Using machine learning for crop yield prediction in the past or the future.

Area of Science:

• Agronomy
• Machine Learning
• Computational Science

Background:

• Machine learning (ML) is increasingly used in agronomy for crop yield prediction using farm data.
• The impact of data partitioning strategies on ML model performance for yield forecasting remains under-explored.

Purpose of the Study:

• To investigate how predictive algorithm choice, data volume, and data partitioning strategies affect ML model performance for crop yield prediction.
• To evaluate ML models using synthetic datasets generated from biophysical crop models.

Main Methods:

• Simulated sunflower and wheat yield data using DSSAT (OilcropSun, Ceres-Wheat) for Spain (2001-2020).
• Analyzed data with regularized linear models, Random Forest, and artificial neural networks using ordered data partitioning (older data for training, newer for testing).
• Compared model performance against a baseline of average farm yield.

Main Results:

• Random Forest exhibited superior performance (RMSE 35-38%) compared to artificial neural networks (37-141%) and regularized linear models (64-65%).
• Even the best ML models offered limited improvement over the baseline prediction (RMSE 42%).
• Ordered data partitioning for training/testing was crucial for evaluating future yield prediction capabilities.

Conclusions:

• Random Forest is a suitable algorithm for yield prediction, but its practical advantage over simpler methods needs careful evaluation.
• Random data partitioning should be avoided for yield forecasting models; ordered partitioning is recommended.
• Comparing AI-driven predictions against a baseline is essential to justify the costs associated with data acquisition and model implementation.