Optimization of Soybean Phenological Models Using Historical Data of Breeding and Comparison of Optimization Methods

Abstract

Phenological models from sowing to flowering time were constructed for 50 cultivars and lines using historical data of soybean breeding over the past 55 years. Three methods for optimizing the parameters: Nelder-Mead (NM), swarm particle optimization (PSO), and genetic algorithm. In cross-validation, PSO was significantly more accurate than the other two methods, although the difference was small. NM was more dependent on initial values than the other two methods, and the optimization had to be repeated multiple times from different initial values to reach the global optimal solution. Although the number of records included in the historical data varied greatly among the cultivars and lines, when more than 300 records were used, the accuracy for predicting days to flowering did not deteriorate drastically. Specifically, the correlation coefficient (r) and root mean squared error (RMSE) between the predicted and observed values were ≥0.92 and ≤4.52 days, respectively. As a result of the prediction of the extra data that was not included in the historical data, the accuracy of PSO was 0.98–0.99 (r) and 3.20–4.34 days (RMSE), suggesting the usefulness of the historical data for flowering time prediction. Comparison of the optimized parameter values among the breeding centers revealed certain trends such as prolonged critical daylength at higher latitudes. The results of this study will be useful for the construction of phenological models from historical data.