An explainable and transferable deep learning framework for spatiotemporal urban flood prediction by integrating...

Area of Science:

• Environmental Science
• Artificial Intelligence
• Hydrology

Background:

• Urban flooding is a growing threat due to climate change and urbanization.
• Current prediction models lack physical plausibility and transferability in complex urban settings.

Purpose of the Study:

• To introduce ViTUN, a hybrid deep learning framework combining Vision Transformer and U-Net.
• To capture spatiotemporal flood propagation characteristics under diverse conditions.
• To improve the accuracy and transferability of urban flood prediction models.

Main Methods:

• Developed ViTUN, a hybrid deep learning framework integrating Vision Transformer and U-Net.
• Trained and evaluated ViTUN using hydrodynamic simulation data for urban inundation in Yueyang, China.
• Utilized Grad-CAM for model interpretability analysis.

Main Results:

• ViTUN significantly outperformed U-Net in urban flood water depth prediction (e.g., 10.2% higher CSI, 53.7% lower MAE).
• ViTUN demonstrated strong transferability to untrained regions, achieving an average R² of 0.940.
• Interpretability analysis showed ViTUN focuses on critical inundation areas and urban features.

Conclusions:

• ViTUN offers a fast, interpretable, and transferable solution for urban flood prediction.
• The model has strong potential for real-time early warning and effective flood risk management.
• ViTUN advances data-driven approaches in hydrological hazard assessment.