Rye Revolution: Boosting Crop Predictions with AI

Scientists have made a significant breakthrough in crop improvement by developing new methods to enhance the accuracy of AI and machine learning models. These models are used to predict how different crop genotypes perform in various environmental conditions. One of the major challenges in crop improvement is accurately predicting environment-specific genotype performance. Researchers have been working on improving machine learning and deep neural network-based predictions by isolating genotype-by-environment interactions from main environmental effects.

For a long time, machine learning and deep neural network-based predictions have been dominated by environmental main effects rather than genotype-by-environment interaction effects. This limitation has hindered their adoption in plant breeding. A team of researchers has been exploring novel approaches to tailor machine learning and deep neural network models toward predicting environment-specific genotype differences and rankings.

The researchers introduced two new methodologies. The first one involves target-variable engineering based on linear mixed-model decompositions to isolate genotype-by-environment interaction effects. The second methodology involves a custom-loss-function implementation of the mean squared error of differences. This approach optimizes models directly for prediction of within-environment genotype differences.

These new approaches were tested using a comprehensive cross-validation framework on a dataset covering major rye growing regions worldwide. The results showed that the new approaches improved predictive abilities of machine learning and deep neural network models by 21 to 62 percent compared to classical machine learning-based yield prediction.

The researchers were also able to increase predictive accuracy for environment-specific genotype rankings by 15 percent and 9.8 percent across different cross-validation schemes over baseline genotypic main effects. By using historical weather records, they were able to predict genotype performances in future, untested years.

This work has significant practical value for plant breeders, growers, and variety testing authorities. The new approaches can help improve crop yields and quality, ultimately contributing to food security. By leveraging environmental covariates and developing tailored machine learning and deep neural network strategies, researchers can make more accurate predictions and informed decisions.