Bidirectional selective state space learning with flood wave optimization for flood disaster factor classification in wireless sensor networks

Flood disaster factor classification from heterogeneous environmental data is essential for reliable early-warning and emergency-response systems. However, existing deep learning approaches commonly depend on manually designed signal decomposition pipelines that cannot be jointly optimised with the classifier. Transformer-based models can capture long-range dependencies but impose quadratic computational complexity, limiting their suitability for resource-constrained Wireless Sensor Network gateways. In addition, conventional sequence models generally use fixed temporal representations and provide limited modelling of interactions among multiple environmental variables. To address these limitations, this paper proposes WSN-FloodMamba, an end-to-end framework for nine-class flood disaster factor classification. The framework integrates an Adaptive Patch Tokeniser for modality-specific multi-resolution representation, a Bidirectional Flood-Selective State Space Model for linear-complexity temporal modelling, and Cross-Sensor Fusion Attention for learning inter-variable dependencies. A class-weighted focal loss addresses class imbalance, while Flood Wave Optimization identifies the model hyperparameters through physics-inspired exploration, exploitation, and convergence phases. Evaluation on the Kaggle Flood Prediction Dataset shows that the proposed framework achieves 98.85% accuracy, 97.62% F1-score, 96.10% Matthews Correlation Coefficient, and 99.74% macro-average AUC-ROC. Ablation experiments confirm the contribution of the individual modules. The results demonstrate the potential of WSN-FloodMamba as an efficient and integrated framework for flood-factor classification and future edge-assisted flood-monitoring applications.