Predictive Modeling of Brown Rust of Sugarcane Based on Temperature and Relative Humidity in Florida

Abstract

Logistic regression models were developed from 5 years (2014 to 2018) of disease severity and weather data in an attempt to predict brown rust of sugarcane at the Everglades Research and Education Center in Belle Glade, Florida. Disease severity (percentage area of the top visible dewlap leaf covered by rust) was visually assessed in the field every 2 weeks for two varieties susceptible to brown rust. A total of 250 variables were derived from weather data for 10- to 40-day periods before each brown rust assessment day. A subset of these variables were then evaluated as potential predictors of severity of brown rust based on their individual correlation or their biological meaningfulness. Analyses of correlation and stepwise logistic regression allowed us to identify afternoon humid thermal ratio (AHTR), temperature-based duration variables, and their interaction terms as the most significant variables associated with brown rust epidemics of sugarcane in Florida. The nine best predictive models were identified based on model accuracy, sensitivity, specificity, and estimates of the prediction error. The prediction accuracy of these models ranged from 73 to 85%. Single-variable model BR2 (based on AHTR) classified 89% of the epidemic and 81% of the nonepidemic status of the disease. More than 83% of the epidemics and 81% of the nonepidemic status of sugarcane brown rust was correctly classified via multiple-variable models. These models can be used as components of a rust disease warning system to assist in the management of brown rust epidemics of sugarcane in south Florida.