Machine learning applied to the prediction of root architecture of soybean cultivars under two water availability conditions

Abstract

The objective of this study was to evaluate the performance of four machine learning models, as well as multitask learning, to predict soybean root variables from simpler variables, under two water availability conditions. In order to do so, 100 soybean cultivars were conducted in a greenhouse under a control condition and a stress condition. Aerial part and root variables were evaluated. The machine learning models used to predict complex root variables were artificial neural network (ANN), random forest (RF), extreme gradient boosting (EGBoost) and support vector machine (SVM). A linear model was used for comparison purposes. Multitask learning was employed for ANN and RF. In addition, feature importance was defined using RF and XGBoost algorithms. All the machine learning models performed better than the linear model. In general, SVM had the greatest potential for the prediction of most of the root variables, with better values of RMSE, MAE and R2. Dry weight of the aerial part and root volume exhibited the greatest importance in the predictions. The models developed using multitask learning performed similarly to the ones conventionally developed. Finally, it is concluded that the machine learning models evaluated can be used to predict root variables of soybean from easily measurable variables, such as dry weight of the aerial part and root volume.