AE—Automation and Emerging Technologies: Neural Network Prediction of Maize Yield using Alternative Data Coding Algorithms

Backpropagation neural networks with five data coding schemes were used to predict maize yield at three scales in east-central Indiana of the Midwest USA, using 1901–1996 local crop-stage weather data and yield data from farm, county, and state levels. Input data included precipitation and air temperature during maize reproductive (R) stages R1 (silking) to R5 (denting of kernels), the year, and, for some nets, the scale of yield data. The five coding schemes were maximum value, maximum and minimum value, logarithm, thermometer (powers of 10), and binary (powers of 2). Root mean squared error over a testing set was determined at farm, county, and state scales. The best version of the network was maximum and minimum value coded and gave a root mean squared error of 10·5% overall (8·6% farm, 12·5% county, 9·0% state yield). The prediction error among the five coding types ranged from 10·5 to 46·9% for the best net of each type. Neural net software usually has a default coding scheme, which is used without considering an alternative. The results of this study suggested that the data coding method had a significant effect on neural net performance, and that sensitivity testing of data representation should be performed when constructing neural nets. The study also confirmed the usefulness of neural nets for yield prediction from simple data sets.