How far can a model stand in for a measurement, and how should its output change a breeding decision?
Models are fitted to trial and sensor data to predict traits a breeder would otherwise measure by hand, and their accuracy is checked on data the model has not seen. Two short reviews set out how prediction, simulation and decision analysis can inform crossing, testing, selection and release decisions, one using sweet corn as the worked example and the other proposing the order in which a low-data crop such as sweetpotato could adopt these tools. This is also the subject of Ph.D. research at the Mathematical Modeling Laboratory, Kyushu University.
Predicting traits from data
In soybean, rice and wheat, the best models built on image traits predicted seed weight or biomass with R² of 0.90 or higher on held-out test data.
Regression and machine-learning models are fitted to image traits and trial records, compared against each other, and tested on genotypes, seasons or areas that were not used for training.
2023 · Frontiers in Plant Science
Scanned seed images predict soybean seed weight
Seed size traits measured from scanned RGB images of 164 soybean genotypes predicted hundred-seed weight; on held-out test data, multiple linear regression reached R² 0.94 and random forest 0.92.
2025 · Plant Stress
Image-based prediction of nitrogen use efficiency in 300 rice genotypes
Plants grown under sufficient and deficient nitrogen were imaged three times with RGB, IR and NIR sensors. Trait dissection singled out NUpE and NUEb as target traits, image-based NUEb agreed closely with the destructive measurement (R² 0.98), and three donors were selected: IC463705, Suweon and Cauvery.
Breeding digital twins
The reviews propose linking predictions to specific breeding decisions that can be checked against later results.
Short reviews on how prediction, simulation and decision analysis can inform crossing, testing, selection and release decisions under climate change, and on the order in which a low-data crop such as sweetpotato could adopt AI and genomic-selection tools.
2026 · Ecological Genetics and Genomics
A proposed definition of a breeding digital twin
A short review that frames predictive plant improvement as a breeding digital twin with five elements (a breeding referent, a state vector, an action space, a utility function and a prospective audit), using sweet corn as the worked example.