Hybrid AI-Based Floodplain Mapping and Early Warning Systems Using Ensemble and Deep Learning Models in Data-Scarce Regions

Abstract

Flooding in Nigeria affects, displaces and damages the livelihood of over 200,000 people every year, causing economic losses of over USD 35 million. The research proposes an AI-based inundated floodplain mapping and warning framework that employs a hybrid dataset of actual hydrometeorological observations from monitoring stations and synthetically augmented data to fill in spatial and temporal gaps. The dataset consists of 150 stations over the period 2018–2022 with 4,500 records daily data set with 12 hydro-geospatial predictors. To avoid data leakage, seven models (Random Forest, XGBoost, CatBoost, SVM, ANN, CNN, and LSTM) were stratified spatiotemporal split (70–15–15) trained with cross-validation. The results demonstrated that the LSTM method performed best, achieving 93% accuracy, and an AUC-ROC of 0.95. Similarly, it also attained a RMSE of 0.28 m and an NSE of 0.89. Comparatively, XGBoost achieved 92% accuracy with an RMSE of 0.30 m, whereas CNN achieved 91% accuracy with an RMSE of 0.32 m. Combining RF and LSTM augmented accuracy to 95% with further decrease of RMSE to 0.25m. Errors predicting flood extent ranged from -5.7% to +6.7%. The Niger Delta model’s regional validation (93%, RMSE = 0.28 m) was superior to that of the North (87%, RMSE = 0.40 m). Following a seasonal analysis, the LSTM model improved from 91% (NSE = 0.86) in the dry season to 95% (NSE = 0.91) in the wet season. Statistical test results show that performance difference between the top models is significant (p<0.05). The results reveal that hybridized data-driven AI models can effectively enhance flood prediction and early warning with robust and scalable solutions in data-scarce regions, particularly ensemble and deep learning approaches