Optimization of Artificial Neural Network Architecture for the Nowcasting of Rain Rate Field From Radar Observation

The possibility of short term prediction of rain rate from radar observations using a Neural Network based approach has been investigated using several case studies of radar observations for Central-North Italy. To this aim the performances of different neural architecture and different approach for the training of neural device in operational condition are tested and compared. A simple feed forward neural network is recognized to be the best neural architecture for the specific application while the increasing of network complexity, in terms of nodes of the network, seems not to significantly affect the performances of the neural nowcasting.