Low statistical data processing for applications in Earth-space paths rain attenuation prediction by an artificial neural network

The paper discusses low statistical data processing tools used to select appropriate input-output pairs to train an artificial neural network. The input-output pairs are constituted by a satellite link's operating parameters, such as the rain rate for a specific time percentage, latitude, elevation angle, polarization angle, station height, frequency as input, and attenuation as output. After several experiments, we observed that the existence of low statistical input-output data contributed to failures in the neural network learning process. In this way, we developed an instrument to identify poor statistical data among experimental data. So, after implementation of this method, no more failures were detected during the learning process and the neural network performed well in the prediction of rain attenuation in Earth-space paths.