Machine learning models for crude protein prediction in Tamani grass pastures

Abstract

Understanding forage quality is essential for meeting animal demands and optimizing production. This study aimed to: (i) test the applicability of machine learning models with tabular data such as climate variables, light interception (LI), nitrogen dose (N dose), interval between grazing (GI), and pre- (HPRE) and post-grazing height (HPOST) to predict leaf crude protein (CP) content of tamani grass pastures; (ii) identify which variables contribute most to CP prediction. A set of 90 instances was used with 80% for training and validation and 20% for testing. The hyperparameters were adjusted with grid-search on the training set. We tested Linear Regression (LR), Multilayer Perceptron (MLP), Decision Trees (DT), Random Forest (RF), and XGBoost. The MLP (r = 0.75, R2 = 44.18%, MAE = 1.55), RF (r = 0.78, R2 = 49.07%, MAE = 1.59) and XGBoost (r = 0.78, R2 = 56.65% MAE = 1.45) models presented the best prediction results (p < 0.001). The variables most important in predicting CP content were GI, followed by N dose, HPRE and HPOST. XGBoost outperformed other tested models (p < 0.001). Tabular data, including N dose, GI, HPRE, HPOST, LI, and climatic variables, is a viable alternative for predicting CP. In conclusion, the results of this study suggest that management practices may have a greater influence on the chemical composition of Tamani grass than environmental conditions, although further research with larger and more diverse datasets is needed to confirm these findings. Link to the API: https://github.com/GabrielaAquino93/Project-BiomassCalculator.