Explainable machine learning models of major crop traits from satellite-monitored continent-wide field trial data

Area of Science:

• Agricultural Science
• Data Science
• Machine Learning

Background:

• Major food crops, primarily grasses, lack accessible biological data hindering yield and fitness trait research.
• Understanding crop performance is crucial for global food security.

Purpose of the Study:

• To develop robust, cross-continent yield prediction models using machine learning.
• To interrogate these models to uncover key drivers of crop behavior and interactions.

Main Methods:

• Assembled a continent-wide database of field experiments over 10 years.
• Utilized machine-phenotyped populations of ten major crop species.
• Trained ensemble machine learning models using diverse variables (weather, soil, sensor, satellite, management data).

Main Results:

• Achieved robust cross-continent yield prediction models with R² > 0.8.
• Identified key drivers and complex interactions influencing crop yield and agronomic traits.
• Demonstrated the capacity of machine learning to interrogate large datasets and generate testable outputs.

Conclusions:

• Machine learning offers a powerful tool for analyzing complex agricultural data.
• This approach can predict crop behavior and inform strategies for improving crop yield and fitness.
• Data-driven insights are essential for the future of food production.