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

The use of ML in agronomy has been increasing exponentially since the start of the century, including data-driven predictions of crop yields from farm-level information on soil, climate and management. However, little is known about the effect of data partitioning schemes on the actual performance of the models, in special when they are built for yield forecast. In this study, we explore the effect of the choice of predictive algorithm, amount of data, and data partitioning strategies on predictive performance, using synthetic datasets from biophysical crop models. We simulated sunflower and wheat data using OilcropSun and Ceres-Wheat from DSSAT for the period 2001-2020 in 5 areas of Spain. Simulations were performed in farms differing in soil depth and management. The data set of farm simulated yields was analyzed with different algorithms (regularized linear models, random forest, artificial neural networks) as a function of seasonal weather, management, and soil. The analysis was performed with Keras for neural networks and R packages for all other algorithms. Data partitioning for training and testing was performed with ordered data (i.e., older data for training, newest data for testing) in order to compare the different algorithms in their ability to predict yields in the future by extrapolating from past data. The Random Forest algorithm had a better performance (Root Mean Square Error 35-38%) than artificial neural networks (37-141%) and regularized linear models (64-65%) and was easier to execute. However, even the best models showed a limited advantage over the predictions of a sensible baseline (average yield of the farm in the training set) which showed RMSE of 42%. Errors in seasonal weather forecasting were not taken into account, so real-world performance is expected to be even closer to the baseline. Application of AI algorithms for yield prediction should always include a comparison with the best guess to evaluate if the additional cost of data required for the model compensates for the increase in predictive power. Random partitioning of data for training and validation should be avoided in models for yield forecasting. Crop models validated for the region and cultivars of interest may be used before actual data collection to establish the potential advantage as illustrated in this study.