Aplicação de redes neurais artificiais para detecção de padrões de temperatura superficial do Atlântico Tropical

Abstract

The Self-Organizing Maps are an unsupervised Artificial Neural Network algorithm that is able to represent patterns within a dataset. It is based on two pillars, competition and cooperation between neurons. In Physical Oceanography, a scientist might find a huge amount of data that not always is used or analysed. Therefore, this methodology proves helpful given its ability to iterate through large datasets, extracting patterns, only dependant on the processing power available. This work aims to input a 30-year Tropical Atlantic’s Sea Surface Temperature (SST) data into a Self-Organizing Map, assess its feasability and quality and compare it with traditional methods. This same methodology will also be applied to SST anomalies. We were able to find both seasonal changes and slight interannual variabilities, which were consistent with previous findings. When plotting centroids for the relative frequency of each neuron, the heating/cooling assymetry of the Tropical Atlantic was well resolved. With regard to SST anomalies, the algorithm had a slightly weak performance, which we conclude to be due to the setted grid size, although it was able to identify patterns similar to the Empirical Orthogonal Functions method.