Dynamic Maize Yield Predictions Using Machine Learning on Multi-Source Data

Timely yield prediction is crucial for the agri-food supply chain as a whole. However, different stakeholders in the agri-food sector require different levels of accuracy and lead times in which a yield prediction should be available. For the producers, predictions during the growing season are essential to ensure that information is available early enough for the timely implementation of agronomic decisions, while industries can wait until later in the season to optimize their production process and increase their production traceability. In this study, we used machine learning algorithms, dynamic and static predictors, and a phenology approach to determine the time for issuing the yield prediction. In addition, the effect of data reduction was evaluated by comparing results obtained with and without principal component analysis (PCA). Gaussian process regression (GPR) was the best for predicting maize yield. Its best performance (nRMSE of 13.31%) was obtained late in the season and with the full set of predictors (vegetation indices, meteorological and soil predictors). In contrast, neural network (NNET) and support vector machines linear basis function (SVMl) achieved their best accuracy with only vegetation indices and at the tasseling phenological stage. Only slight differences in performance were observed between the algorithms considered, highlighting that the main factors influencing performance are the timing of the yield prediction and the predictors with which the machine learning algorithms are fed. Interestingly, PCA was instrumental in increasing the performances of NNET after this stage. An additional benefit of the application of PCA was the overall reduction between 12 and 30.20% in the standard deviation of the maize yield prediction performance from the leave one-year outer-loop cross-validation, depending on the feature set.