy, while the cross-validation results yielded consistent performance, with a mean MAE of 0. 17 and a standard deviation of 0. 02. The results demonstrate that the MTL framework consistently outperforms single-task models in predicting flood severity and water levels, while providing competitive performance in economic impact estimation. In high-risk situations, the framework might cut emergency decision-making timescales by 25–40%, which would be a paradigm change with important ramifications for climate resilience. This research underscores the transformative potential of AI in disaster management and contributes to Nigeria’s sustainable development and resilience against climate change-induced disasters, while highlighting the need for future work on uncertainty quantification and scenario-based validation.
Woherem et al. (Mon,) studied this question.