Accurate short-term flood forecasting is essential to reduce the risks associated with extreme rainfall events in vulnerable river regions. Although some existing deep learning studies target precipitation forecasting or stream flow discharge prediction, direct prediction of water levels remains almost underexplored, despite being the most critical variable for flood warning. This study addresses this gap by investigating water level prediction at gauging stations across Piedmont, Italy, using a multi-modal deep learning framework that combines hydrometric station data and radar imagery for 30 minutes to 6 hours lead times. Inspired by late fusion approaches from multi-modal forecasting, multiple architectures are compared based on LSTM, MLP, 1D-CNN, 3D-CNN and combined architectures.