On a Novel Methodology for South Atlantic Convergence Zone Events Classification Using a Multimodal Convolutional Neural Network

The precipitation in the Southwest and Midwest regions of Brazil is strongly influenced by the South Atlantic Convergence Zone (SACZ). However, predicting and detecting this phenomenon is a difficult task. It is typical in Brazil, and existing world circulation model tools have limited accuracy for its prediction. Therefore, a discrete valued bi-dimensional singular transform is proposed to be applied in a region of interest of the SACZ occurrence to extract phase-coherent structures from data. With these manipulations, a novel architecture of a multimodal convolutional neural network was conditioned for the classification of SACZ day by day. The proposed methodology achieved 93% accuracy considering a day-by-day classification for 2012.