Análise e previsão de eventos críticos de precipitação com base no SPI e em redes neurais artificiais para o Estado de Pernambuco.

Abstract

The identification and classification of areas susceptible to critical events be it rainy or dry events, has become a frequent need in the current context of climate variability, esponsible for natural disasters in several countries in the world. The diagnosis based on meteorological, agricultural and hydrological impacts can be measured by climatic indices. The Standardized Precipitation Index (SPI) was developed to categorize and make the diagnostic the variability of the rainfall based on different temporal scales. The application of SPI methodology to 57 stations distributed about the state of Pernambuco, Northeastern Brazil, for the years 1963 to 2015, was able to highlight and rank the main anomalies of rainfall through its intensity and duration. Smaller scales the SPI (monthly and quarterly) indicated the start and trend of each event, the semiannual scale identified the behavior of rainy period and the annual and biennial scales it defined the strongest and most enduring events. Positive and negative events were diagnosed in the scale categories: low, moderate, severe and extreme. Were analyzed the events that occurred more widely and thus more significant. Were highlighted the critical rainfall events of 1963, 1973, 1984 and the dry events of 1993, 1998, 2012. The cluster analysis using the metric of Ward was applied to SPIs to delimit the two well-defined groups to any timescale of the SPI. The division of Pernambuco state was as follows: Group 1, from Coast to Agreste and Group 2 represents the entire Sertão. The values of the temperature anomalies of the sea surface were correlated with each SPI scale and used as input in models based on Artificial Neural Networks (ANN) to predict the variations of this index in the study area. The results showed that the model had a good forecast with the standard of behavior of the quarterly SPI scale, but did not get the same level of performance for the monthly and semi-annual scales, but the model the ANN was able to absorb the trend of the values of these scales and find a good association.