A spatial-temporal graph deep learning model for urban flood nowcasting leveraging heterogeneous community features.

Area of Science:

• Environmental science
• Computer science
• Artificial intelligence

Background:

• Flood nowcasting is crucial for situational awareness during extreme weather.
• Existing models often lack the ability to integrate diverse data streams effectively.

Purpose of the Study:

• To develop and test a novel deep-learning model for urban flood nowcasting.
• To integrate physics-based and human-sensed features for improved prediction accuracy.

Main Methods:

• An attention-based spatial-temporal graph convolution network (ASTGCN) was developed.
• The model integrates real-time physics-based (rainfall, water elevation) and human-sensed (resident reports, activity) data.
• Static, physics-based dynamic, and human-sensed dynamic features were used for nowcasting flood inundation.

Main Results:

• The ASTGCN model demonstrated superior performance in nowcasting urban flood inundation at the census-tract level (precision 0.808, recall 0.891).
• Integrating heterogeneous human-sensed dynamic features significantly improved model performance compared to using only physics-based features.
• The attention mechanism effectively focused on the most influential, dynamically varying features.

Conclusions:

• The proposed ASTGCN framework shows significant promise for enhancing urban flood nowcasting.
• The integration of human-sensed data offers a valuable addition to traditional physics-based approaches.
• Further investigation with more historical data could lead to a robust predictive tool for community responders.