Machine learning algorithms to predict the dry matter loss of stored soybean grains (Glycine max L.)

Predictive modeling is useful to estimate future events based on collected data and discipline-specific knowledge to guide decision making. Machine learning is one of the most frequently used methods to develop predictive models. There is limited literature in which machine learning techniques have been applied to soybeans' behavior during storage, and thus, the goal of this study was to test different model frameworks that estimate dry matter loss of soybeans utilizing data acquired from a dynamic grain respiration system with different levels of moisture content (12, 14, 18, and 22%, wet basis-w.b.) and temperature (25, 30, and 35 degrees C). Five different models were trained and tested with a 2,625 point data set, which was partitioned into a training set (90%) and a testing set (10%). In the training step, 10-fold cross-validation was used to choose the best hyperparameters for each of the models. All fitted models were evaluated using standard metrics of Root Mean Square Error (RMSE), coefficient of determination (R-2), dispersion of residual values, and a coefficient of performance. The Random Forest model performed best in terms of the distance between predicted and observed values, with three predicting variables. The use of predictive modeling in the recent Agriculture 4.0 scenario is a promising development for risk management and decision making in the context of grain and oilseed storage. Practical Applications Soy is one of the most important crops in the world. The dry matter loss through respiration is a major problem during storage. The results of this study can aid researchers and the soybean industry in managing storage as a risk management and decision-making tool.