Modeling large-scale flood inundation requires weeks of calculations using complex fluid software. The state-of-the-art in operational hydraulic modeling does not currently allow flood real-time forecasting fields. Data driven models have small computational costs and fast computation times and may be useful to overcome this problem. In this paper, we propose a new modeling approach based on a coupled of Hydrodynamics finite element model and Multi-headed Deep convolutional neural network (MH-CNN) with rain precipitations as input to forecast rapidly the water depth reached in large floodplain with few hours-ahead. For this purpose, one first builds a database containing different simulations of the physical model according to several rain precipitation scenarios (historic and synthetic). The multi-headed convolutional neural network is then trained using the constructed database to predict water depths. The pre-trained model is applied successfully to simulate the real July 2014 flood inundation in an 870 km2 area of La Nive watershed in the south west of France. Because rain precipitation forecast data is more accessible than discharge one, this approach offers great potential for real-time flood modelling for ungauged large-scale territories, which represent a large part of floodplain in the world.