Flood Discharge Prediction Based on Remote-Sensed Spatiotemporal Features Fusion and Graph Attention

Floods have brought a great threat to the life and property of human beings. Under premise strengthening flood control engineering measures following strategic thinking sustainable development, many achievements been made in forecasting recently. However, due complexity traditional lumped model distributed model, hydrologic parameter calibration process is full difficulties, leading long development cycle reasonable prediction model. Even for modern data-driven models, spatial distribution characteristics rainfall data are also not fully mined. Based on this situation, paper abstracts into graph structure data, uses remote sensing images extract elevation information, introduces attention mechanism rainfall, employs long-term short-term memory (LSTM) network fuse temporal prediction. Through well-designed experiments, effect peak value arrival time verified. Furthermore, compared with LSTM BIGRU without feature extraction, advantages spatiotemporal fusion highlighted. The specific performance that RMSE (the root means square error) R2 (coefficient determination) GA-RNN significantly improved. Finally, we conduct experiments observed ten events history target watershed. According hydrological specifications, can be evaluated as Class B