South American Intraseasonal Oscillation: EOF and Neural Network Approaches

Abstract This study introduces four univariate regional indices to improve the representation of intraseasonal rainfall variability across South America throughout the year, focusing on Brazil. These indices are constructed using two distinct approaches: the linear Empirical Orthogonal Functions (EOF) method and the unsupervised machine‐learning Self‐Organizing Maps (SOM) technique. Both methods are applied to Outgoing Longwave Radiation (OLR) and precipitation‐filtered anomalies in the 30–90‐day band over the South American domain. Results demonstrate that regional indices provide valuable insights into intraseasonal South American rainfall variability, including phase and strength, compared to global indices of the Madden‐Julian Oscillation (MJO). Despite being computed using only the South American domain, the regional indices capture the tropical‐tropical MJO teleconnection through the zonal wavenumber‐1 structure. The diversity in amplitude and evolution of precipitation, primarily influenced by tropical‐extratropical teleconnections through Rossby wave trains, is more evident when using the non‐linear SOM index. The regional indices also accurately measure the impacts of intraseasonal variability on extreme precipitation events over Brazil. Case studies, such as the 2013/2014 summer drought episode, highlight this ability, when a deficient rainy season severely affected the Southeast Region of Brazil, impacting agricultural production and hydroelectric power generation. During this episode, the regional indices show agreement between drought periods and suppressed precipitation phases, while global indices indicate an inactive MJO phase. These findings underscore the effectiveness of regional indices in capturing intraseasonal variability, offering significant implications for extreme weather prediction and their impacts on South American water resources and socio‐economic activities.