Predicting Rice Heading Date Using an Integrated Approach Combining a Machine Learning Method and a Crop Growth Model

Accurate prediction of heading date under various environmental conditions is expected to facilitate the decision-making process in cultivation management and the breeding process of new cultivars adaptable to the environment. Days to heading (DTH) is a complex trait known to be controlled by multiple genes and genotype-by-environment interactions. Crop growth models (CGMs) have been widely used to predict the phenological development of a plant in an environment; however, they usually require substantial experimental data to calibrate the parameters of the model. The parameters are mostly genotype-specific and are thus usually estimated separately for each cultivar. We propose an integrated approach that links genotype marker data with the developmental genotype-specific parameters of CGMs with a machine learning mode, and allows heading date prediction of a new genotype in a new environment. To estimate the parameters, we implemented both a frequentist approach and a Bayesian approach with the advanced Markov chain Monte-Carlo algorithm called the differential evolution adaptive metropolis and conducted the estimation using a large amount of data on heading date and environmental variables. The data comprised sowing and heading dates of 112 cultivars/lines tested at multiple locations for multiple years and the corresponding environmental variables (day length and daily temperature). We compared the predictive accuracy of DTH between the proposed approach, a CGM, and an unintegrated single machine learning model. The results showed that the extreme learning machine (one of the implemented machine learning models) was superior to the CGM for the prediction of a tested genotype in a tested location. The proposed approach outperformed the unintegrated machine learning method in the prediction of an untested genotype in an untested location. We also evaluated the potential of the proposed approach in the prediction of the distribution of DTH in segregation populations. The results showed a high correlation coefficient (ca. 0.8) of the 10th, 50th, and 90th percentiles of the observed and predicted distribution of DTH. In this study, the integration of a machine learning model and a CGM was better able to predict the heading date of a new rice cultivar in an untested potential environment.