A machine learning modelling framework for Triticum turgidum subsp. durum Desf yield forecasting in Italy.

The forecasting of crop yield is one of the most critical research areas in crop science, which allows for the development of decision support systems, optimization of nitrogen fertilization, and food safety. Many tested modeling approaches can be differentiated according to the models and data used. The models used are traditional crop models that require data that are often difficult to measure. New modeling approaches based on artificial intelligence algorithms have proven to be of high performance, flexible, and can be tested based on available data. In this study, four independent field experiments conducted on Triticum turgidum subsp. durum Desf. in central–southern Italy were used to train a set of machine learning (ML) algorithms to predict the yield using 16 variables: fertilization, nitrogen management, pedoclimatic, and remote sensing data. Four ML algorithms were calibrated and validated over two independent sites, and a linear regression model was used as a control. The calibrated models can predict the grain yield in the two regions by using ancillary data, topsoil physical and chemical properties, multispectral drone imagery, climatic data, and nitrogen fertilizer applied at the site. Among the four ML algorithms, stochastic gradient boosting (root-mean-square error = 0.58 t ha−1) outperformed others during calibration and transferability. Nitrogen application rate, seasonal precipitation, and temperature are the most important features for predicting wheat yield.