Four-Layer Spherical Self-Organized Maps Neural Networks Trained by Recirculation to Simulate Perception and Abstraction Activity-Application to Patterns of Rainfall Global Reanalysis

This work is intended to organize a big set of time series. To do that a self-organized map is implemented in four spherical layers trained by recirculation. This way tries to simulate aspects of perception and abstraction. The methodology and the fundamentals are described. About the fundamentals, both from the problem point of view and the neural aspects as brain functioning, perception and abstraction concepts, psycho genetics and grouping ideas, and from the architecture of the network, scheme of training, spherical layers of the maps and algorithms involved in the iterative training. Then, it is used to organize a big set of time series of rainfall reanalysis on grid point around the Earth to show how it functions. After removing the average from the series, the annual cycle in shape and amplitude is the main criterion for organization. It is shown how the successive layers contain more general abstractions, their representativeness around the Globe and in regional scale. It is compared with individual series in some points of grid. A possible change of behaviour is found in global scale around 1973 and with a variant in the methodology a possible change in the annual cycle the same year.