MLSEL: A tuned multilayer stacking ensemble learning with meta feature for flood risk prediction

While stacking ensemble methods have been widely adopted for disaster risk classification, most traditional implementations are limited to shallow single-layer architectures, lacking generalization capacity and adaptive meta-feature design. This study proposes Multilayer Stacking Ensemble Learning (MLSEL), a novel approach that addresses these limitations through three key innovations: (i) a deep three-layer stacking architecture, (ii) Meta Feature Augmentation (MFA) to enrich inter-layer representations, and (iii) automated hyperparameter optimization using Optuna to enhance meta-learner performance. The model is evaluated on a multiclass flood risk dataset consisting of 50,000 structured records. Results reveal that the proposed MLSEL particularly the configuration using XGBoost + Optuna + MFA achieves superior accuracy of 98.69%, significantly outperforming both the best-performing baseline and conventional stacking models. This research demonstrates that combining deep-layered learning, meta-feature engineering, and adaptive optimization effectively overcomes overfitting, feature redundancy, and scalability issues inherent in traditional stacking. The MLSEL framework establishes a robust foundation for accurate and reliable disaster risk prediction systems.