Soybean rust is the most aggressive soybean disease in Brazil. Despite its epidemiology is known, there are few studies about factors that cause it based on field data. This paper aimed to report influence of weather variables on rust occurrence using the decision tree technique. The models were developed based on disease detection dataset during harvests (2007/08 to 2010/11), temperature and rainfall variables at varied time windows prior to disease detection. For each disease "occurrence" record, a corresponding "non-occurrence" was generated based on the assumption that disease was not present at the thirtieth day prior to the report date, due to unfavorable weather conditions. The training set for modeling consisted of 45 rainfall and temperature variables and 12,591 records. The chosen predictive model resulted in a decision tree with approximately 78% of accuracy and 108 rules, determined by cross-validation. The interpreted model, with 28 rules, considered the temperature variables as more important, of which temperatures below 15 °C and above 30 °C were related to events of non-occurrence, while temperatures within the favorable range have been associated with events of occurrence, showing consistency with the literature.
data mining; plant disease forecast; epidemiology; decision support systems
