Integrating agronomy and machine learning to generate high-resolution global maps of yield potential with local relevance

Abstract Reliable data on yield potential is crucial for identifying areas with opportunities for production improvement. Here, we integrated an agronomically robust bottom-up approach with machine learning to generate high-resolution global maps of yield potential for maize, wheat, and rice. Our machine learning metamodel leverages site-specific yield potential derived from locally evaluated crop growth simulations and gridded climate, soil, and cropping system global databases. The metamodel showed high accuracy in predicting yield potential for the three crops, but the prediction uncertainty was higher in regions where local estimates of yield potential were missing. Our work demonstrates the benefits of integrating bottom-up and machine learning methods to achieve global coverage at high spatial resolution and ensure local relevance. The novel global yield potential maps can help to identify areas with large room to increase crop yields and serve studies assessing food security, land use, and climate change from local to global levels.