Adaptive neuro-fuzzy inference systems for epidemiological analysis of soybean rust

Abstract

The objective of this work was to develop and to evaluate adaptive neuro-fuzzy inference systems as methodology to describe the severity of soybean rust (Phakopsora pachyrhizi) monocyclic process in soybean [Glycine max (L.) Merr.], under effects of leaf wetness, temperature, and days after fungi inoculation. The experiment was conducted in growth chambers with mean air temperatures of 15, 20, 25 and 30 °C and leaf wetness periods of 6, 12, 18 and 24 h. The plants were inoculated by spraying a suspension of P. pachyrhizi inoculum at concentration of 104 uredinospore mL−1. A disease assessment key was adopted for estimate amounts of soybean rust at 0, 6, 9, 12 and 15 days after fungi inoculation. A hybrid neural network training with 3 and 3000 epochs was applied to disease severity data for optimization of fuzzy system parameters used to describe the severity of soybean rust based on leaf wetness, temperature and days after fungi inoculation. Higher accuracy and precision of the neuro-fuzzy systems estimates were obtained after training with 3000 epochs. Nevertheless, training with 3 epochs produced smoother estimates. The neuro-fuzzy systems enabled to describe the severity of soybean rust monocyclic process under effects of leaf wetness, mean air temperature and days after fungi inoculation and was better applied for Conquista cultivar, followed by Savana and Suprema cultivars. Higher soybean rust severity was verified under temperatures among 20 °C and 25 °C, leaf wetness above 6 h, with higher values above 10 h, and 15 days after fungi inoculation. Temperatures near 15 °C increased the latent period of the disease but not inhibited its development after 10 days of fungi inoculation.