Detecção de secas e visualização de padrões climáticos com aprendizado de máquina

Abstract

Climate change has serious implications for the occurrence of extreme events, with significant social, economic and environmental effects. Predicting drought events is essential for efficient planning of water resources, as well as for the adoption of preventive measures. Also, understanding the climatic patterns that result in dry years can be important for decision makers. In this study, we propose a drought classification model using the machine learning technique called support vector machine. The model was applied to the state of Ceará, which is often affected with long and severe droughts. The explanatory variables consist of global gridded temperature data, which allowed the identification of the regions that most influence the occurrence of droughts in Ceará. The model presented an accuracy of 77%, i.e., 47 out of the 61 years evaluated by the model were correctly classified as dry or non-dry years. The evaluation of the weights attributed by the model confirms the influence of the surface temperature of the Pacific Ocean and the South Atlantic on the drought regime in Ceará.