Satellite and model-based data integration for crop yield estimation and interpretability in Europe

Currently, a variety of methodologies exist for monitoring croplands and estimate yield, with different levels of automation at regional and global scales. They use a diversity of data sources, mostly field observations and crop growth simulation models, to generate accurate crop yield estimations. Earth observation (EO) data is a unique source of information to monitor key variables of soil, plant and atmosphere conditions in a temporally resolved and spatially explicit manner, yet it is often not fully exploited in operational services.

We propose a machine learning approach based on Gaussian Processes (GP) to develop crop yield models that integrate the EO and model-based bio-geophysical parameters driving crop yield of main cereals (corn, barley and wheat) grown in Europe. The developed models are interpretable and allow us to detect anomalies as well as to analyze sensitivity to changing environmental conditions, such as warmer temperatures and drier soils. Our results exemplify how the synergistic use of EO data and models with interpretable ML opens up new possibilities to face present challenges in agriculture and food security.