Warning models for coffee rust (Hemileia vastatrix Berkeley & Broome) by data mining techniques

• Autor(a) principal: Girolamo Neto, Cesare
• Data de Publicação: 2014
• Outros Autores: Rodrigues, Luiz Henrique Antunes, Meira, Carlos Alberto Alves
• Tipo de documento: Artigo
• Idioma: por
• Título da fonte: Coffee Science (Online)
• Texto Completo: https://coffeescience.ufla.br/index.php/Coffeescience/article/view/687
• topic: Predictive models; random forest; support vector machines; artificial neural networks; decision trees; Alerta de doenças; florestas aleatórias; máquinas de vetores suporte; redes neurais artificiais; árvores de decisão

Coffee rust can cause severe yield losses if control measures are not taken. Warning models are capable of generating useful information regarding to the application of fungicides, decreasing economic losses and environmental impacts. The aim of this study was to develop, compare and select warning models developed by data mining techniques in order to predict the coffee rust in years of high and low fruit load. For 13 years (1998-2011), data was collected from an automatic weather station. The independent variables were 23, obtained from the weather station, and the dependent variable was the monthly progress rate for the coffee rust, which was generated by the values of disease incidence. The most important features were refined by feature selection techniques, and the modeling was performed using four data mining techniques: support vector machines, artificial neural networks, decision trees and random forests. For high fruit load years the best accuracy was 85.3% and for low fruit load years it was 88.9%. Other performance measures like recall and specificity also had high and balanced values. The warning models developed on this study provide further information for monitoring the disease on high fruit load years than other models previously developed, and also provide a possibility for the monitoring on years of low fruit load.