Optimization of Rainfall Intensities Classification Based on Artificial Intelligence Using Recurrent Neural Network

In this paper, a classification of precipitation intensities is performed using the recurrent neural network (RNN). The latter is used to take into account temporal information in order to verify the contribution of the previous states to future states in the prediction of precipitation classes. The structure of an RNN introduces a mechanism of memory of the previous entries which persists in the internal states of the network and can thus impact all its future exits. The RNN is learned by using the mappings between Meteosat Second Generation (MSG) data as network inputs and radar data as network outputs. With the RNN, the outputs are also recombined at the inputs. To classify a precipitation scene at time t, the results depend on the MSG inputs at time t and the classification results at time t - 1. The model was evaluated by making comparisons with radar data considered as reference data. To see the contribution of RNN, we have also implemented the ANN, and a comparison between RNN results and ANN results is performed. Thus, all the comparisons show very interesting performances in terms of good classification rate obtained by RNN.