Development of a Neural Network for Soybean Rust Epidemics

The objective of this study was to develop a neural network to predict soybean rust disease severity for asingle soybean cultivar. Data available for development consisted of sequential weekly plantings of TK 5 soybean cultivarin 1980 and 1981, planted at the Asian Vegetable Research and Development Center in Taiwan. Plant and diseasedevelopment were measured weekly throughout the season for each planting date experiment. Data from 73 epidemicswere collected with a total of 577 scenarios (observations) over the two years. A total of seven inputs related to soybeanand rust progress and weather were available for development. Validation was performed for two different assumptions about the independence of the epidemics. In the firstvalidation, it was assumed that all epidemics were independent. The 73 epidemics and their associated observations ofrust severity were selected so that 70% were used for training, 15% for testing, and 15% for validation. Several differentthree-layer feedforward neural networks were developed using different combinations of inputs and hidden nodes. Thebest NN had four inputs and four hidden nodes, and gave the highest coefficient of multiple determination (R2) of 0.793for the validation scenarios. This configuration gave the lowest average error of 11.0% per observation. Cumulativeprobability of percent error between predicted and observed rust severity showed that 50% of the validation scenarioshad errors within 10% of measured values, and 80% had errors within 20%. In the second validation, it was assumed that epidemics in 1980 were independent of epidemics in 1981. Epidemics for1980 were used for training and testing, and epidemics in 1981 were used for validation. Several different three-layer feedforward networks were developed with different combinations of inputs and hidden nodes. Networks with three and fourinputs gave very similar results. The three input network gave an R2 of 0.858, average error of 9.7%, and a maximumerror of 43% for the validation scenarios. Cumulative probability of percent error between predicted and measured rustseverity showed that 54% of the validation scenarios had errors within 10%, and 84% of the scenarios had errors within20% of measured values.