Comparative Analysis of Machine Learning Models for Predicting Forage Grass Digestibility Using Chemical Composition and Management Data

Accurate prediction of forage digestibility is essential for efficient livestock management and feed formulation. This study evaluated the performance of machine learning (ML) models to estimate the in vitro digestibility of leaf and stem components of Brachiaria hybrid cv. Ipyporã, using three datasets composed of pasture management variables, chemical composition variables, and their combination. Artificial neural network (Multilayer Perceptron, MLP), decision trees (REPTree and M5P), Random Forest (RF), and Multiple Linear Regression (LR) were tested. The principal component analysis revealed that 61.3% of the total variance was explained by two components, highlighting a strong association between digestibility and crude protein content and an opposite relationship with lignin and neutral detergent fiber. Among the evaluated models, MLP, LR, and RF achieved the best performance for leaf digestibility (r = 0.76), while for stem digestibility the highest accuracy was obtained with the LR model (r = 0.79; MAE = 2.42; RMAE = 2.87). The REPTree algorithm presented the lowest predictive performance regardless of the input data. The results indicate that chemical composition variables alone are sufficient to develop reliable prediction models. These findings demonstrate the potential of ML techniques as a non-destructive and cost-effective approach to predict the nutritional quality of tropical forage grasses and support precision livestock management.