MDRK-RNN: A Multimodal Data and Runoff Knowledge-Driven RNN for Riverine Flood Extent Forecasting

Effective flood warning methods can significantly reduce the loss of life and property caused by extreme flood events. In this paper, we propose a new multimodal data and runoff knowledge-driven recurrent neural network (MDRK- RNN) for flood extent forecasting missions in large-scale riverine flood events, induced by concentrated heavy rainfall. Specifically, the multimodal data, i.e., historical flood extent mapping results from different satellites and digital elevation model of the target area, and runoff knowledge, i.e., estimated accumulated flood flow in the target area, are simultaneously fused to generate a combined conditional data, which is important for increasing the accuracy of the proposed MDRK- RNN model. Experimental results using measured data demonstrate the effectiveness of the proposed MDRK-RNN in forecasting flood extents.