Rainfall frequency and seasonality identification through artificial neural networks

The artificial neural network technique is experimented to cope with the study of the sub-annual seasonal non-stationarity of the rainfall process. The homogeneity of the climatic signals inside each of the natural 12 monthly classes is analyzed, adopting a multilayer feed-forward network with error back-propagation, The possibility of identifying 'monthly based seasons' from only daily rainfall data is found to be quite limited. The coupling of rainfall and temperature statistics is instead confirmed to be a fundamental climatic indicator. Contrary to what is commonly expected, the season uncertainty appears higher in summer and in winter than in spring or autumn. The hypothesis of defining any monthly based pluviometric regime is however demonstrated to be generally difficult to sustain, revealing the necessity of adopting an unsupervised criterion to identify any seasonal filter of the rainfall process.