NEURONAL NETWORKS TO DESCRIBE EPIDEMICS OF COCOA WITCHES’ BROOM

Abstract

Artificial Neural networks (ANN) were evaluated as tools to describe epidemics of cocoa’s witche's broom and as a potential method to forecast the disease. The ANN were built with data collected in Altamira-PA-Brazil, between January 1986 and December 1987, and were compared by regression analysis. The variables studied were basidiocarp production, disease intensity, and 16 climatic variables. Seven climatic variables were selected at 1 to 10 weeks before basidiocarp production and 11 variables at the 8th and 9th weeks before evaluation of disease intensity. Temporal series were also analyzed. A total of 37 regression models were tested and 100 ANN built. Neuronal networks could forecast disease intensity more efficiently than regression equations. The best ANN used 11 climatic variables, in the 9th week before disease occurrence.  The best ANN, with two intermediary layers of artificial neurons, and regression equation to describe basidiocarp production included the variable rainfall duration, in hours.