ASTUNN: An enhanced spatiotemporal uncertainty guided neural network for flood management in mountainous areas

Mountain flood disasters in rugged terrains pose significant challenges due to rapid onset, complex spatiotemporal dynamics, and data scarcity, where traditional hydrological models and pairwise graph neural networks struggle to capture multi-scale dependencies and uncertainty in propagation patterns. This study proposes the Adaptive Spatiotemporal Uncertainty-Guided Neural Network (ASTUNN), a hybrid framework that synergistically combines Bidirectional Gated Recurrent Units (BiGRU) for temporal modeling, Spherical Manifold Graph Learning (SMGL) for non-Euclidean spatial analysis, Fractional-Order Dynamic Attention (FODA) for long-memory patterns, Stochastic Variational Inference (SVI) for uncertainty quantification, and Adaptive Feature Synthesis (AFS) for multi-scale fusion. Key innovations include: (1) hyperedge-aware spatiotemporal message passing with fractional-order attention to model higher-order interactions and long-range dependencies in river networks and terrain gradients; and (2) stochastic variational uncertainty estimation to provide calibrated probabilistic forecasts and prevention capability rankings. These contributions overcome limitations of static graphs and deterministic models under rapid environmental changes. Validated on multi-source hydrological datasets from seven high-risk mountainous regions in southwest China, ASTUNN achieves an AUC-ROC of 0.947, MAE of 0.103 for prevention capability rankings, and ECE of 0.029, outperforming state-of-the-art baselines by 15–25 % while reducing false alarms by 18 % and enabling early warnings up to 48 h ahead.